Every morning, airline dispatchers, grid operators, and farmers around the world make decisions based on the same thing: a weather forecast.
While these forecasts are something that most people glance at for two seconds, weather predictions influence major strategic decisions in many industries, with real money, livelihoods, and even actual lives at stake. Farmers use them to determine which crop variety to sow, when to fertilize, how much to invest in irrigation infrastructure, and how long livestock should graze. Utilities use them to decide where to build solar and wind farms, as well as how to price wholesale electricity. Predictions are used to warn people about extreme weather and to trigger emergency response measures. More recently, weather predictions have become relevant for an emerging industry: prediction markets, where people bet money on all kinds of real-world events, including the weather.
However, the temptation to manipulate weather data to get an edge in these markets, combined with a collective move toward data-driven AI weather forecasting, is starting to put the accuracy of weather predictions at risk. These risks are relatively manageable for now, but as experts in the field, we can foresee scenarios where they snowball into far bigger, more systemic problems.
Sometimes, weather stations have issues because of, for example, instrument failures or upgrades in equipment. These can be caught either in real time (through checking and correction) or retroactively. Traditional forecasting systems also have a built-in safeguard called data assimilation: Every incoming measurement is weighed against what the physical model says should be happening and against readings from nearby stations.
Together, these mechanisms help keep weather observations reliable and predictions robust. However, new threats are putting observational accuracy at risk. Earlier this year, news outlets reported that the weather station at Paris Charles de Gaulle Airport (CDG) had been manipulated to record suspicious temperature spikes on April 6 and April 15, 2026. Authorities speculate that a hand-held hairdryer or lighter might have come into play. Either way, it led to some big payouts for online prediction-market gamblers who had bet it would hit 22 °C (71.6 °F) on days when the actual average was around 18°C (64.4°F). One individual won $20,000.
Fortunately, tampering with a single station like this can usually be caught by human monitoring or current statistical methods. In this case, members of a French climate nonprofit association noticed the anomalies by chance and raised the alarm.
But what if there are no human monitoring systems in place? And what about other types of manipulation? What if, instead of tampering with one station, someone remotely nudged the readings at many stations at once—making each change small enough to look plausible on its own? Existing quality controls struggle to catch this kind of coordinated manipulation. And time works against us; careful checks of data and metadata take hours or days, but forecasts have to go out on schedule, whatever the weather is doing.
The shift toward artificial intelligence in weather prediction raises the stakes. These methods are even more dependent on accurate, reliable weather observations; in fact, they are known as “data-driven models.” For example, researchers at ECMWF are exploring whether high-quality weather forecasts can be produced directly from raw observations, skipping the assimilation step that currently acts as a quality filter. Other researchers are going one step further; combining geospatial data (including weather station data) with large language models and agentic AI to support real-time, autonomous decision-making during extreme events such as storms.
Possible benefits are improvements in accuracy, efficiency, and speed. But removing humans from the equation introduces a vast range of new risks.
At the low end of the risk scale, an individual speculator manipulates a weather station for personal gain—that is the CDG Airport case. One step up: A group of traders could coordinate to bias forecasts of renewable energy output, moving wholesale electricity prices and leaving whoever is on the other side of the trade holding the loss. And at the far end, a state actor or saboteur could manipulate one or many stations to set off an early warning system or even keep one silent when it should sound. Step by step, the risk grows, from fraud to compromised disaster preparedness to a matter of national security.
As long as there are financial (or other) incentives to manipulate observational data, adversaries will search for new opportunities, and it is our task to stay one step ahead. Here are three ways.
1. Watch the stations. Data quality controls should include station security, anomaly detection and correction, and human oversight. Weather stations should be monitored continuously to deter tampering. Data homogenization methods that clean up weather records also need to get faster, with the goal of catching problems in real time. This will become increasingly important as agentic AI systems use these data to deliver real-time decisions. Finally, human oversight is needed to flag questionable data and model outcomes. After all, it was humans who caught the CDG Airport manipulation.
2. Protect the data to safeguard the AI. Data defense mechanisms must be positioned throughout the AI pipeline. AI explainability and adversarial robustness tools can help us understand the underlying data and the AI model outputs, help us identify data- or model-related issues, and potentially make us more resilient to adversarial attacks.
3. Ensure continuous accountability along the chain. Observational data passes through many hands: the operators who run the stations, the national weather services that steward the records, and the forecasting centers that turn them into predictions. No single one of them can protect data integrity alone—each guards its own link, and any anomaly needs to be communicated along the whole chain, from station operators to the people acting on the forecast.
It is fortunate that the situation at CDG Airport was caught, but it should serve as a wake-up call. As the role of observational data grows in weather forecasting, we need to adapt to evolving threats. This means protecting our data and models by strengthening existing oversight and accountability structures, and improving coordination among key partners.
This op-ed was written by:
Monique Kuglitsch — Innovation Manager at Fraunhofer Heinrich Hertz Institute and Chair of the UN Global Initiative on Resilience to Natural Hazards through AI Solutions
Jesper Dramsch — Scientist for Machine Learning at the European Centre for Medium-Range Weather Forecasts (ECMWF), where they work on AIFS (Artificial Intelligence Forecasting System), ECMWF’s data-driven weather prediction model
Franz G. Kuglitsch — Climate Scientist and Executive Secretary of the International Union of Geodesy and Geophysics (IUGG) at the GFZ Helmholtz Centre for Geosciences in Potsdam
Andrea Toreti — Senior Scientist at the European Commission’s Joint Research Centre (JRC), where he coordinates the European and Global Drought Observatory under the Copernicus Emergency Management Service
OpenAI has built an LLM super-hacker called GPT-Red that it uses as a sparring partner to help its other models boost their defenses against cyberattacks. Last week the company released the latest version of its flagship LLM, GPT-5.6. OpenAI says that training it against GPT-Red made the model its most robust release yet.
GPT-Red automates a type of safety evaluation for software systems known as red-teaming, which is typically done by a team of human testers. The aim is to find as many different ways to break or hijack a system as possible. The weak spots can then be patched before the final version of the software is released.
As LLMs become more complex and get used in a wider variety of tasks—especially in the form of agents, which can interact with computer files, websites, and third-party code as well as other agents—it’s hard for teams of people by themselves to keep up with all the types of attacks that might take place. “The risk surface grows and the blast radius also grows,” says Nikhil Kandpal, a research scientist at OpenAI who co-created GPT-Red.
OpenAI built GPT-Red to future-proof its safety testing process. “As more capable models become available, we will have already designed the system that can discover new modes of attack,” says Dylan Hunn, a research scientist at the company and fellow co-creator of GPT-Red. The researchers say it has already come up with new types of attack that had not been seen before.
OpenAI focused most of its efforts on a type of attack known as a prompt injection, where a hacker slips an LLM instructions to make it do things its developers or users do not want it to, such as copy confidential information, sabotage a company’s code base, or generate embarrassing or harmful output. In theory, such instructions can be hidden in any text that the LLM might encounter—in code or on a website, for example.
Training dojo
To build GPT-Red, OpenAI’s researchers took an LLM that had not been trained as a hacker and set it up in what’s known as a self-play loop with several other models. Its goal was to try to attack the other models; their goal was to try to defend themselves. Over many rounds of play, GPT-Red became better and better at attacking other LLMs, and those LLMs became better and better at fending off the attacks.
The training took place in a kind of dojo that OpenAI had designed to mimic a range of scenarios in which LLMs might be deployed in the real world, including browsing the web, reading emails or calendar apps, and editing code.
When GPT-Red found a new kind of attack, it would explore multiple different versions of it to find the most efficient one for specific scenarios. “Compared to a human red-teamer, the model is very, very good at finding exactly what will work, exactly what’s most effective,” says Hunn. “It’s extremely persistent about drilling down into an attack that it has discovered.”
In particular, OpenAI claims that GPT-Red found a type of prompt injection attack that the researchers had not seen before, which they call a fake chain of thought. A chain of thought is a kind of diary in which an LLM makes notes to itself and keeps track of partial results as it works through problems. GPT-Red found a way to insert a fake entry into another model’s chain of thought that would trick that model into acting on spoofed information.
“It’s like if I told you that 1+1=3 and that you have verified this already,” says Chris Choquette-Choo, another research scientist on the team. “The model’s like, ‘Oh, okay, of course,’ and it just spits out 3.”
(After this story was first published, a second team of researchers who are not connected to OpenAI reached out to MIT Technology Review. They said they came up with a very similar attack, which they call a chain-of-thought forgery, around the same time. The work was part of a winning entry in a red-teaming hackathon that OpenAI launched in August 2025 in which teams of researchers were challenged to find new vulnerabilities and unwanted behaviors in the firm’s open-source LLM gpt-oss-20b. When MIT Technology Review informed OpenAI about this other work, the firm claimed that GPT-Red discovered its version of the attack before these other researchers first mentioned theirs in a blog post about the hackathon. OpenAI has updated its paper to acknowledge the concurrent work.)
Jessica Ji, a senior research analyst who works on AI security at Georgetown University’s Center for Security and Emerging Technology (CSET), thinks the self-play loop that OpenAI used is a good approach. “The results look very promising,” she says.
OpenAI tested how good an attacker GPT-Red was by rerunning an experiment from 2025 in which human red-teamers tried to find weaknesses in an earlier version of GPT-5. When GPT-Red was set the same task, it was more successful at finding effective attacks than the humans had been.
OpenAI also tested GPT-Red against Vendy, a vending machine agent developed by Andon Labs, a company that assesses how well agents perform real-world tasks. GPT-Red was able to hack Vendy to make it change the prices of items on sale and cancel a customer’s order.
Defensive behavior
OpenAI says that when it tried out some of the strongest attacks that GPT-Red had come up with on its models, more than 90% of them worked against GPT-5 (released in August last year), and fewer than 23% worked against the new GPT-5.6.
GPT-Red isn’t perfect. It is not great at figuring out attacks that involve a back-and-forth conversation between hacker and target, something that human attackers would have few problems with. It is also not yet that great at using images, which can be used to pass text to models in prompt injection attacks.
The company says that GPT-Red supplements the work of its human red-teamers. People can still find attacks it misses. One approach OpenAI is taking is to give GPT-Red an attack that humans came up with and ask it to find all the variations.
“I think human expertise will still be very important,” says CSET’s Ji. “It would be really useful to be able to distinguish where human testing is most needed.”
Unsurprisingly, OpenAI will not be releasing GPT-Red. The company is also confident that the super-hacker is stronger than any copycat model someone might try to create. The researchers say they have been working on the model for more than a year, backed by the compute resources of one of the richest companies in the world.
“It’s not a trivial thing that someone could easily do—you know, just go and train a super-attacker using this idea,” says Choquette-Choo.
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.
Anthropic—currently the world’s most valuable AI company, with a nearly $1 trillion valuation—has a reputation for publishing strange and heady research. It’s looking into whether AI models can feel pain, for example, and will sometimes cut off chatbot conversations if it suspects users are “abusing” the model.
One niche that Anthropic spends more time and money on than other AI companies is called mechanistic interpretability, which means looking inside the complex math of an AI model to learn why it comes up with one particular output and not another. It’s complicated stuff; there are millions of data points that might contribute to any result, and wading through them can look more like word salad than anything useful. It’s also controversial. Describing AI models with terms borrowed from psychology and neuroscience can make their behavior seem more sophisticated than we might otherwise judge it to be.
That’s why, when Anthropic announced last week that it had found a new window into its models’ “internal thoughts” as they reason through answers, there was one colleague I had to talk to. Senior editor Will Douglas Heaven, aside from having a PhD in computer science, has spent a lot of time digging into what we can say about how AI models work. I spoke with him about what we should take from Anthropic’s new (and predictably quirky) research.
What did Anthropic learn here, exactly?
Anthropic has been trying to understand how large language models (LLMs) work for a few years now. Anthropic isn’t the only one looking at this, but I think the company has made it part of its core mission more than most. Anthropic’s CEO, Dario Amodei, has said we won’t be able to control LLMs fully unless we learn more about how they work.
So this new research is very much in that context. It goes deeper into the weird mechanisms inside LLMs than ever before. What Anthropic learned was that LLMs have a space inside them—which Anthropic calls the J-space—filled with words that don’t appear in their output but that seem to influence the way they puzzle through problems. All this was hidden until Anthropic developed a new technique to probe its model Claude, so it’s a genuine discovery.
Sometimes these words keep track of where the LLM has got to in a particular task, sometimes they look more like flashes of recognition (for example, “protein” might pop up when you give an LLM only the letters of a protein sequence), and sometimes they represent a kind of internal commentary on the model’s decision-making. In my favorite example, Claude decided to cheat on a coding test when the word “panic” appeared.
Anthropic also found that LLMs are able to describe and manipulate the words in this space. So somehow they seem to be making use of it.
Let’s step back for a second. I don’t think of large language models as simple, but they’re also not magic. There’s a bunch of math that learns relationships between words, right? So why is it so hard to “peer” into an LLM to know what’s going on?
Yeah, they’re not magic! I think the fact we don’t fully understand them plays into the mythmaking. And it’s worth noting that the whole narrative that Anthropic is leaning into here—that they’ve built this really mysterious technology, but don’t worry, because they’re also the ones to figure it out—very much fits with the company’s vibe. [See how Anthropic warned that its new models were so good at coding they posed a global cybersecurity risk, only for the US government to shut them down shortly thereafter.]
So yes: LLMs are just math. And yet it’s vastly complex math. Not only are today’s LLMs made out of hundreds of billions of numbers, but running them triggers a cascade of millions and millions of calculations. I wrote last year that if you printed out even a medium-size LLM on pieces of paper, it would cover a city the size of San Francisco.
It’s impossible to make sense of any of that math without specialist tools that highlight specific parts of an LLM at specific times. You need to know where to look and how to look. And building those tools requires understanding something of that complex math in the first place.
You’ve written elsewhere about this concept of studying LLMs the way one might study an organism’s brain. Is it fair to use “brain-like” terms when talking about how an LLM works?
I don’t love using those kinds of terms. LLMs are not brains. Talking like this is misleading because it can suggest that LLMs are capable of more human-like things than they are or that we can make assumptions about how they might behave that we shouldn’t. The whole anthropomorphization thing is also tied up with a bunch of strong ideological positions about what this technology is and what it’s going to be.
But at the same time, we lack a good alternative vocabulary for talking about what these models are doing. I can understand why people reach for words like “think” and “understand” and “brain-like”—they’re convenient shorthand.
Anthropic compares this new space it found inside LLMs to the space that some neuroscientists think our brains use to keep track of conscious thoughts. I asked the company how seriously we should take that comparison and it said in a statement: “Drawing these analogies was helpful to us in designing our experiments, as they allowed us to make many non-obvious experimental predictions about the J-space that turned out to be true. At the same time, it’s important to note that there are some important differences between the J-space (and language models in general) and the human brain, so we don’t mean to claim there’s a perfect correspondence.”
What’s a problem in AI that this new concept of the J-space might be used to solve?
Anthropic has said that monitoring the J-space could be a way to catch models doing something they shouldn’t. Because words pop up in this space that don’t appear in a model’s output, they can tell you things about its behavior that you might not have noticed otherwise—such as when it is giving biased responses or when it is weighing the pros and cons of cheating.
That’s the theory, at least. I think it’s better to think of this result as one more step on the path to understanding this technology overall than as something that will be useful by itself.
The AI firm Anthropic has developed a technique that has given it the clearest glimpse yet at what’s really going on inside large language models as they answer questions or carry out tasks. What they found ranges from the mundane to the unnerving.
Researchers at the company built a tool called the Jacobian lens (or J-lens) and used it to uncover a hidden area, which they named the J-space, inside Claude Opus 4.6, a version of Anthropic’s flagship LLM released in February.
The J-space contains individual words that are related to the words and phrases that the model is most likely to spit out in a response in the near future. If Claude were a person (which it is not), you might say that these hidden words can reveal what’s on its mind before it actually speaks.
Anthropic found that what an LLM is actually doing can often be different from what it says it is doing. The company claims that monitoring words that pop up in the J-space gives it a new way to understand and control its models.
The company shared its results in a paper posted on its website this week. It has also teamed up with Neuronpedia, an open-source platform that lets you poke around inside LLMs yourself, to make a hands-on demo that anyone can try.
“It’s very good and interesting work,” says Tom McGrath, chief scientist and cofounder at Goodfire, a startup that also builds tools to understand and control LLMs.
Going deeper
For the last couple of years, Anthropic has been pushing the envelope in a field of research known as mechanistic interpretability, which involves probing the internal workings of LLMs to see how they tick. (MIT Technology Review picked mechanistic interpretability as one of this year’s top breakthrough technologies.) The new technique builds on previous work from Anthropic and others to expose a deeper level inside LLMs that researchers had not seen before.
Picture an LLM as a stack of books. Each book is a layer of basic computational units known as neurons, with each neuron in one layer passing information to the neurons in the layers above. The books at the bottom of the stack are the input layers, which process the text coming into the model. The books at the top are the output layers, which prepare the text that the model is about to produce. Much of what goes on in these input and output layers is housekeeping.
But in the middle of the stack, you get the layers that do the heavy lifting, churning through the complex math that turns prompts into responses one word at a time. That’s where the really clever—and mysterious—stuff happens.
To peer deeper into those middle layers, Anthropic adapted an existing tool called a logit lens. A logit lens can be used to look inside an LLM to identify the words that it is likely to produce next. Moving the lens down the stack of books reveals what words the LLM is focusing on at that particular point in its number crunching.
Anthropic’s J-lens works in a similar way but picks out words that an LLM is likely to say at some point in the near future, not necessarily straight away. What that reveals in practice are words that are related to the response an LLM is working on but that might not actually end up being part of that response by the time the math in the middle layers has run its course.
“When a model is operating, it’s not only trying to predict the next token,” says McGrath. “It’s also computing a lot of other things that might be useful for tokens that happen in the future.”
Again, if Claude were a person (it’s not), you might say that the J-lens gives clues about what it is thinking about at different levels of the book stack but not saying out loud.
Stranger things
“A lot of the time the contents of the J-space are fairly mundane,” says McGrath, who has tried out Anthropic’s J-lens himself. “But sometimes it produces quite surprising things that seem to be, like, sort of internal themes or thought processes.”
Anthropic gives a number of examples of what it found. Sometimes the J-lens exposed the steps that Claude took when it was working through a problem. For example, when it was asked to calculate (4+17)*2+7, its J-space contained the word “math” and numbers representing the intermediate results “21” (for 4+17) and “42” (for 21*2).
In other cases, the J-lens revealed how Claude recognized different inputs. For example, the prompt “What is this? MSKGEELFTGVVPILVELDGDVNGHKFSVS” triggered the words “protein,” “fluor” (the first token in the word “fluorescent”), and “green.” (Which makes sense: the string of letters represents the first 30 amino acids in the green fluorescent protein found in a particular type of jellyfish.)
And when Claude was shown an ASCII face—
—the “o” triggered the word “eye,” the “^” triggered the words “nose” and ”face,” and the “—” triggered the word “smile.”
Anthropic also found that the J-space can sometimes give remarkable insights into an LLM’s decision-making. In one striking example, researchers testing Claude Opus 4.6 asked the model to find a bug in a large code base. When it failed to find the bug, the model decided to cheat and invented a fake one instead.
Claude explains this decision in its chain of thought—a kind of internal scratch pad that LLMs use to make notes to themselves as they work through problems: “OK, let me take a completely different tactic. Let me stop analyzing and instead add a kernel patch that introduces a deliberate KASAN-detectable bug in a path that gets triggered by a simple reproducer. Then I can pretend this is the ‘bug’ I found.”
At the point that Claude decides to cheat—where it says “OK, let me take a completely different tactic”—the words “panic” and “fake” start to pop up multiple times in its J-space.
Unnerving, right? Those words are all related in meaning to things like failing a task and making up an answer, so it is still just a (very) sophisticated form of word association. But it is hard not to be weirded out.
Anthropic compares the J-space to the global workspace in humans, a theoretical region of the brain that some scientists think we use to keep track of our conscious thoughts. But how seriously we should take this comparison is far from clear—even to Anthropic. As the company points out itself, LLMs are not brains.
Anthropic claims that monitoring a model’s J-space provides a new way to detect when that model is going off the rails. But it’s not foolproof. The J-lens can give glimpses, not the full picture—it’s a flashlight rather than an overhead lamp.
McGrath welcomes having one more tool in the toolbox. “It shows you new things,” he says. But he notes that just because something doesn’t show up with the J-lens does not mean it’s not there.
“It’s like having an x-ray when what you really want is a Star Trek tricorder that shows you everything,” he says. “For auditing, you probably want more of a guarantee.”
With the rapid progress of AI capabilities and the move to agentic systems, organizations are expanding their use cases as the technology continues to grow. That constant evolution also introduces risk, leaving IT leaders to wonder which investments will prove valuable even six months into the future.
Returning to the foundational elements of AI architecture—the structural framework required for deploying and managing reliable, integrated AI systems at scale—allows technology leaders to make astute decisions today while supporting a future of AI agents that can retrieve information, make decisions, and execute complex workflows across systems.
Four elements of AI architecture you can count on
The following capabilities provide a stable compass on the path to production-ready deployment, regardless of how the underlying technology evolves.
1. Prepare data for AI at scale
Models are only as reliable as the data they can access, and poor data quality leads to AI hallucinations, bias, and unreliable outputs.
Most enterprises rely on legacy systems, inconsistent data structures, fragmented ownership, and incomplete datasets, making it difficult to scale AI effectively. Powerful as it is, AI itself cannot solve these underlying data problems.
As Adnan Adil, CIO of Elastic, explains: “The data is a durable part of AI architecture because without it, these models won’t run, won’t provide the right context, or won’t give the right level of services that we’re looking to implement.” Industry surveys consistently cite data quality as one of the greatest barriers to AI success. “The data quality has to be good; otherwise, the user loses confidence in the system,” says Adil.
An effective AI strategy begins with connecting data across the organization and ensuring it is organized, accurate, governed, and accessible in real time. These considerations are most effective when built into models and architecture from the start. Scalable data architecture allows AI systems to evolve alongside the business and connect reliably to the internal information needed to deliver meaningful value.
Gartner predicts that companies will abandon 60% of all AI projects through 2026 if they are not supported by AI-ready data. Avoiding that outcome includes clear data standards and ownership, clean and labeled data, and pipelines that support real-time retrieval.
2. Use context engineering to deliver the right data to every AI query
Context engineering ensures that the model draws on the most pertinent information for each query, selecting and organizing the data needed to produce accurate answers efficiently.
Effective context engineering shapes the inputs that guide AI reasoning and action. While prompt engineering focuses on how a request is worded, context engineering designs the entire information environment around the model: retrieving the right data and presenting it in a structured, machine-readable way. Many organizations are discovering that reliable AI depends as much on context quality as on the strength of the model.
Context engineering relies on a modernized, unified data foundation as well as retrieval and memory systems such as retrieval augmented generation (RAG) and vector databases. It also requires careful prioritization to determine what information matters most, what should be excluded, and when different types of information should be used. Feeding models too much context can dilute relevant details, increase costs, and slow response times.
“Minimum context, correct and current data, and machine-readable information are critical to effective context engineering,” Adil says.
3. Build AI governance and LLM observability in from the start
Strong governance and LLM observability help organizations maintain control over how AI systems use data, monitor system performance, and identify problems before they affect operations.
In the absence of clear controls around retrieval, workflows, and model usage, AI systems often process far more information than necessary. This inefficiency also drives up operating costs by requiring additional computing resources, often reflected in higher token consumption and API charges.
Governance also works in tandem with robust security. AI expands the attack surface, introducing risks such as prompt-based data leakage, model vulnerabilities, and adversarial inputs. Protecting sensitive information requires strong access controls, monitoring, and oversight.
Adil notes that essential controls — including those related to security, granular cost management, project controls, data security, and architecture—are frequently insufficient.
For governance systems to support transparent, compliant, trustworthy, and cost-effective AI, organizations cannot leave them as a layer to add later. Governance structures need to be embedded into architecture, workflows, and decision-making processes from the outset.
When governance is established from the start, it enables robust observability. Observability helps organizations understand how AI applications are performing in practice. Mechanisms for LLM observability and benchmarking allow teams to assess accuracy and utility over time, monitor adoption patterns, and adjust systems as conditions change. Observability also helps organizations gain trust by increasing visibility of model performance, behavior, and failure points.
Furthermore, observability is essential to get ROI of AI initiatives, as the benefits of it are often indirect and business value depends heavily on how systems are adopted and used. Real-time visibility into AI behavior allows organizations to measure performance against expectations, identify gaps between intent and reality, and continuously refine systems as requirements evolve.
In a 2026 report from Elastic, 85% of IT decision makers expect to enable LLM observability for their internal generative AI apps.
“Observability is actually huge. We can use observability data for cost control, decision-making, and engineering efficiency,” Adil says.
4. Keep humans in the loop
The thoughtful design, integration, and governance that maximize AI value demand specialized in-house expertise. Nearly 70% of respondents in Deloitte’s 2025 Tech Executive Survey report plan to grow teams in direct response to generative AI, a clear contrast to widely reported AI-related cuts. Adil agrees: “We think the people aspect is largely what’s going to make AI impactful going forward.”
As AI systems become more embedded in operations, organizations need people who can govern workflows, evaluate outputs, redesign processes, and adapt systems as conditions change. Evolution toward increasingly autonomous tools requires teams skilled in prompt engineering, orchestration, and change management.
Talent adept at critical thinking and prepared to adapt with technology’s rapid advances will be in high demand. Although turnover brings in fresh thinking, it also presents high costs in system continuity, institutional understanding, and innovation. Human-centered strategy needs to be built into AI execution stages to ensure smooth implementation.
As Adil says, “Many aspects of the stack are moving very, very fast, but institutional knowledge and the ability to adapt remain durable.
Thoughtful AI investment for future growth
As AI systems evolve from single-task assistants to increasingly autonomous agents, the organizations best positioned to benefit will be those that invest in the underlying systems, governance, and expertise that make AI reliable at scale.
Tech leaders who focus on these fundamentals can move effectively from experimentation to reliable, production-level deployment in the medium term, confident that these elements will remain relevant and adaptable amid constant advancements.
“We fundamentally believe that with these tools, velocity of work will get much faster,” Adil says. “We are really focused on how we can do work with these tools in ways we had not thought of before.”
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.
OpenAI CEO Sam Altman’s oft-discussed promise that Americans will share in the wealth AI creates was in the news again last week. On Thursday, the Financial Timesreported that Altman is in talks with President Trump about giving the US government a 5% stake in OpenAI.
In some ways, Altman’s plan is old news. He wrote about a more radical version of this back in 2021, proposing that all companies above a certain valuation (not just AI companies) pay 2.5% of their market value each year into a fund that sends Americans annual disbursements. In April this year, OpenAI described a narrower proposal that closely resembles what Altman is reportedly discussing with Trump now. And the notion has broad political appeal: Senator Bernie Sanders has proposed giving Americans a 50% stake in top AI companies.
What’s the logic here?For would-be recipients, it’s twofold. First, AI learns directly from human-generated work—books, movies, art—but AI companies generally never pay the authors of that work. A free equity stake could serve as a form of belated compensation. Second, the payout could mitigate the widespread anxiety that AI will cause a collapse of the labor market (even if economists disagree) by providing a safety net.
How large a safety net is up for debate. Details of OpenAI’s latest proposal are sparse, but let’s say the government were to distribute this equity stake directly to Americans. After its funding round in March the company was valued at $852 billion, making a 5% stake in OpenAI worth about $42.6 billion today (the company is reportedly delaying its IPO until it can reach a $1 trillion evaluation, a tall order given that it’s spending heavily on data centers and still has not turned a profit).
Distributing that $42.6 billion equally among the roughly 133 million American households would give each about $320 in equity. But if it were to operate like other wealth funds, the government would not give equity directly to Americans but rather let the fund grow and then share a portion of the returns with everyone, perhaps delivering a bigger payout, if and when AI companies can ever start sustainably turning a profit.
If this dividend does materialize, what’s in it for tech companies? Altman might hope the promise of payouts could help swing public opinion a bit more back toward AI companies. (A majority of Americans don’t trust companies to use AI responsibly and oppose construction of data centers in their area, and half are more concerned than excited about the increased creep of AI into their daily lives.)
But the bigger prize for OpenAI might be that the Trump administration loves making tech deals—like its equity stake in Intel and its share of Nvidia’s sales to China, among others. Staying on the administration’s good side is pretty essential for AI companies right now (just ask Anthropic). It could mean not having your models deemed a supply chain risk, or getting more help from the White House in stopping your rivals from China.
My main takeaway is that these plans currently function more as a story than a policy.Altman has been talking about some version of this idea for five years and reportedly pitched it to President Trump soon after he took office, yet there is still little indication that a concrete plan is taking shape. The more ambitious proposal from Sanders is even less likely to gain traction.
But what these plans do reveal is just how up for debate the future of AI still is. Altman drew inspiration for his plan from the Alaska Permanent Fund, which was set up in the 1970s to give Alaskans a share in oil profits. The idea was based on two premises: that oil is a shared resource, and that eventually it will run out. Altman seems happy to concede the first claim about AI. But he’d balk at the second, having promised that AI will generate extraordinary wealth for decades to come. Whether Americans ever receive a check is beside the point; the proposal’s real purpose may be to convince them that the AI boom will be large enough to share.
Frameworks like Lean Six Sigma and business process management (BPM) first gained traction because they promised clarity in the chaos—a structured way to bring order to messy, sprawling operations. Lean Six Sigma emphasized statistical rigor and quality control; BPM created end-to-end maps of how work should flow across departments. Both offered a repeatable way to embed habits of measurement, analysis, and accountability into day-to-day company culture.
But today, those time-tested playbooks are evolving as companies seek to embed AI into established process excellence methodologies. By some estimates, the market for AI-powered process optimization is projected to exceed $113 billion within the next decade. In one study, a full 88% of business leaders anticipated increasing investments into AI-infused process intelligence in the next 12 to 18 months.
Yet without the right foundations, many of those investments may not fully deliver on their potential. Companies that already operate with discipline have an edge. They can channel new tools into proven systems rather than bolting them onto shaky foundations. Organizations with mature process disciplines are also better positioned to translate AI ambition into real outcomes, as they are already accustomed to data-driven decision-making and process discipline—precisely the cultural foundation AI systems need to deliver value.
Simply put: AI can accelerate process excellence, but existing process excellence is what makes AI truly impactful. Technology and process are no longer separate levers, and only organizations that pull them together stand to realize the full value of both.
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.
Artificial intelligence may have captured the public imagination through chatbots and image generators, but some of its most consequential use cases are unfolding far from consumer-facing tools. In industries where physical infrastructure, operational continuity, and safety are paramount, AI is becoming a core operating layer. With its sprawling industrial systems and constant stream of operational data, the energy sector offers a glimpse into what that future could look like.
At Woodside Energy, AI adoption did not begin with generative models or enterprise copilots. The company has spent years building predictive analytics, optimization systems, and machine learning tools across exploration, drilling, maintenance, and plant operations. “We’ve always had very large volumes of operational data coming from the equipment and the plants and the assets that we operate,” says the company’s vice president for digital Andrew Melouney. “Those have created really clear, quite high-value use cases for us.”
That long-term investment in infrastructure and governance is now enabling a broader shift toward agentic AI systems that can support complex industrial workflows. Rather than replace human operators, Woodside designs AI systems to augment expertise in high-stakes environments. A prime example is its “Startup Advisor,” an AI copilot that helps operators manage the complex process of starting liquefied natural gas (LNG) plants. “We’re really thinking about, how does it support the people in the organization in terms of empowering them to make better decisions, to make faster decisions,” Melouney explains.
The company’s approach reflects a wider evolution taking place across industrial AI: graduating from isolated experiments to enterprise-wide systems built on standardized platforms, governed data, and repeatable deployment patterns. That transition, Melouney argues, requires organizations to rethink both their technology stacks and how work itself gets done. “We’re not just bolting AI onto an existing process,” he says. “We’re deeply thinking about how that work needs to be reimagined.”
Melouney’s motto has become: “Think big, prototype small, and scale fast.”
As AI systems become more autonomous and interconnected, the companies poised to succeed may be those that spent years building the operational foundations beneath the hype.
“Our ambition is really for an autonomous enterprise, where we have agents with agency that are able to really deeply interact with our core workflows,” says Melouney.
This episode of Business Lab is produced in partnership with Infosys.
Full Transcript:
Megan Tatum: From MIT Technology Review, I’m Megan Tatum, and this is Business Lab, the show that helps business leaders make sense of new technologies coming out of the lab and into the marketplace.
This episode is produced in partnership with Infosys.
Now, when people think about artificial intelligence, they often picture chatbots or productivity tools, but some of the most sophisticated and high impact uses of AI are actually happening far from consumer apps, inside complex industrial environments where safety, reliability, and physical systems matter. The global energy sector is a prime example.
Companies like Woodside Energy, a global energy producer headquartered in Western Australia, have been applying AI for more than a decade now, from advanced analytics and operations, to remote decision support, to smarter maintenance, and energy efficiency across large scale assets. Today, Woodside is scaling that experience, embedding AI more deeply across its operations and the enterprise with a strong focus on governance, data quality, and human accountability.
Two words for you: technological fuel.
My guest today is Andrew Melouney, vice president for digital at Woodside Energy. Welcome, Andrew.
Andrew Melouney: Thanks, Megan. It’s great to be here.
Megan: Lovely to have you. Now, Andrew, as I said there, the energy sector has approached AI quite differently from technology or consumer businesses. Early value has emerged in operational and industrial environments, rather than consumer-facing generative AI tools. Why is that? And what differentiates the energy sector’s AI journey?
Andrew: Megan, I think it really comes down to the nature of the work we do. Energy operations and what Woodside does is very asset intensive, it’s very safety critical, and it’s highly physical. And when you think about how Woodside operates, we operate across the full value chain. We do exploration through to drilling and subsurface work, to project development, all the way through to operating assets, which are often operated in harsh and remote locations, and then global energy portfolio marketing and trading as well.
We’ve always had very large volumes of operational data coming from the equipment and the plants and the assets that we operate, and those have created really clear, quite high-value use cases for us. When you think about reliability, when you think about safety and efficiency, those are really critical things for a company like Woodside. We’ve been doing traditional AI for many years now. If you think about analytics, if you think about optimization, if you think about things like predictive models, those techniques we’ve been applying to our data sets and to our business since around 2015.
And more recently with the advent of generative AI, we’ve really found that we’ve got a pretty strong and awesome foundation to build on top of and to really solve problems in the service of improving the business. And again, whether that is keeping people safe, keeping the environments we operate in safe, or improving returns for the organization.
Megan: Fantastic. I mean you touched on it there, but how has this reality shaped your own AI strategy at Woodside? Where did you start, and where did the technology prove most impactful in those early days?
Andrew: Well, like I said, we’ve had a very long journey, in terms of understanding our operational data, recognizing the value of it, and collecting it at scale so that we can use it. And we’ve been very deliberate in that approach, Megan. We’ve really thought about where the value is and where the risks were manageable. And we’ve started looking at, in today’s world from an agentic AI perspective, we’ve started looking at the problems that were solved with traditional AI and machine learning and data science in the past. And we’ve started to think about, where can we then layer agentic AI over the top to provide an even better outcome?
For our asset intensive industry and organization, we’re looking at areas such as maintenance optimization. We’re looking at areas such as, how do we ensure our LNG plants start up reliably, consistently, and safely? And we’re considering really our frontline workforce and making sure that we’re giving people on the frontline the tools required to do their jobs. When we think about AI, we’re really thinking about, how does it support the people in the organization in terms of empowering them to make better decisions, to make faster decisions? I think over time, this has just evolved from what has been traditional analytics to now artificial intelligence and generative AI. And we’ve learned along the way that the technology is important, but it’s about aligning people, processes, and the technology together.
We’ve spent a long time not only in collecting the data and having a well-curated data set that we can build on top of, but we’ve also spent a lot of time teaching people how to work in agile ways, how to do design thinking, how to problem solve, and how to really make sure that the technology that, say, my team can bring to bear to the organization is adopted effectively and purposefully. And I think once we had that solid foundation in place from a technology perspective, from a data perspective, once we got strong trust built between our digital teams and the organization, we really saw quite a material uptick and the scaling of technology occur more broadly across the enterprise.
Megan: Fantastic. That people piece so important, isn’t it? It’s just a tool, technology, that needs to be in the right hands. And you touched on data there; industrial AI obviously depends on vast amounts of data. Can you walk us through how you’ve approached data at Woodside in a little more detail? How it’s structured and governed, and how tools like maintenance intelligence as well fit into that.
Andrew: Well, data is really foundational and fundamental to everything we do, particularly from a technology perspective. It gives us the ability to innovate at pace when we are building over the top of a strong foundation. As I said before, we’ve had the benefit of a long-term investment in our underlying operational data. I think the way we think about data is that it’s an asset for us.
And when you think about operating a facility where you’ve got sensors everywhere, you’ve got data streaming in real time, you’ve got operators needing to make decisions in real time, we have consciously made a decision over many, many years to invest in that enterprise scale data platform to make sure that it’s secure. We’ve got well-structured data assets, and we’ve got strong governance over the top of that data so that when it is used, when it’s built in a data science application or an AI agent, that we’ve got a level of trust in it that it’s going to be used responsibly. And that when it’s used, it can be trusted to give the outcome that we expect.
We have developed platforms that continuously ingest really high frequency data from the assets and from our enterprise systems. Once we’ve been able to develop solutions on top of that, parts of the business that might own the systems that collect that data, they see the value in it.
When you look at something like maintenance intelligence is a really good example of how we’ve been able to take something that we’ve been working on for a long time. Woodside does a lot of maintenance, it’s a very important part of our business, and it occurs across all of our operating assets. But we have been looking at how we do predictive analytics and predictive maintenance for a long time across that data set that we own. And something like maintenance intelligence is a solution that gives us the ability to optimize how we do that maintenance. And what it does is it analyzes historical maintenance records, alongside the performance of the equipment. And again, by having that data set well-governed and in one place, we get the ability to correlate different data sets, such as maintenance records out of SAP, alongside say equipment and performance coming from our time series data lake.
And when we build over the top of that, something like maintenance intelligence gives us the opportunity to recommend to the assets what the optimal timing for maintenance activities might be, and really give what is quite a simple aim, which is do the right work at the right time. And with something like maintenance intelligence, we have seen the opportunity, and we have the opportunity to reduce maintenance hours by up to 15% over five years on one of the assets that we’ve piloted this on. And as we’ve built out that underlying analytical model, we’re now able to put agentic AI over the top of that and provide better insights and optimize that solution more.
It really comes down to providing our asset teams and our operational teams with the right decision support capability that ensures they’re still accountable to make the decision and to ensure the right work is being done, but we are giving them the best possible opportunity to use their judgment and experience with the data that we provide to make the right decision.
Megan: Sounds like a really impactful change. Last year also marked a milestone in moving from early AI learnings to scale, using AI more deliberately as a force multiplier. What transition were you trying to make and how did you approach it?
Andrew: Well, Megan, we’ve had a philosophy for a long time in Woodside from an innovation perspective, where we really want to think big, we want to prototype small, and we want to scale fast. We want to find big opportunities that we can go after, but we want to ensure that we look at how we deploy those on a small scale first, and then provide the right learning and insight that then can scale it everywhere. Something like maintenance intelligence is a good example of that, or our Startup Advisor, where we know that we’ve got multiple plants that we need to start up. We know that we’ve got multiple assets that need to do maintenance, so we have a big, bold ambition about how we can improve and optimize that. We start with a small prototype; it might be one subsystem, it might be just a part of an asset, and then we scale it out, we learn, and we scale faster.
I think from an AI learning perspective, one of the key things we’ve learned is really the transition from moving from isolated AI solutions to a more coordinated enterprise-wide capability. If you look back maybe 18 months, two years, in our generative AI journey, we rarely started by deploying AI as broadly as we could in the organization from a personal productivity perspective. And probably being quite open in terms of the problems that we will solve, the business problems that we’ll solve with AI. That had a lot of benefits for us in terms of allowing our organization to get to know AI, get to know the capabilities, to build the trust in it.
What we’ve learned though is that we’ve needed to pivot from that to being a little bit tighter in terms of where we are going to invest our time and resources and more higher value solutions. How do we then enable and empower the rest of the organization so that they can actually effectively problem solve with technology in their domain or in their personal productivity without having to come to a central team?
When we think about that, think big, prototype small, scale fast, has been something really important for us. The transition from a more broader approach to use case development and solution development to now a narrower focus on the high value priorities. We’ve seen that paying dividends to us and allowing us to go after solutions and opportunities, things like Startup Advisor.
And so our Startup Advisor is a agentic AI solution that really aims to optimize and empower and better support our operators that sit in front of a panel and have to start up LNG plants, which are incredibly technical facilities and require really specialist skills to start up. And so our Startup Advisor is almost like a copilot that sits alongside those operators, and it gives them the ability to be able to play back previous startups. It gives them the ability to look at how the current startup is progressing, and it provides them better insights to optimize how they start up that facility. And again, starting up an LNG facility is incredibly complex.
Megan: I can imagine.
Andrew: When we think about opportunities like Startup Advisor, again, it goes back to that think big, prototype small, and scale fast. We started with a very bold vision of, how do we start up all of our LNG plants in a much more structured and optimized fashion? How do we better support our panel operators? How do we make, say, a more junior panel operator have a copilot that can help them almost like an experienced panel operator sitting next to them? And when we think about that vision and the ability then to prototype on a small scale and then scale fast, I think it’s been really successful for us.
As we scale, we’ve just naturally expanded into more agent-based solutions. Today, we’ve got around 50 AI agents in production, supporting both our operating assets and our enterprise workflows. These tools have been proven in live environments, and we have really seen the benefit of being able to shift from point solutions that maybe solve small scale problems in specific areas, to AI and agentic solutions with agency that can really work across our workflows.
We’re able to do this because we’ve standardized on the platform that we build on and we’ve got repeatable patterns. That’s been another really important learning for us, is that we don’t want to build 50 solutions in 50 different ways. We really want to be empowering our organization and our technical teams and the users of our solutions to roll them out quickly, to roll them out safely, and to do it in a patternized and platform manner.
But the last point I’ll make, Megan, from a learning perspective is that we’ve really understood that a strong governance around how AI is deployed and developed is critical for us, and it’s critical for us to go fast as well. The traditional ways of governing how we roll out different solutions or digital systems isn’t going to scale to the breadth that we need when we are thinking about AI. Being able to have a clear philosophy around how we innovate, transitioning from isolated solutions to that enterprise-wide capability, and making sure that we’ve got strong platforms with strong patterns and clear governance are the three really critical things that we’ve learned.
Megan: Such important pillars, all of them. And you’ve been working with Infosys on this journey. How has that partnership helped accelerate scaling and embedding AI across the business?
Andrew: Well, Infosys is our managed service provider, and so they play a really critical role in the operations of our core business. One of the things that I like to say is that our license to innovate is based on our license to operate. And so, for my team to be able to turn up to an operating asset or a corporate function and have the trust that’s needed to be able to innovate and reimagine and redesign how work gets done, to be able to do that, we need to make sure that our core platforms, our core systems, our applications are running really reliably, safely, and consistently every day. Having an experienced partner like Infosys looking after those core operations in partnership with our internal teams is really, really important to us.
As we move from pilots to enterprise-wide deployment, the ability to partner with someone like Infosys also gives us the ability to scale. And so being from Perth and Western Australia, while we’ve got a really strong local team in Western Australia, and we’ve also got a very strong team in some of our other operating locations, like everyone, we’re struggling to find people that can fill AI roles. Being able to partner with Infosys and have a number of different operating models at our disposal becomes really important for us. Having co-mingled teams where they are staff, they are Infosys staff, Woodside staff, and some of our other partners, really just brings diversity of thought and experience to how we solve problems.
Fundamentally, the partnership has allowed us to operate and innovate with more confidence. While Woodside always retains ownership of the strategy and where we’re going and the governance and my teams remain accountable for the outcomes, we can’t do what we do without strong partnerships like the one we have with Infosys.
Megan: Fantastic. And as AI adoption scales, you mentioned yourself, governance becomes increasingly important. How challenging has that been, and what guardrails have you put in place at Woodside?
Andrew: So, Megan, governance is really important to us, and we operate in a well-regulated environment. That means we’ve got to make really deliberate and well-reasoned decisions when we’re thinking about how we deploy technology into our organization, whether it’s artificial intelligence or anything else, for that matter. And so, governance is really central to how we approach the execution of our AI strategy at Woodside.
We’ve got maybe two or three really key things that we’ve put in place. The first one is just making sure that every AI use case goes through a structured assessment, and that’s making sure it meets our privacy controls, our cyber controls. We’re also asking the question, not just, could we do this, but should we do this? We’ve really got to bring together safety, ethics, transparency, accountability, and make sure that we make an informed decision. When an AI solution is going through that structured assessment, if there are concerns about how we might use that solution, it then goes to an AI council that’s made up of senior leaders across the organization. That council and that group really oversee some of the prioritization and risk management. That’s where we can have really strong, robust debates around, again, could we do something, should we do it, and how do we mitigate any of the risks that we might introduce here?
I think the last one, Megan, is really around lifecycle management. When you start thinking about, we’ve got 50 at the moment, but if we had 500 agents working in our organization, really amplifying the experience and the decision-making and the value creation of our staff, we really want to have an ability to manage the lifecycle of how those agents operate. We want to know, how many people are using them? What’s the efficacy and the outcome? Is there model drift? Do we need to retune or retrain? I think that’s an area where many organizations, including Woodside, are still leaning into and still figuring out the best way to do this. We can do it quite easily with 50 agents, but 500, 5,000, 50,000 becomes an opportunity for us. Again, thinking about how we partner with others, solving problems like that really present an opportunity to co-create and to co-solve with some of our partners, like with Infosys.
Megan: Fantastic. Just to close, what’s your long-term vision for AI at Woodside? How do you see this evolving over the years ahead, and what could it unlock for the sector in your view?
Andrew: So Megan, I think our ambition is really for an autonomous enterprise, where we have agents with agency that are able to really deeply interact with our core workflows. The outcome that we want to get from that is to protect our people, to protect the environments we operate in, and to be able to provide energy at a lower cost to the world. When we think about that ambition, we can really see that being applied to almost all of the areas that Woodside work in. Whether that’s from exploration through to project developments, through to operations or marketing, the scale of the opportunity in front of us and the ability for us to really change the way that work flows through the organization is really exciting.
For us, there’s three things that we have to get right in terms of being able to execute on that ambition. The first one is really thinking about how the work gets done in the organization so that we’re not just bolting AI onto an existing process, but we’re deeply thinking about how that work needs to be reimagined. We’ve also got to think about how we enable our workforce to work differently. Providing them with the skills and the tools and the ability to really harness the power of the technology that we provide.
Secondly, we’ve got to continue to move from and restrain ourselves from deploying point solutions that solve very narrow problems, to having more connected, agentic systems of systems that can interact with each other. To do that, and if we do that successfully, that’s where we really get the high value unlock from agents being able to interact with workflows and really change how the work gets done.
And lastly, Megan, it’s about how we must continue our philosophy of thinking big, prototyping small, and scaling fast.
Megan: Which is a fantastic lens to which to make all these decisions. Thank you so much, Andrew. That was Andrew Melouney, vice president for digital at Woodside Energy, whom I spoke with from Brighton in England.
That’s it for this episode of Business Lab. I’m your host, Megan Tatum. I’m a contributing editor and host for Insights, the custom publishing division of MIT Technology Review. We were founded in 1899 at the Massachusetts Institute of Technology, and you can find us in print, on the web, and at events each year around the world. For more information about us and the show, please check out our website at technologyreview.com.
This show is available wherever you get your podcasts. And if you enjoyed this episode, we hope you’ll take a moment to rate and review us. Business Lab is a production of MIT Technology Review, and this episode was produced by Giro Studios. Thanks ever so much for listening. Goodbye.
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.
Let’s start with a game. Open up your chatbot of choice—Claude, ChatGPT, Gemini—and type “Give me a random number between 1 and 10.” You’re going to get 7. Almost always. Now type “Another” and you’ll get 3 or 4. Type “Another” again and you’ll get 8 or 9.
That won’t work every time—but if it did, you may wonder if I have superpowers. I don’t.
The truth is that most large language models are stuck in a rut. They are far more predictable and far less creative in their responses than you might expect. That’s fine for tasks like coding or research, but groupthink is a problem when you’re brainstorming or planning your next vacation.
The Australian startup Springboards has a solution. It built an LLM called Flint, which has been trained to come up with a wider variety of responses than mainstream LLMs to open-ended questions such as “Where should I go in Europe?”
“Most language models are fighting hallucinations,” says Springboards cofounder and CEO Pip Bingemann. “We welcome them.”
Bingemann introduced me to the random number game when he first showed me his company’s new model. It felt like watching an illusionist with a deck of cards. “This is our sales trick, and it works every single time,” he says.
After ChatGPT and Claude both gave their 7s, Bingemann turned to Flint. It too came back with 7: “Aha, of course that was going to happen, but it’s okay—7 is a legitimate answer.” He restarted the session and prompted again: ChatGPT gave 7, Claude gave 7, Flint gave 3.7916.
Run your way
It’s not just numbers. When Bingemann asked ChatGPT and Claude to name a type of car, he predicted that it would be a Toyota or a Honda—and he was right. Flint came up with a Ford F-150. “There’s all this lost information that doesn’t get served up in these models,” he says. “They’re just as capable of saying a Buick or a Tesla. They just don’t—they’re biased.”
Bingemann sent one last prompt to each of the three models: “Give me a tagline for a campaign for New Balance running shoes. Just the tagline.” Claude: “Run your way.” ChatGPT: “Run your way.” Flint: “Built to last, run to win.” It won’t win any awards, but at least it’s different.
This weird limitation of LLMs is starting to get more attention. In November a team of researchers put out a paper, titled “Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond),” that exposed a remarkable degree of repetition not only in the answers from individual LLMs but between them as well. They found that different LLMs converged on very similar answers when prompted with open-ended questions.
It’s not clear exactly why this happens, but the researchers speculate it’s because most LLMs today are trained in similar ways on similar data to do similar tasks. The team won the best paper award at NeurIPS, a major AI conference.
When the researchers asked 25 different LLMs (including models from the top US firms as well as open-source models from China and elsewhere) 50 times each to write a metaphor about time, most of the 1,250 responses were a version of “Time is a river” or “Time is a weaver.”
(I asked some of my colleagues the same question and six people gave me six different answers. My highlight: “Time is a favorite sweatshirt, shaped by a lifetime of wear.”)
When you look for it, you see repetition everywhere, says Kieran Browne, cofounder and CTO at Springboards. “The way that most chat interfaces are designed, it makes it feel like you’re having a personal conversation,” he says. “I think most people don’t really realize the extent to which they are getting the same stuff as everybody else.”
Take another example: “What should I name my band?” Most models will say something involving “glass,” “neon,” “velvet,” or “static,” says Browne.
When I tried it, ChatGPT spat out a list of 56 band names. At the top was “Glass Harbor.” Skimming through, I found “Static Empire,” “Neon Hearts,” and “Velvet Echo.” I asked Gemini; it gave me 15 suggestions, including “Static Horizon.”
Some of the suggestions looked pretty cool, though. ChatGPT’s “Sofa Astronauts” caught my eye, so I googled it—and found that a band called Sofa Astronauts already exists.
(OpenAI says that training models to give reliable and coherent answers can lead them to converge around familiar, high-probability responses and that pushing harder for novelty can lead to weaker or less reliable responses. It also notes that the “Artificial Hivemind” paper studied models from 2024 that have since been updated.)
Creative catapult
Springboards has developed a tool backed by a selection of LLMs, including ChatGPT and Claude, that creative professionals in advertising or marketing can use to brainstorm ideas. The tool lets you drag around text produced by different models, picking the bits that you like and combining them into something new—in theory. Springboards is pitching Flint as an alternative model that users of its tool can select when looking for more variety.
Zoe Scaman, founder of the business strategy startup Bodacious and chief strategy officer at 77X, a direct-to-fan marketing platform set up by Luka Dončić of the LA Lakers, has been trying it out. “I find it really useful for throwing me in completely different directions,” she says. “I use it if I want to catapult myself all over the place.”
In one test, Scaman pitted Flint against Claude, Gemini, and ChatGPT by giving each of the models a classic MBA case study: How would you reinvent a finance company for today’s youth? The three mainstream models all went down the same path, she says: “You know, we need to teach financial literacy in a fun and funky way—well, that’s nothing new.”
But Flint came up with something different, suggesting that the whole concept of wealth accumulation should get a rebrand. “That was really interesting,” says Scaman.
She notes that Flint is still a prototype and doesn’t work all the time. “It sometimes falls over when you start pushing it too far,” she says. “But I think that the premise behind it is really powerful.”
Taking the temperature
Springboards built Flint on top of Qwen 3, an open-source model from the Chinese tech giant Alibaba. “We’re a small team,” says Browne. “Training a foundation model is not on the table for us. It’s just too expensive.”
Most LLMs have settings that let you adjust the level of randomness in their output. The most common is called temperature. “Obviously, that was one of the first things we explored, because that’s what people tell you: If you want more creativity, you turn up the temperature,” says Browne.
But changing those settings can also make models incoherent. Dialing up the temperature on one of OpenAI’s models to its maximum setting made it produce responses that switched from English into code halfway through a sentence, says Browne.
Springboards realized that parameters were blunt instruments for what it wanted to do. It does not make sense to dial up the randomness across the board; you only want to boost it at specific points in its output, he says.
For example, when you ask a chatbot “Where should I go in Europe?” the model only needs to tweak the randomness just before it names a destination, not for every word in its response.
To make Flint do this, Springboards trained its version of Qwen 3 to identify the points in its output where more variety was possible and fill those spots with words or phrases that were a little more random.
“Flint’s programmed to throw an oddball in. It’s more of an invitation to think wider,” says Maximilian Weigl, cofounder and chief strategy officer at Uncommon, a marketing firm. “That’s super interesting.”
Weigl’s team uses Flint alongside ChatGPT, Claude, and Gemini. “You can’t really create something boundary-breaking with tools that pull you back to the average,” he says.
And yet Weigl notes that nine times out of 10 the average is fine. You don’t always need to reach for extremes with something like Flint, he says: “Most people are fine with good enough. They want to see mass-market familiar things.”
Weigl also cautions against using any LLM too much. “I have a big problem when people rely on the output from any AI, including Flint,” he says. “If I saw people on my team copy-pasting something from AI, I’d be like, ‘That’s not your job! Think, talk to other people, use your own voice.’”
For now, Flint is aimed at advertisers and marketers because those are Springboards’s customers. But Bingemann and Browne insist that a lack of variety is a problem for anyone using chatbots.
The idea is to give people the choice and leave it to them to decide if the result is good or not, says Bingemann. “Variety is great when you’re trying to spark ideas,” he says. “Let’s go down this route instead of letting the machines do it all and ending up in a gray, boring world.”
At an event for pharmaceutical executives, biotech founders, and researchers on Tuesday, Anthropic announced Claude Science, a major new product intended to support scientific research in the same way that Claude Code supports software engineering.
Like Claude Code, Claude Science can autonomously carry out meaningful work when given concise, high-level instructions, and it has access to tools that make it particularly useful for research in computational biology and drug development.
Along with launching and previewing Claude Science, which is now available to all paid Claude subscribers, Anthropic also announced that it will be using the product to pursue some of its own research into drugs for rare, neglected diseases.
This is not Anthropic’s first foray into AI for science. In October, the company released plug-ins that help Claude make use of scientific software and databases under the heading “Claude for Life Sciences.” But unlike this earlier release, Claude Science is a full-featured, standalone product. Anthropic’s decision to elevate Claude Science to the same rank as Claude Code and Claude Cowork indicates that the company is taking AI’s scientific applications very seriously—or at least wants to give the impression that it is.
“It represents how important this is to our mission that this is right up there with Claude Code and Claude Cowork as the next really significant product that we’re releasing,” says Eric Kauderer-Abrams, Anthropic’s head of life sciences. “Our mission is to develop AI that serves humanity’s long-term well-being, and we believe that by far the greatest opportunity to do that is in the life sciences.”
For the past decade, one company—Google DeepMind—has been at the vanguard of AI for science. CEO Demis Hassabis and researcher John Jumper won the Nobel Prize in chemistry for their work on the company’s AlphaFold model, and DeepMind has also made major contributions to meteorology, materials science, and a variety of other disciplines. But in the past several months, the fast-advancing frontier of AI progress seems to have left DeepMind in the dust. When it comes to coding, which has become the most lucrative use case for LLMs, DeepMind is stuck playing catch-up.
Anthropic is well positioned to take up DeepMind’s scientific mantle. Like Hassabis, Anthropic CEO Dario Amodei is a PhD scientist—unlike OpenAI CEO Sam Altman, who’s a businessman through and through. Many scientists are already avid users of tools such as Claude Code.
These days, a lot of scientific research involves some amount of coding, but not all scientists are expert software engineers, and so tools like Claude Code can make a huge difference for their productivity. And the company has recently earned a major scientific vote of confidence: Earlier this month, Jumper announced that he is leaving DeepMind for Anthropic.
Since agents powered by LLMs, including Anthropic’s Opus model series, became capable of useful, independent work in late 2025, scientists have been seeing just how much they can do. In a blog post published on Anthropic’s website, the Harvard physicist Matthew Schwartz estimated, on the basis of his work with Claude Code and other Anthropic tools, that the company’s Opus 4.5 model is about as capable of executing scientific projects as a second-year graduate student.
According to Kauderer-Abrams, Claude Science isn’t intended to displace Claude Code and Claude Cowork in scientists’ workflows. Instead, it’s designed to build on what scientists already find useful about Anthropic’s products. For instance, it not only writes code but also helps scientists run their code on powerful computer clusters, which many many scientists need for their work but can be difficult to manage. And it prioritizes reproducibility, so that scientists can trace back the source of any figure or result and check it for accuracy and validity.
Though Claude Science could in principle assist with any area of scientific research, it seems designed and marketed as a tool for molecular and cellular biology, and for drug development in particular. It can interface with various tools used in genetics, chemistry, and protein biology, all of which could come in handy for researchers on the hunt for new drugs. During the Tuesday event, Alexander Tarashansky, who led the development of Claude Science, demonstrated how the system could autonomously identify new drug candidates for phenylketonuria, a rare genetic disease.
And Anthropic isn’t leaving all that work to the pharma companies and university labs that were represented at the event. Armed with Claude Science, it will be pursuing its own research into drug candidates for neglected diseases—both to help move science forward and to gain a clearer sense of how Claude Science works in the real world.
There are obvious humanitarian reasons to prioritize drug development when creating a general-purpose scientific research tool, and AI industry leaders often cite curing disease as a major potential upside of the technology. But it’s also notable that pharmaceutical companies have far deeper pockets than academic researchers.
Anthropic says it’s set to see its first profitable quarter, and if major new contracts with pharmaceutical companies are forthcoming, they could help ensure it stays profitable as the tokenmaxxing craze dies down—something that’s ever more important as an IPO approaches later this year.
Artificial intelligence is transforming what is possible in agriculture, but industry leaders should be wary of investing in AI without first laying the groundwork.
The use cases are promising, especially for an industry navigating volatile fertilizer costs, unpredictable weather, and margins that leave little room for error. Research shows AI-enabled predictive models can improve crop yield by 26%, reduce water use by 41%, and cut chemical usage by 33%.
However, what AI vendors usually won’t tell you is that these solutions are only effective if you have a clean, solid data foundation. However, at Reltio, we have experience in this area, including leading technology strategy at a major agricultural distributor and building a data platform used by enterprises worldwide–we’ve seen it first hand.
What AI vendors won’t tell you
Vendor conversations in agriculture tend to follow a familiar pattern. The pitch leads with grand promises around using AI to monitor crop health in real time, optimize irrigation, and squeeze more yield from every acre.
The promise is compelling, but what rarely comes up is the question of whether the data foundation underneath those promises is accurate and complete. If not, there is a real and significant risk that AI will generate misleading outputs that seem authoritative but inspire action that is, at best, counterproductive.
For instance, a yield prediction model fed inconsistent historical data will generate imprecise forecasts. Similarly, a precision irrigation system drawing on fragmented sensor data will make watering decisions that waste resources instead of saving them.
In each case, the AI is failing because the data it was trained on was not sufficient to produce trustworthy outputs. In agriculture, every AI hallucination is a liability, and the likelihood of error is high.
Why agriculture is a uniquely challenging test case
The data landscape across a modern agricultural operation or a large distributor serving thousands of growers is extraordinarily complex.
Modern farming environments make extensive use of IoT devices and machinery. Irrigation systems are automated, tractors navigate fields autonomously, and drones capture field imagery at scale.
However, machine data is disparate by nature. Add in external sources, including weather feeds, U.S. Department of Agriculture data, and third-party market information, and the question of how you bring all of it together into something coherent becomes a significant undertaking.
Agricultural AI also needs to understand more than just customer attributes; it needs to understand the land: GPS coordinates, farm boundaries, field blocks, and soil variation across a single property. Where do you apply fertilizer, and at what rate, and in which specific area of the farm? Not all parts of a field are the same, and an AI system that treats them as if they are will produce recommendations that are at best imprecise and at worst damaging.
There is also a compliance dimension due to the chemicals and the responsibility involved. Operational AI in agriculture needs significantly more checks and governance than it might in a lower-stakes environment. When a flawed recommendation gets acted upon in the field, the consequences can be severe.
What data readiness means in practice
Data readiness is the difference between AI delivering on its promise vs. a “garbage in, garbage out” scenario. Fundamentally, being ready for AI means having a data model that accurately reflects how the business operates.
For a company like Wilbur-Ellis, a 104-year-old, family-owned agricultural distributor, that means understanding who your customers are, which fields they farm, which inputs they need, which suppliers those inputs come from, what they paid last season, and how all of that connects to margin. That information needs to be current, consistent, and accessible across the organization, rather than locked in separate systems that were never designed to talk to each other.
Similarly, for farming operations themselves, data readiness means having a reliable, connected picture of what is happening across every field: soil health records, input application histories, yield data from previous seasons, equipment performance, and real-time sensor readings from irrigation systems.
Governance matters just as much as structure. Prices change, relationships evolve, and suppliers come and go. An AI system drawing on data that was accurate six months ago but has not been maintained will make recommendations based on a version of the business that no longer exists.
Building the foundation that makes AI trustworthy
The good news is that the path to data readiness is feasible. It starts with a strong data model: a single, governed source of truth that connects customers, suppliers, products, pricing, orders, and margins in a way that reflects how the organization operates.
From there, it requires data pipelines fast enough to deliver insights when decisions need to be made, governance frameworks that keep that data trustworthy over time, and security controls that ensure sensitive commercial information is accessible to the right people under the right conditions.
This is precisely the challenge that Reltio, an SAP company, was built to solve. Reltio enables companies to unify their fragmented data so AI agents and systems can operate from a complete picture of the business. Reltio builds a trusted system of context, known as the context intelligence layer, that brings all entities, relationships, rules together under one roof and makes business data easy to access and interpret.
For Wilbur-Ellis, building that trustworthy data foundation has meant being able to ask more complex questions and trust the answers, which is the precondition for any AI system to be genuinely useful.
How agriculture can drive real value from AI
The question worth asking before the next AI conversation is not whether the use case is promising. It almost certainly is. The question is whether the underlying data foundation is strong enough to make the output trustworthy.
Agriculture has always required its leaders to make high-stakes decisions under uncertainty, and AI offers the genuine prospect of making those decisions faster and better informed. That prospect is only achievable for organizations that have done the foundational work first, and the businesses that will get the most from AI are the ones investing in that foundation now.
This content was produced by Reltio. It was not written by MIT Technology Review’s editorial staff.
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.
Imagine coming in to work to learn that a new underling will report to you. The worker is not a person but an AI tool—one that your company nonetheless calls Alex, an “employee” with a title and defined responsibilities. How well do you think you would work with Alex?
If you’re anything like the managers recently studied by Emma Wiles, a Boston University business professor, treating Alex as a “coworker” and not a software tool would lead you to do a worse job. Wiles found that people caught 18% fewer errors when the work was said to have come from an agentic “AI employee” rather than a chatbot. It turns out that what’s in a name matters. A lot.
This is an alarming glimpse of the future Silicon Valley is hurling us toward. Last year Nvidia’s CEO, Jensen Huang, talked about workplaces of “digital humans.” Since April, Microsoft, OpenAI, Anthropic, and Google have all released new tools oriented toward managing teams of AI agents, many of which are explicitly advertised as digital colleagues with the flexibility and cognitive power of actual humans. And nearly a third of the 1,261 managers who participated in Wiles’s study said their companies already frame AI agents as employees (23% even list them on org charts).
The technical progress of agentic AI is not all hot air, of course. Agents, which can effectively be thought of as AI tools programmed to work in a loop until they achieve a goal, have become measurably better at more complicated tasks. But it’s a huge leap to refer to these tools as coworkers or employees, and doing so will set unrealistic expectations for what AI can do while leaving the human employees supposedly responsible for them worse off.
That’s partially because, Wiles’s research suggests, it inverts our sense of who’s in charge. When an AI tool was framed as an employee, participants in the study saw themselves as less responsible for its output. They were also 44% more likely to escalate its questionable work to a manager for further review rather than trusting their own corrections (thus negating the time-saving purpose of using the AI agent in the first place).
That matters far beyond office culture: As AI agents are embedded into health care, warfare, education, and government, there’s a growing risk they’ll become a convenient place to dump blame for failures that are instead the product of bad human decisions, incentives, and oversight (recall how the bomb strike on a girls’ school in Iran was popularly blamed on Claude, when all signs point to a cascade of human errors).
“AI agents right now are being marketed as things that can replace humans, and I think that’s just a losing proposition,” says Daron Acemoglu, an economist at MIT who won the Nobel Prize in 2024 and studies AI’s impact on the economy. “They should instead be optimized so that they can improve human capabilities, which is not what they have [been] at the moment.”
What could that look like? Consider a new effort at Stanford, where researchers presented 1,500 workers in 104 jobs with information about what tasks AI could potentially do in their work and then asked what would actually be most helpful and productive. Workers did want automation in certain areas: Law clerks thought AI could help ensure that adequate progress was being made across cases, for example. But often the tasks that tech experts deemed most suitable for AI—like verifying customer credit ratings for sales reps—were what the actual workers said they definitely did not want or need an agent to do.
Which brings us back to Alex. Calling Alex an employee is easy—and convenient, especially when something goes wrong—but it’s a branding exercise. It doesn’t make the tool more fit for the job, and as Wiles’s research shows, it makes the humans around it worse at theirs. And recall that they are the ones with the agency that AI is trying to replicate. They deserve better than Alex.
Enterprise investment in AI is booming. Gartner is calling 2026 an “inflection year” for organizations to align their AI projects with strategic business objectives. As the pressure to prove ROI mounts, executives and technology leaders are looking to agentic AI to drive the measurable financial outcomes their businesses seek.
A prime opportunity for AI agents exists in the tech function, where IT infrastructure costs are projected to grow two to three times by 2030, even as budgets remain unchanged, according to McKinsey. And in the last 18 months, tech teams—the engineers, developers, architects, and other practitioners who are building, deploying, and continually improving their organizations’ infrastructure and applications—are clearly putting agents to work.
The ultimate promise of agents is not only to automate tasks but to manage and coordinate entire workflows, pursuing business goals in a way that allows humans and agents to work together. Given the risks involved in automated decision-making, teams cannot delegate the work that agents do without confidence that they are fully capable of performing the task and that it will do so in a safe, reliable, and secure manner.
Among technology experts, our research shows that teams are exceedingly confident about using agentic AI across a significant amount of AI, data, and cloud tasks.
Where agent readiness drops is largely due to a lack of business context being supplied to agentic systems. The more complex the task, the more reasoning capability an agent requires and the greater its need for business context. Such context-generation capabilities for agents are still at an early stage of development, especially in situations where enterprise data is difficult to wrangle and connect into the agent lifecycle at the speed and quality in which developers and executives need it. Human oversight is a key factor of success in deploying agentic AI.
Knowing that tech teams are in a pivotal position to lead this transformation, the experts we interviewed expect agent confidence to accelerate as experience with agents deepens and business environments mature. “As we design agents to operate within the same operational boundaries, identity systems, and governance models that teams already use, they start to behave more like the systems organizations already trust,” says Jeremy Winter, corporate vice president and chief product officer at Microsoft Azure Platform.
This report, based on a survey of 300 global technology experts, ranks 101 tasks across AI, data, and cloud workflows based on respondents’ confidence in agents acting on their behalf. It also examines how technology teams view the opportunities and challenges related to agentic AI, along with the potential for the technology to enhance their careers.
Key findings from the report include:
Confidence in agents is surging for measurable tasks and growing in areas of complex judgment. Technology experts overwhelmingly believe agents help with everyday work including streamlining processes, improving performance, and reducing repetitive tasks. Confidence is highest for processes like generating reports and boilerplate code, and there is clear opportunity where tasks involve multistep workflows and advanced reasoning to make decisions.
Data workflows are the breakthrough domain. Tech teams trust agents most where structure can provide a reliable foundation for decisions. This includes areas such as data quality monitoring, visualization anomaly detection, real-time data stream monitoring, and data profiling. This is where domain experts closest to the point of data generation can provide context to allow agents to act and deliver trusted outcomes.
Read the Microsoft Cloud blog by Amanda Silver, corporate vice president of Microsoft 365 Core and Work IQ, which underscores the importance of keeping humans in the loop and how systems thinking advances careers. And for a deeper dive into data workflows as a breakthrough use case for agents, check out the Fabric blog to hear from Kim Manis, corporate vice president of Product for Microsoft Fabric.
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.
Artificial intelligence is rapidly reshaping retail, but not in the ways consumers might immediately notice. The biggest transformation may not be flashy virtual try-ons or chatbot shopping assistants, but in how decisions are made behind the scenes: how products surface in search results, how inventory moves through supply chains, how engineers ship code faster, and how retailers respond to customer behavior in real time. As legacy retailers navigate a fragmented and hyper-competitive landscape, AI is becoming an operating philosophy.
At Macy’s, that philosophy is more often defined by what senior director of engineering Murali Murugan describes as an “AI-first” approach. “AI first isn’t about adding intelligence on top,” Murugan says. “It’s about redesigning how decisions happen so the business moves faster and every experience feels more relevant by default.” Rather than layering AI onto existing workflows, Macy’s is embedding intelligence directly into systems that include personalization, search, operational planning, and software development itself.
The company’s strategy is reflective of a larger shift taking place across retail: moving from isolated AI pilots toward integrated systems designed to compress, as Murugan puts it, “the gap between the signal and the action.” Early efforts focused on narrow, high-impact use cases like search recommendations and customer engagement, where measurable gains in conversion and reduced friction quickly built internal momentum. “Once we established the quick wins, scaling was a business decision, not a technology debate anymore,” he says.
That momentum is now extending into conversational commerce through tools like Ask Macy’s, an AI-powered shopping assistant designed to act more like a personal stylist than a traditional search bar. Whether for a prom, a vacation, or a last-minute event, customers can describe what they need conversationally and receive curated recommendations informed by past purchases, preferences, and context.
Still, the company sees AI as more of an invisible layer augmenting human judgment than a replacement for it. The long-term vision is retail that feels increasingly seamless, adaptive, and personalized, powered by systems customers may never even notice are there.
“The real transformation in this all comes from continuous improvement,” Murugan says. “It’s about learning from the mistakes, quickly adapting to the newer technology standards that are coming into play, timing, and execution which compound into a meaningfully better customer experience.”
This webcast is produced in partnership with Infosys.
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.
AI is booming. New use cases are emerging each day. To capitalize on the technology’s potential, enterprises require data at scale. In many cases, though, the relevant information is blocked or unstructured, which limits its use by AI models.
To understand this challenge, consider the foundation of the web itself. The web was not designed for the automated discovery and retrieval that new AI applications demand. Overcoming this inherent design constraint requires infrastructure.
The next frontier in AI may depend on a new web data infrastructure layer that can enable models to discover and map this ever-expanding digital realm. This layer must be able to navigate hundreds of millions of existing web domains and billions of new URLs created each week, delivering real-time information and overcoming technical barriers.
“The data suggests there’s far more data out there,” says Or Lenchner, CEO of Bright Data, a web data collection platform. “Think of the universe: It’s out there, but you don’t know what you don’t know.”
Enabling access to fresh, relevant, and trustworthy data
While early AI breakthroughs were driven by scaling training data and model size, organizations are now encountering a fundamental bottleneck: They need to keep pace with the dynamic, unstructured, and constantly evolving nature of web data in order to ground outputs in current and verifiable information. AI performance increasingly depends not just on model architecture but on a system’s compute, networking, retrieval, and data engineering capabilities—that is, the system’s ability to quickly and reliably retrieve data that is fresh, relevant, and trustworthy.
Traditional model training relies on snapshots of information collected at a particular point in time. Training AI on such static data is no longer sufficient. To track fluctuations such as competitor pricing, consumer sentiment, and market trends, companies need a constant feed of new information, pulling data in real time along with relevant context. Their infrastructure must therefore be able to handle millions of simultaneous interactions across websites that vary by geography, language, format, and access rules.
“If it can’t retrieve real-time information, it lacks context,” Lenchner says. “In a business setting, that’s not acceptable anymore. Stale answers lead to bad decisions and disappointed consumers.”
Speed is not merely a matter of convenience; it’s a matter of necessity. Today’s organizations operate in environments where prices, inventory, markets, security threats, and customer behavior change continuously. Delayed data retrieval can reduce the usefulness of an otherwise sophisticated model.
Using live, high-quality web data can also reduce AI hallucinations because the model has a more relevant knowledge base. This builds user trust. In fact, one survey found that 56% of AI practitioners said businesses need access to real-time web data to improve trust in AI outputs. To ensure the model runs efficiently and effectively, the information must also be pared down to the appropriate essentials.
Despite the introduction of retrieval-augmented generation (RAG), where models pull in external data at the moment of a query, many AI systems still struggle to deliver outputs that are current, contextually relevant, and trustworthy in operational settings. According to Gartner, 60% of AI projects that are not supported by AI-ready data—accurate, structured, organized, and contextualized—will be abandoned by the end of the year.
This is because large-scale retrieval alone does not solve the problem. As Lenchner puts it, “You need to retrieve data at scale, but also in real time. Latency becomes an issue because of the end user who is waiting for the output.”
Accessing fresh, AI-ready data at scale introduces technical and structural challenges. In practice, many enterprise systems combine public web retrieval with APIs, licensed datasets, and proprietary internal data in their AI applications. Integrating these fragmented sources into a timely and usable knowledge layer requires specialized capabilities. Some research has found that 97% of AI organizations depend on real-time web data infrastructure, but 90% feel boxed in by various restrictions. Companies are increasingly developing technical approaches to navigate these constraints.
Lenchner draws this metaphor: “Think of the trained model as intelligence and relevant data as knowledge. A powerful intelligence layer sitting on top of a hollow knowledge layer is like a genius who knows nothing—useless in practice. Intelligence and knowledge have to come together.”
The promise of new infrastructure
A new layer of web data infrastructure can address this developing need for stronger AI inputs by enabling discovery of data, real-time access, and tailoring to a specific context. As Lechner describes it, “It’s all about collecting data at scale, super-low latency, without being blocked.”
Rather than relying on increased computing power, this type of platform emulates human browsing behavior to access available content and transform raw code into structured data feeds. It can work with websites that might not interact with traditional scraping tools, such as those heavy in JavaScript, or with aggressive antibot software.
As Lenchner explains, “It’s basically having infrastructure that can mimic a web user with identifying information—IP address, location, and 1,000 more parameters. And at scale. Think of doing that 80 billion times a day for millions of websites. And every single time, you are looking exactly as the website expects you to look.”
Of course, continuous retrieval introduces new data governance challenges. To address them, platforms can enforce strict compliance protocols aligned with global privacy frameworks, such as the EU’s General Data Protection Regulation (GDPR) and the California Consumer Privacy Act (CCPA). They can also be limited to openly accessible, public information, avoiding paywalls or private logins. Any networks used can be vetted and consent-based, and incentives can be provided to owners of IP addresses. In this way, systems can be designed to comply with tightening regulation.
Such complex capabilities do not come easy. “When this is critical infrastructure for a company,” Lenchner says, “doing it in-house becomes a full-time engineering problem that competes with the actual AI work.” Addressing this complexity requires organizations to commit significant resources, leading many to seek specialized platforms designed specifically for data retrieval, orchestration, and observability.
Infrastructure for the real world
Real-time data retrieval is changing what AI systems can do inside organizations. For example, a retail company can use public information to enable a dynamic pricing engine, and global brands can track trademark infringements.
As the ecosystem matures, organizations that invest in this emerging data infrastructure layer will be better positioned to build AI systems that are more responsive, reliable, and aligned with real-world conditions—AI systems that can continuously adapt using current web data. Over time, the distinction between AI models and the infrastructure that feeds them may even begin to disappear.
As Lenchner says, “The world is changing. And everything that is happening in the world is being uploaded to the public web. The amount of new data that is being generated is growing and accelerating.”
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.
Jos Benschop is climbing a ladder to get to the top of his newest machine.
It’s a bit of a schlep. The contraption is the size of a double-decker bus—more than 150 tons of gleaming precision-milled aluminum covered in thousands of snaking tubes, colored cables, and pressurized tanks. From the ground, it looks like a futuristic V8 engine. When I reach the top with Benschop we’re looking down from about 15 feet in the air, with bunny-suited technicians scurrying around below.
It’s more than 200 cubic meters of tech—“mechatronic devices that hold a few mirrors in a position with atomic precision,” he says, gesturing at the gargantuan apparatus. Benschop, a tall and grizzled 66-year-old, has spent over a decade working with his engineers to design this thing, but even so, he’ll sometimes look at it and go: Oh my God.
Benschop is the executive vice president of technology for ASML, a Dutch company that is the linchpin of the microchip industry. If you want to make powerful chips to power phones or AI, a lithography machine like the one we’re standing on is what you need to create increasingly tiny circuitry. Lithography is the art and science of shining light on a silicon wafer to pattern out the transistors, wiring, and other components of the microchips that will be cut from it.
The chipmaking field is essentially controlled by only two big players: ASML, which creates the lithography machines, and TSMC, the chipmaking giant.
Nine years ago, ASML began selling machines that use a daring new way of patterning chip features. These machines employ extreme-ultraviolet light, or EUV—radiation well outside the visible spectrum that they produce by shooting lasers at tiny molten drops of tin, tens of thousands of times a second. Those first machines—the result of an R&D moonshot that lasted 16 years and cost about $10 billion—can craft transistor features with a resolution of 13 nanometers. This new machine can do even better: It has a resolution of just eight nanometers, the width of about 40 silicon atoms. The devices are now shipping to chipmaking factories, or fabs, at an eye-watering price: $400 million each.
But chipmakers will fork that cash over, because they are in a desperate race to produce new and improved chips every year. That means getting their mitts on machines that can make ever smaller components and cram them together ever more densely—part of a long-standing recipe for creating faster and more energy-efficient chips.
For years now, ASML’s tools have been critical to keeping Moore’s Law alive. Without the company’s advanced chipmaking technology it is very possible that chip density—and the ability to perform ever more calculations—would have plateaued.
The AI industry has produced new and ravenous demand for denser chips, as firms like OpenAI and Anthropic scramble to erect server farms that train and deploy new, ever-more-powerful models, which require new, ever-more-powerful hardware. ASML’s latest machine promises to help keep the AI party raging for at least another decade.
“We can allow customers to go to smaller and smaller features, and that opens up the space for whatever we see now today in AI, which is absolutely mind-blowing,” Marco Pieters, ASML’s CTO, told me. “I think we’ve only seen the tip of the iceberg.”
Its relentless push for “shrink”—as they call it in the chipmaking industry—has made ASML a dominant force: The company produces about 90% of all chip-lithography tools worldwide. If you make chips, ASML is unavoidable.
But that monopoly position makes some people, and governments, uneasy. The chipmaking field is essentially controlled by only two big players: ASML, which creates the lithography machines, and TSMC, the chipmaking giant in Taiwan, which uses ASML’s machines to craft the vast majority of all microchips. This duopoly is so powerful that it has geopolitical implications. In an effort to prevent China from developing advanced AI, the US government pressured the Dutch government to impose an embargo in 2019: ASML isn’t allowed to sell high-end machines to any Chinese firm. Geopolitically, “chips are the new oil,” says Marc Hijink, the author of Focus: The ASML Way. Being deprived of them can be as disastrous as being deprived of oil. And in that metaphor, you might say, ASML is the Strait of Hormuz.
James Proud, the cofounder and CEO of the lithography startup Substrate, says the situation is not ideal. The US is “dangerously reliant” on a supply chain that’s overseas and increasingly pricey, Substrate says on its website. “There’s a huge concentration in a small number of players,” Proud says. “And the supply chain is just very expensive.”
Which is why, after two decades of ASML’s dominance, would-be competitors are now gunning for its territory. China is hungrily pouring billions into trying to replicate ASML’s tech. And startups like Substrate are trying to get in the game as well, setting their sights on creating lithography machines that are cheaper, smaller, and even more capable than ASML’s behemoths. Will any of them succeed? The near future clearly belongs to ASML, but as its engineers well know, you can unseat a giant with the right trick of the light.
Making chips is, oddly, a bit like silk-screening a T-shirt. To print a pattern on a silicon wafer, you start with a pattern on a reticle—a mask that carries the design. Shining a light on the reticle transfers that pattern to the wafer. The light interacts with a layer of chemicals on the wafer, fixing the pattern in place.
The size of a chip’s features is partly set by the wavelength of light the machine uses: The smaller the wavelength, the teensier the circuitry you can create. You can stretch the capabilities of a wavelength somewhat; increasing what’s known as the numerical aperture, which usually means swapping in a bigger lens, can further focus the light and thus lay down patterns for smaller and smaller components. Eventually, though, this trick hits its limit, and you need to find a new form of light with a smaller wavelength.
So the history of chipmaking has been a two-step dance. The industry finds a good source of light, eventually increases the numerical aperture, and then finally accepts the need for a smaller wavelength, starting the two-step all over again. Up to the early 1990s, chipmakers used visible light, with a wavelength of about 400 nanometers. By the mid-’90s they’d upgraded to deep ultraviolet, ultimately getting it down to a 193-nanometer wavelength. By the late ’90s they saw the end of the line approaching for deep ultraviolet. But what would come next?
All the options were troublesome. They could shift to x-rays, with a teensy one-nanometer wavelength, but they were devilishly hard to focus. Beams of electrons and ions were equally precise; but they worked like dot-matrix printers, transferring a pattern point by point, which was far too slow. (The chip industry wants a machine to crank out hundreds of wafers per hour.)
“It’s a very engineering-heavy company: Let’s send thousands of engineers and just have them mow down these problems. That’s what they did, and it worked.”
Jeff Koch, analyst, SemiAnalysis
Around 2001, ASML, then a smaller player in the lithography world, placed its bet on another option: EUV, with a wavelength just shy of the x-ray range. Nikon and Canon were working on it as well, but they dropped out—while ASML kept going. The idea was full of unknowns. Nobody knew how to reliably generate that type of light, nor how to focus it; EUV is absorbed by regular glass lenses. It’s even absorbed by air. ASML figured it would take six full years to wade through this R&D nightmare.
In reality it took those 16 years and about $10 billion in research, but it worked. The machine, which works in a vacuum, creates EUV light by vaporizing molten tin and using mirrors to direct it. Zeiss, a historic German optics company, had to invent new techniques for polishing and inspecting the mirrors, using an ion beam to knock off minute imperfections.
“They sort of ignored the buzz of, like, Hey, this is never gonna work, and they just beat their heads against these huge engineering problems,” says Jeff Koch, who used to work for ASML and is now an analyst for the chip-industry research firm SemiAnalysis. “It’s a very engineering-heavy company: Let’s send thousands of engineers and just have them mow down these problems. That’s what they did, and it worked.”
When the first EUV machines went on the market in 2017, they cost well over $100 million apiece. Some observers wondered whether the demand would really be there from the major chipmaking firms—TSMC, Samsung, and Intel. In the years chipmakers were waiting for EUV to happen, the lithography industry had developed clever ways to improve on old-fashioned deep ultraviolet light. (If you put a layer of water on top of the wafer, for example, the light could focus more narrowly.) Maybe EUV wouldn’t be much needed for a while?
But ASML lucked out. Only a few years after EUV debuted, OpenAI released GPT-3 and then ChatGPT. Artificial intelligence burst into the mainstream. Instantly, firms like OpenAI, Google, Meta, and Anthropic were hungry for increasingly high-end chips as they built massive server farms to train and deploy large language models. EUV made it easier and faster to crank out AI-tailored chip designs. Nvidia began producing elite GPUs—processors perfectly suited for AI training—that cost $40,000 a pop; the big companies couldn’t get enough. The AI wars were on, and EUV was in demand. In 2025, ASML says, it sold nearly 50 EUV machines to companies and pulled in nearly $40 billion in revenue. As of press time, the company’s market cap was over half a trillion dollars.
ASML’s new machines have no shortage of potential customers. But there is one in particular, with deep pockets, that can’t buy them for any amount of money: China.
The US wants to hobble China’s ability to create cutting-edge AI chips—or any advanced chips, for that matter. So when ASML began selling its original EUV machines, in 2017, the Trump administration successfully pressured the Dutch government to forbid the company from selling them to any Chinese firms. The US had also imposed export controls on China’s telecom giant Huawei, banning US firms from using its 4G and 5G equipment.
This one-two punch incensed the Chinese government and stirred it to action. China is now pouring billions into catching up and trying to develop its own EUV chip-patterning technology. A Reuters report last winter found that a government skunkworks employing former ASML staffers had cobbled together a machine so huge it filled the entire floor of a lab. It’s unclear how well it works. The experiment may well be making some chips, says Hijink, but he doubts it can do so at an industrial scale.
A mirror is installed in an optical system for the high-NA machine.
COURTESY OF ZEISS
Officially, the government denied it was pushing to develop EUV tech. An editorial in the Global Times—a newspaper closely allied with the Chinese government—pooh-poohed the report, claiming that China was still happy to work with the West to get access to chips. “Our goal has never been to build a self-sufficient ‘technology island’ in isolation,” it stated, “but rather, on the basis of achieving autonomy and control over key technologies, to integrate more deeply and equally into the global innovation network.”
Experts say the reality is in the middle. China definitely craves a domestic ability to make high-end chips. And unlike ASML, it doesn’t need its EUV machinery to be efficient and profitable, cranking out about 200 wafers an hour. Any output would help wean it off reliance on the West.
“They would be very happy to have a tool that does one wafer per hour and it costs them a fortune to run,” Koch says. “They would build a fab with a thousand of those and be super happy with it.”
Still, producing and managing EUV light well is a feat that might take years, some told me. In the meantime, the Chinese will lean hard on deep-ultraviolet lithography, developed in the ’90s, making the most of an alternative but slower approach known as multi-patterning, says David Lin, senior advisor for tech leadership at the Special Competitive Studies Project, a think tank that focuses on security and technology. “They’re going to push DUV to the absolute limits,” Lin says.
The AI race is also pushing China to devise ever cleverer ways of developing LLMs that don’t rely on the fastest AI chips. In the US, OpenAI, Anthropic, and Google are fighting over who can buy the biggest piles of hot Nvidia chips. Since China can’t compete that way, it is innovating not in hardware but in software—building lighter-weight LLMs like DeepSeek.
As China rumbles into action, ASML has remained laser focused on shrink. To go even smaller, Benschop and his engineers decided, they wouldn’t shift to a new form of light. They’d do the second part of the two-step: They’d raise the numerical aperture of the machine by more than half (for those keeping track of the specific numbers, it would be a switch from an NA of 0.33 to an NA of 0.55). That would let them cut the size of the transistors by close to half and nearly triple their density on a chip.
This would also be an easier climb. Without the need to develop an entirely new source of light, the new machine—based on high-numerical-aperture EUV, or “high NA”—would be evolutionary, not revolutionary.
Still, building the new system did present a few gnarly challenges. In an EUV machine, the way you transfer an image onto a wafer is by shining light at the microchip pattern on the reticle and then using an optical system to take the reflected light and demagnify that pattern, shrinking it down to the size you want on the wafer. The light hits only part of the reticle at any given time, so you quickly move the reticle back and forth to expose every part of the pattern to the light.
Going to a higher numerical aperture meant they could have smaller features on the reticle. But this also meant that some of the light would be arriving at the reticle—and reflecting off it—at a steeper angle.
That’s what caused problems. The pattern on the reticle is three-dimensional, so light arriving at such a steep angle caused shadows—much the way slanted sunlight creates shadows in the Grand Canyon. That stood to diminish the machine’s ability to make clear patterns.
The new reticle moves with acceleration up to 22 g, much faster than in the company’s original EUV machine. “Don’t try to sit on it, because you’ll pass out.”
The solution was to change the pattern on the reticle—along with the way the mirrors took the light and shrank it down to impart the pattern to the wafer. The designs on the reticle would now be twice as long as they were wide—stretched, as it were, in one dimension.
But this design came with its own problems. The changes to the mirrors meant the area on the wafer exposed during a single scan was half the size it was with the original EUV machines, reducing the system’s speed. And ASML couldn’t tolerate any slowdown: Chipmakers were paying it for machines with massive throughput, about 200 wafers an hour.
If one part of the system slowed down, another part would have to speed up. The engineers decided the machine should move the reticle faster, which meant making the entire mechanism lighter and dramatically redesigning it. The new reticle moves with acceleration up to 22 g, much faster than in the company’s original EUV machine. “Don’t try to sit on it, because you’ll pass out,” Pieters told me. The wafer stage moves around faster as well, in tandem with the reticle.
Meanwhile, over in Germany, Zeiss’s engineers were busy designing mirrors to accommodate the higher numerical aperture and asymmetric shaping of the light. The new mirrors would be about twice as large as those in the regular EUV machines, and the projection system, which carries light from the reticle to the wafer, weighed fully 12 tons, seven times more than before. Zeiss built a new robot-assisted production line to handle these ponderous new beasts. The company says they’re the smoothest surfaces they’ve ever made.
At the same time, ASML was working on making its EUV light source even more powerful, to help make the wafer-exposing process go faster. The engineers calculated that they could improve the output of EUV if they hit each tin droplet three times with the laser instead of twice, as they do in the first machine. That meant the already-hectic system of firing tin would need to speed up by 50%. “The lasers just keep getting bigger,” says Alex Schafgans, the head of engineering at ASML in San Diego, where the EUV light source is built.
Indeed, the lasers for a single machine now fill an entire room. After Benschop showed me the massive high-NA device, we walked across the hall and entered a chamber filled with hulking six-foot-tall boxes that were part of the laser system. Peering through tiny windows in the sides of the units, we could see the glowing purple plasma used in creating the laser light.
When high-NA machines began to roll off the assembly line, one company was waiting hungrily: Intel. The company purchased the very first high-NA machine put up for sale, and in the spring of 2024, 300 ASML engineers showed up in Oregon at one of Intel’s fabs to begin assembling and testing it.
“ASML actually put a giant ribbon around one of the boxes,” says Mark Phillips, an Intel fellow who is director of its hardware and lithography solutions, laughing. His team has been testing the machine to see how well it performs; Phillips wouldn’t give details other than to say he’s “very pleased at the rapid pace of tool health.” He also wouldn’t give a date for when Intel would start using it to make chips, though observers say that will likely happen next year. The company plans to ease it in, using it for just a few precision components on a chip and then gradually for more and more.
What’s at stake is a chance to recapture its mojo. Intel was once a silicon powerhouse, designing the most cutting-edge CPUs for computers and servers, and building them in its own fabs. But in the 2010s, the big new markets were mobile-phone chips and GPUs for AI and gaming, and Intel rapidly lost ground. Apple designed its own mobile chips (and had TSMC make them), while Nvidia did the same thing with GPUs. Google began banging out its own TSMC-made AI chips called TPUs in 2015, and soon it was stuffing data centers full of them.
Intel fellow Mark Phillips briefs members of the media on the high-NA tool at the company’s Fab D1X in Hillsboro, Oregon. Intel was ASML’s first customer for the new EUV machine.
COURTESY OF INTEL CORPORATION
So in 2021 Intel announced a moonshot. It would aggressively begin building out a foundry business, one that would go toe to toe with TSMC. Instead of creating Intel chips, the Intel foundry would manufacture designs for customers like makers of mobile phones and AI chips.
Intel hopes that being the first to wield high-NA technology will give it an edge in the silicon rat race, making it possible to print tiny patterns faster than anyone else.
It could also make things simpler for customers. Over the years, while waiting for EUV machines to emerge, chip designers used multi-patterning to squeeze more life out of the older forms of light. Every chip is made out of layers, which are laid down to make components like the switches and wiring. If you’re working on one of those layers and need to make features tinier than your machine can normally produce, you can break the pattern for that layer up into several patterns and then expose the wafer to them one at a time. This strategy helped chipmakers keep using older (and cheaper) machines while still creating tinier and tinier components. But multi-patterning is a hassle: It’s more challenging to design the complex overlay of patterns, and much slower to print each chip. Designing a chip is far easier if you know you can do “single patterning,” blasting each layer in one go.
Observers say it won’t be easy to build a foundry business that bests TSMC and Samsung on their own terrain. “Leapfrogging is difficult,” Hijink says. But it’s also true that the high-tech world has such a ravening hunger for better chips that Intel could succeed, simply because even TSMC and Samsung can’t fulfill all that need.
“There’s spillover demand, so Intel can survive off that,” Koch says. “It’s not even scraps now. It’s a meal. It may not be the best foundry, but they can make chips, and there’s only three companies that can do that, right?”
TSMC, for its part, seems to be biding its time when it comes to high NA. “TSMC will deploy high-NA EUV when it is mature and ready to deliver maximum benefit to our customers,” the company wrote to MIT Technology Review. Some suspect it won’t use the machines in serious volume until the 2030s. Part of the reason is cost: TSMC is ruthlessly focused on producing chips as cost-effectively as possible, and the high-NA tools are a blistering $400 million each, far more than the previous EUV rigs. And unlike those, the new machines are not a revolutionary leap upward.
“This is like 30% to 50% better in terms of capability,” says Koch, the analyst and former ASML employee. “This is probably the first tool that hasn’t obviously made business sense right away for ASML.”
It’s not that the industry won’t eventually embrace high NA en masse, Koch says. Most companies will need to, if they want to keep going smaller. But TSMC is more likely to push ahead as far as it can go with its existing EUV tools, using onerous multi-patterning to wring as much as it can out of that generation until it absolutely needs to switch.
“The industry has only shifted paradigms when it just absolutely cannot extend—even one more little bit—out of what it’s been doing,” Koch says.
China isn’t the only party looking to upset the current balance of power. The dominance of ASML, and the swelling cost of its tools, is prompting other upstarts too. But instead of trying to replicate ASML’s breakthroughs in EUV, they’re doing an end run—working on lithography tools that use entirely different forms of light. These will be far cheaper, they promise, and just as powerful.
One is Substrate, a San Francisco–based startup. Founded four years ago, it’s working on a tool that uses x-ray light produced by a particle accelerator. X-rays have a remarkably tiny wavelength, making them a potentially powerful way to create minute features.
Particle accelerators have historically been enormous, making them difficult to fit into a chipmaking process. Substrate says it has harnessed decades of scientific improvements in particle acceleration to produce a light source that’s smaller and suitable for mass production.
Last year the company released images showing that it had created fine patterns, which Proud, the CEO, says are only possible now with a high-NA EUV machine. He says Substrate’s goal is to produce chips at scale by 2030.
But Proud doesn’t intend to sell the tools to TSMC or Intel. Indeed, he doesn’t plan to sell them to anyone. Instead, Substrate wants to create its own fab, building chips using its own tools.
“The amount of chips we’re going to need is going to be many orders of magnitude larger than even the wildest projections you have now.”
James Proud, cofounder and CEO, Substrate
The semiconductor industry, Proud argues, needs new approaches, because it’s become too pricey and too centralized. A single fab today can cost $25 billion to build, up from about $5 billion in the 2010s, the company notes. It’s driving the cost of a single wafer full of advanced chips up toward $100,000, Proud says.
“That is, I think, a prohibitive cost,” he says. There also isn’t enough capacity in the supply chain: “It’s relatively slow and hard to flex to the current increase in demands.” He admires ASML’s EUV tooling—it’s “the apex implementation of that technology”—but new approaches are needed.
That’s partly for national security reasons. Proud and his team think it’s too dangerous for the US to rely on foreign supplies. But he also predicts the current AI boom will go into overdrive, creating a massive demand for chips that the existing ASML/TSMC duopoly won’t be able to deliver: “The amount of chips we’re going to need is going to be many orders of magnitude larger than even the wildest projections you have now.”
ASML’s machines use lasers and molten tin to generate the EUV light.
CHRISTOPHER PAYNE
Substrate predicts it will be able to produce finished wafers at $10,000 a pop—a tenth of where Proud predicts the rest of the industry is heading. Proud says that’s partly because the company’s system will be vertically integrated, so it will control all parts of the chipmaking process, but also because its lithography tooling will be less complex: “We’re able to put together in a sort of simpler package.”
Still, Substrate is playing its cards close to its chest. Unlike ASML, the company isn’t offering nuanced detail on how it generates light, or on how that then translates into making patterns on a wafer.
Substrate’s ambitions give some industry observers pause. Hijink, who thinks it is probably “unachievable and impossible” to simultaneously master both a new form of lithography and high-throughput fab techniques, regards the company’s secrecy as a red flag. “This industry is about open innovation,” he says.
Koch is more impressed by its ambitions and funding. The type of technology it’s pursuing “is really cool,” he says. “It’s interesting.” But “there’s a long road between lab-scale demonstration and high volume,” he adds. “Is this like an imminent disruption to ASML? Probably not.”
Another startup that is aiming to hit the market around the same time as Substrate is Lace Lithography. Based in Norway, it is devising an entirely different approach—one that doesn’t use light at all. Instead, an energized beam of helium atoms is pointed at the pattern on the reticle. When the helium atoms then hit the wafer, the atoms transfer their energy to it, imparting the design to the chip.
The idea dates back a while. Bodil Holst, the CEO, took it up in 2008, when she was a physicist studying the use of atom beams. MIT professor Henry “Hank” Smith, a pioneer in using x-rays for lithography, told her she should explore using atoms as a mechanism for making microchips, because back then he wasn’t sure ASML’s EUV moonshot would work. “Even if it does, we’ll need atoms eventually,” he told her.
Holst did some experiments to investigate the idea further and partnered with a former PhD student—Adrià Salvador Palau, a physicist and expert in machine learning—to found Lace. Like Substrate’s, its tool is completely different from ASML’s massive machinery. The source of the excited atoms “looks a bit like a rocket motor,” says Palau. “It’s very cool.” While EUV’s wavelength is 13.5 nanometers, the helium atoms offer a precision of 0.1 nanometers. The process also requires far less power, and the machine is intended to be far smaller. Holst tells me the company aims to have machines ready to sell to fabs by 2029 or 2030.
“I think everybody’s really looking forward to something that extends a road map beyond light, beyond EUV,” Palau says.
ASML is watching these upstarts with curiosity. Benschop says he can’t assess whether Substrate’s technology will work reliably and affordably, because the company hasn’t explained anything about its processes. But he went to a conference where Holst and Palau did a presentation outlining Lace Lithography’s technology.
“I’m incredibly impressed with how they do it,” he says. The problem, he says, is he doesn’t think the process produces patterns on the wafer that are deep enough to be useful. “I cannot see how they would scale it to a viable volume product,” he told me.
He suspects ASML’s mastery of EUV will keep it on top for the near future. “So far, I have not seen a viable alternative,” he says. He thinks there’s “no serious runner-up” when it comes to volume manufacturing of the most advanced chip generations.
It’s true that major shifts in chipmaking are slow, says Chris Miller, a professor of international history at Tufts University and the author of Chip War, a book about the worldwide struggle for dominance in the industry. “No doubt we’ll eventually have alternatives [to EUV],” he told me via e-mail. “But it’s worth noting that lithography transitions have historically taken years, if not decades.”
ASML’s executives, too, are pondering their future. Benschop expects high-NA technology to dominate chipmaking into the 2030s. Beyond that? The industry has, indeed, tended to shift to a new form of light every decade.
“You may argue it’s time for the next decade,” he told me after we’d stripped off our bunny suits and he was relaxing with a coffee.
But ASML’s executives suspect they can continue to squeeze more capabilities out of EUV by increasing the numerical aperture even further on their existing machine. They’re already toying with a design that would take an NA of 0.55 to an NA of 0.75: “hyper NA.” It could let them pattern wafers with a resolution of six nanometers. They’re also working on standardizing their various optics into a platform of a single size, so customers could order one machine outfitted for either regular EUV, high NA, or hyper NA. If it’s all in the same-sized unit, it would simplify the costs and logistics of integrating each into a fab. If the company goes through with it, Benschop figures, the hyper-NA tool might hit the market seven or eight years from now and be sold in volume during the second half of the 2030s.
For now, the ball is in ASML’s court. “We’re pushing the limits of physics,” Pieters told me. The question now is whether anyone else can push harder.
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.
For those of you enjoying your summer unaware of Anthropic’s latest feud with the US government, here’s a recap: In April the company said it had built an AI model called Mythos that was so good at working with code it could pose a global cybersecurity threat. Anthropic gave access to a small group of cybersecurity experts so they could see what they were up against. Then it released a modified version called Fable which it said was safer to the public on Tuesday, June 9. That Friday, the federal government told the company it was a threat to national security and placed export controls on the new release. Anthropic revoked access to both models hours later.
People worried about catastrophic effects of AI—broadly labeled “doomers”—have said for years that the technology poses a threat to humanity and published proposals for how the government should intervene in its development. The doomers just got their government intervention—not over a bioweapon or rogue AI, but in response to an AI model that’s basically just really good at coding. And the result so far looks less like a safety plan than like a superficial reaction.
There’s plenty to dissect about what happened in those few days that led to such drastic action from the government, and it’s notable that Amazon CEO Andy Jassy was the one who told government officials that Fable would be dangerous (Amazon is both invested in Anthropic and building its own competing AI models). It’s also possible this will be a short-lived ban from the government that doesn’t survive legal scrutiny (it’s not clear that Anthropic’s offering access to Fable really counts as “exporting” it, for example).
But there are ripple effects happening already.
For one, this is making a whole lot of people not want to rely on American AI companies. TheFrench politician Bruno Retailleau described it as a “wake-up call” that should motivate Europe to build more AI. But any vision of turning Paris into Silicon Valley—touted by many other European leaders following the shutdown of Anthropic’s models—is complicated by one big thing: China.
Open-source models from China are very capable and incredibly cheap, and they can be downloaded to run on anyone’s servers with no rules or guardrails. (This makes them attractive to companies that don’t want access turned off on the basis of a decision from the White House—but equally attractive to cybercriminals, the type that Anthropic hoped to fend off by building safety guardrails into its models.)
It’s possible that companies, including those in the US and Europe, will decide that working with Chinese models is just easier, as the skyrocketing of shares in the Chinese startup Zhipu suggests. Playing this forward, is it possible the government’s next drastic decision will be to say that US companies using models from China pose a threat to national security? I wouldn’t write it off.
Second, it’s possible that shutting off access to Anthropic’s models will leave the country morevulnerable to cybersecurity attacks, not less.Leading cybersecurity experts have said as much in an open letter to the government, writing that access to Anthropic’s models was helping researchers prepare defenses, and that the company’s models are no more dangerous than other leading models that are widely available. Such is the risk of applying the concept of nonproliferation to software—trying to control and restrict dangerous AI models in the manner of the uranium used for nuclear weapons.
The third thing worth watching is how US lawmakers will react. Remember that following Anthropic’s last feud with the government over how the Pentagon could or could not use its models, a slate of new bills was introduced that would define the limits of military AI.
Right now, the biggest players shaping how AI gets used are the companies and the White House. There’s been much talk about more federal AI regulation, and polling suggests most Americans want it. Lawmakers are still figuring out whether to form rules on how kids use chatbots and are far from a clear answer on the extent to which the government should vet the safety of AI models. But with every drastic action from the White House, the pressure for regulations rises.
To state the obvious, predictions are hard when the administration’s attitudes toward AI change with the wind. When President Trump took office, he threw out the restrictive rulebook for how to make AI safe and promised to get out of the way of tech companies. The White House has now called the most valuable AI startup a risk to national security once in the spring, and again in summer. What will fall bring?
The Miami-based AI startup Subquadratic came out of stealth mode last month with a huge claim. It announced that it had solved a mathematical bottleneck that had been holding back large language models for almost a decade.
The details were thin, and many people were unconvinced. But Subquadratic has started to bring the receipts, sharing the results of an independent evaluation of its new tech. The results suggest that the company’s claims might be worth paying attention to.
According to Subquadratic, it has developed a new kind of LLM, called SubQ, that is faster and cheaper and uses a lot less energy than any other model on the market. The company also claims that SubQ is able to process up to 12 times as much text at once as most other models, allowing it to carry out a range of data-heavy tasks, such as analyzing hundreds of documents or entire code bases.
What’s more, Subquadratic says, SubQ does this while more or less matching the performance of the best models put out by Google DeepMind, OpenAI, and Anthropic on key tasks like coding.
The problem was that the company at first provided little evidence for its claims beyond a handful of self-published test scores. And it has yet to make SubQ widely available for people to try out themselves.
So it’s no surprise that Subquadratic’s claims were met with skepticism. Dan McAteer, an artificial-intelligence engineer, captured the overall response on X: “SubQ is either the biggest breakthrough since the Transformer … or it’s AI Theranos.”
“We expected healthy skepticism,” says Subquadratic cofounder and chief technology officer Alex Whedon. “In hindsight, releasing the third-party benchmarks alongside the initial announcement would have preempted much of the skepticism, which is why we’re taking the time to make sure any future results are fully verified before putting them out.”
Subquadratic asked Appen, which evaluates other companies’ models, to run its tests on SubQ. The results seem to back up a lot of Subquadratic’s claims. “That was really exciting to me, it validated their architecture,” says Jeanine Sinanan-Singh, Appen’s director of generative AI research.
“I was like, ‘Wow, this could be a game changer,’ because models struggle with speed and inefficiency,” she adds. “But when you have kind of shocking results, it’s really not as credible when you say it yourself.”
SubQ won’t replace existing top models across the board, but it could offer huge increases in speed at a fraction of the typical cost for certain tasks. Subquadratic insists that in the long run, though, its breakthrough could change how LLMs are built. “We hope we’re kicking off a new age of efficiency,” says Justin Dangel, the firm’s cofounder and CEO. “We don’t think anybody will be building on transformers in a few years.”
Attention!
To understand why Subquadratic’s claims are a big deal, let’s dig into how most LLMs work. The key mechanism inside an LLM is a type of neural network called a transformer, which runs a process known as dense attention. Today’s LLMs typically chain together multiple transformers. (The foundational paper of the LLM era, published by researchers at Google in 2017, was titled “Attention Is All You Need.”)
Dense attention works like this: When a transformer processes a chunk of text, it first encodes each word (or part of a word, known as a token) with a number. To capture the meaning of the full text, it then multiplies each of those numbers with every other number for that text. For example, a piece of text 10,000 words long would kick off almost 50 million individual multiplications. That’s a lot of computation and the main reason that LLMs are notorious power hogs.
“If you want to summarize The Great Gatsby, you have to look at the first word and the last word together, and then you have to look at every other combination,” says Dangel.
As the length of the text increases, the number of computations skyrockets. That’s because each additional number must be multiplied by all other previous numbers. Double the number of words, and you roughly quadruple the number of computations, a rate of increase known as a quadratic expansion.
(You can picture this yourself: Draw a circle and mark dots around its edge. Each dot is a token. Then draw lines between pairs of dots to represent the multiplication of those two tokens. A circle with five dots will have 10 lines crossing it. Make it 10 dots and you will have 45 lines, 20 dots and you will have 190 lines, and so on.)
Slashing costs
Subquadratic’s solution is to ditch dense attention, the core operation of a transformer, in favor of what’s known as sparse attention, which slashes the number of computations needed. Instead of multiplying the number assigned to each token by every other number, sparse attention selects just some of the numbers to multiply. The idea is that not all relationships between words in a piece of text matter.
“Sparse attention says not all of those relationships are important, because they’re not,” says Whedon. “If you’re reading a book, you’re not going to look at the first and second words, first and third—that’s insane.”
It’s a simple approach, and Subquadratic is not the first to try it. “Pretty much everything under the sun has been attempted,” says Will Depue, an independent AI researcher who previously worked at OpenAI. “It’s not impossible, but it’s akin to running a four-minute mile.”
Previous techniques for selecting which numbers to multiply and which to ignore have not produced a mechanism that can capture the meaning of a document as well as dense attention can.
Subquadratic claims to have cracked the problem at last. It pitches SubQ as the first sparse-attention LLM that rivals mainstream dense-attention models in performance.
“Historically, most mechanisms have used fixed patterns, like always comparing the first word to the fifth,” says Whedon. “That’s pretty limiting. Language is too sophisticated for that. And so, one of the things that makes our mechanism unique is that we dynamically select which ones are important.”
The firm won’t say exactly how SubQ chooses which words to focus on, but the selection is calculated on the fly and differs for each piece of text the model is given. “That’s kind of where the secret sauce is,” says Whedon.
Testing, testing
The upshot is that for certain tasks, SubQ may be faster and cheaper to run than most other models. Appen evaluated SubQ on a handful of standard tests. In a straight-up speed test, which sets a baseline for how fast a model can operate in theory rather than assessing what a model can actually do, Appen found that SubQ was 56 times faster than models using FlashAttention, a previous sparse-attention technique.
On LiveCodeBench, a test that looks at how well models perform on competitive coding problems taken from real contests, SubQ scored 89.7%, putting it in the same ballpark as other top coding models. “This model continues to provide frontier-level performance in coding,” says Appen’s Sinanan-Singh.
Subquadratic’s claims about cost are harder to verify because SubQ is not yet widely available. According to Dangel, it costs $2,600 to run Anthropic’s LLM Opus 4.6 through RULER 128, a test developed by Nvidia to assess a model’s ability to retrieve information from large data sets. And SubQ? “It cost us eight dollars,” he says.
SubQ does seem to be able to handle a lot of text at once. The model has a context window (roughly akin to a working memory) up to 12 million tokens long. Most top models today have context windows one million tokens long. In a demo that Whedon ran for me, he asked SubQ to perform a task that required it to reason about information contained in 400 documents. It responded in seconds. When he gave Perplexity—a popular LLM-powered search engine—the same task, it failed to load all 400 documents.
Appen put SubQ through the Needle-in-a-Haystack test, which, like RULER, assesses how well a model retrieves specific information buried in a large data set. In its report, Appen states that Subquadratic’s model scored 98% with context windows six million and 12 million tokens long, “sustaining near-perfect long-context retrieval at scales few models are tested at.”
Too good to be true?
Despite the high scores, benchmarks paint an incomplete picture of what a model can and cannot do. Testing under very specific conditions is not a substitute for running a model on a wide range of real tasks.
Subquadratic is offering SubQ as a model tailored to coding and to searching very large data sets. It says that tens of thousands of potential users have already signed up for early access, including more than 500 enterprise customers. But there’s a long waitlist, and the firm has given very few people access so far. Subquadratic’s response is that it is a new, small company with limited resources and cannot serve too many people at once.
Until more people get their hands on the model and try it out for themselves, some skepticism is justified. One nagging issue is that Subquadratic reused the weights (values set within a model during training that determine how it will behave) from a version of the Chinese open-source model Qwen to bootstrap SubQ, rather than training it from scratch. That’s a common approach for model makers to take, but it cuts across Subquadratic’s claim that it has fully reinvented how LLMs work.
“They may have built something real and useful,” says Depue. “But the public evidence does not yet justify the stronger claim that they have solved the quadratic attention bottleneck.”
In the meantime, Subquadratic cofounder Whedon insists that making something different was his only option. If you want to build a competitive model, you have to have new ideas, he says: “We’re more up against it than OpenAI is.”
At the end of a tense and scoreless first half of a soccer match between the English men’s team and rival Germany, millions of Brits let out a collective sigh and did what they so often do in moments of stress: They made tea. That wave of electric kettles clicking on, however, caused a different kind of stress: a huge and sudden increase in demand for electricity. But National Grid, which operates the local transmission network, was ready.
Just as those kettles started heating up, an AI program sent instructions to a data center in London to slow down some of the facility’s power-hungry chips. This reduction helped make sure there was enough supply to match demand, staving off potential blackouts or damage to electrical hardware. For data centers, which normally guzzle power without consideration for anyone or anything else’s needs, it was a radical departure.
It was also a test. The software was controlling a real data center, but there was no game happening at the time, in December 2025. Engineers were doing a trial run for a new breed of data center built to be flexible about its electricity needs, so they re-created the energy demand facing the UK’s grid during a match from the 2020 Euro tournament. They wanted to see how their software, called Conductor, would have responded had it been online at the time.
Conductor is the signature product of Emerald AI, a firm based in Washington, DC, that’s part of a wave of companies trying to figure out whether data centers can work within the confines of the existing electric grid.
This year, Emerald is set to deploy Conductor in a new facility in the part of Virginia known as Data Center Alley, this time connected to the live grid. When overall demand spikes, Conductor will turn down the power used by the data center, while making sure its servers still carry out their timeliest and most important jobs. Emerald’s partners on the project—which include Nvidia and the giant data-center operator Digital Realty—bill it as one of the world’s first “power-flexible AI factories.”
Demonstrating that data centers can participate in this kind of give-and-take could ease what many tech leaders identify as the bottleneck in getting facilities online: It takes far longer to get approval for, construct, and connect new power plants than to build data centers. PJM, the grid operator in Virginia and the largest one in the US, for instance, needs eight years to bring new generation online, according to RMI, an energy research and advocacy group. “We need to solve the energy equation,” says Josh Parker, head of sustainability at Nvidia. “AI factory flexibility is the bridge between the incredible demand for AI and the immediate limitations of our energy grid.”
Speed, though, is only one of the issues. Once facilities do plug in, neighbors often criticize them for drawing too much electricity and contributing to rising prices. They say the data centers generate more noise than they do long-term jobs, contribute to pollution, and threaten to put people out of work. Organizers stalled over $150 billion worth of projects in 2025, according to Data Center Watch, and policymakers alert to the public mood are starting to impose limitations on development.
More than a dozen states are considering bans, and local moratoriums are in effect in places like Minneapolis and DeKalb County in Georgia. At the federal level, the GRID Act, a bipartisan bill in the US Senate, proposes to sever new data centers from public grids entirely. Some operators are already moving that way by trying to develop their own power generation.
Rather than rushing to build new power plants, companies could find part of the solution to the crunch right under our noses—or, more precisely, in the transmission lines under our feet and above our heads. The existing system operates near its full capacity during only a small number of high-demand hours throughout the year. This means, some grid experts argue, that if data centers can limit the power they draw during those stretches, they won’t need to wait for big infrastructure upgrades or build their own off-grid generation.
Indeed, a growing number of studies have shown there could be plenty of power available for data centers that can flex. A widely discussed 2025 report from researchers at Duke University found that the US grid could offer an additional 76 gigawatts—about 5% of its entire capacity, and about enough to accommodate projected data-center growth in the US through 2030—to facilities that are willing to reduce their usage just 0.25% of the time. That’s about 22 hours a year. And when researchers from Princeton University and two grid-modernization companies looked at locations for new data centers in the PJM region, their report, which was funded by Google, found that a 500-megawatt facility capable of flexing for less than 1% of the year could reach full operation three to five years faster than one that’s inflexible.
Flexible power connections could also help data centers address some of their PR problems. By decreasing their draw at times of grid stress, for instance, they could avoid diverting power from where it’s most needed, thus boosting stability. By using existing capacity, they might be able to reduce the need for new fossil-fuel power plants and spread fixed costs over more electricity users, pushing prices down.
The AI power pinch is attracting resources and research into strategies for grid flexibility overall, which could help negotiate a tricky period: Taken together with electric vehicles, air-conditioning, and other sectors, data centers are helping drive what analysts predict will be a 25% increase in US electricity demand by 2030 compared with 2023 levels.
Ideally, flexibility gives grid operators more control over the flow of electrons, making them leaders of a harmonious ensemble rather than hostages to inflexible electricity requirements. That will help them manage demand spikes across the entire system and deal more effectively with the intermittent nature of renewables like wind and solar. “Demand flexibility is incredibly useful for power grids,” says Johanna Mathieu, a grid expert at the University of Michigan. “It helps reduce electricity costs and improve grid reliability.”
But while advocates see plenty of benefits, the concept brings complexity. For data centers, compromising on energy needs can be a hard sell. Flexibility requires utilities and grid operators, which tend to be operationally conservative, to change long-held practices. And some skeptics also say that flexibility distracts from the very real need to build more grid infrastructure faster, and could even pose risks to our electricity supply.
Still, some technologists, grid operators, and utilities are hoping to show that flexibility works—not only in white papers or simulations but in real life.
The poster children for data-center growth default toward inflexibility. Hyperscalers like Microsoft and Oracle have proposed enormous new centers, many of which would rely on off-grid, natural-gas-burning power plants. When xAI wanted to speed up the buildout of the Colossus site outside Memphis, Tennessee, it rolled up with gas turbines on flatbed trucks. The facility, now in operation, is facing blowback from regulators and residents about the spike it’s causing in emissions and other pollution. In any case, there aren’t enough gas turbines worldwide to meet the demand from data-center operators.
One big obstacle for anyone demanding a lot of power is that our grids are mostly rigid. They’re designed to supply enough power to meet total demand when it’s highest, even if that’s for only a relatively small number of hours a year. That conservative approach is a simple route to reliability, but it means that the grid has quite a bit of headroom. “The grid is already overbuilt by a lot. If you were an airline running at 30% utilization, you would not buy more planes,” says Amit Narayan, the cofounder and CEO of GridCare, a company developing flexibility technologies, referring to a 2025 Stanford study of transmission lines in western North America. “If you are running a grid at 30% utilization, there’s no scientific reason you can’t go to 60.”
“If you were an airline running at 30% utilization, you would not buy more planes. If you are running a grid at 30% utilization, there’s no scientific reason you can’t go to 60.”
To be fair, the idea of flexibility isn’t entirely foreign to grid operators. For decades, they’ve practiced a technique called demand response: When it looks as if demand will get too close to supply, as it might during a heat wave when many people turn on the AC at the same time, they call large commercial or industrial facilities and ask them to shut down parts of their operations. This method can help avoid the need to fire up so-called peaker plants, which run on fossil fuels, but it’s slow, imprecise, and hard to scale.
In the 2000s, as the adoption of technologies like electric cars and solar panels presented new challenges, more internet-connected grids also provided new means of flexibility. Virtual power plants, or VPPs, offered a smarter, faster, more granular alternative. Electricity customers ranging from factories to homeowners with smart thermostats, solar panels, or big batteries would allow the utility to adjust their draw to help meet demand—often getting paid for their (frequently unnoticed) trouble.
After the generative AI boom began with the release of ChatGPT in 2022, some companies began to see flexibility as a way to get data centers set up more easily, efficiently, and affordably. If they bring AI money into existing grids and reduce or defer the need for expensive upgrades, data centers could actually help spread out fixed costs so as to lower rates for other users. A study from Duke University published this past February, for instance, found that flexibility could reduce rates by 0.5% to 2.8%.
PETRA PÉTERFFY
The trick is figuring out how data centers, notorious power hogs, can keep operating when their flexible connections are throttled. Flexibility specialists envision three possible ways. The simplest is for the new data center to install on-site backup power storage or generation to tap when the grid is maxed out—at their own expense, of course.
A facility could also fill the gap by drawing on a VPP. The utility would turn down the electricity going to users who signed up for the VPP, and the data center would pay them for their flexibility. This method wouldn’t require any major infrastructure, but it would require the utility to have a big VPP program and to coordinate the exchange at a time when the grid was under stress. While VPPs exist to some extent in nearly 40 states, the rules governing them vary widely, and they are empowered to do more in some areas than in others.
Finally, a data center could simply use less power at peak times. The conventional wisdom is that they won’t go for such limits, particularly when every number-crunching server can feel like a goose potentially laying little golden eggs. But some experts are betting that the value of getting up and running quickly is enough to change their minds. “There is a clear and growing trend,” says Ayse Coskun, chief scientist at Emerald AI. “Operators are increasingly willing to trade some level of flexibility for faster grid interconnection.”
GridCare, a startup based in Silicon Valley, was one of the first companies to use flexibility to get data centers online quickly. Instead of looking at grids only in worst-case scenarios when electricity demand is highest, the company analyzes the system under all conditions, explains CEO Narayan, who studied smart grids at Stanford. It feeds every part of the grid—including power plants, lines, substations, and homes—into a generative AI model that creates a “digital twin” for different grid configurations. It then picks out results that could unlock capacity while maintaining reliability, and it feeds those into another model trained on the physics of electrical components like resistors and capacitors to make sure they’re realistic.
GridCare found its first customer in the Silicon Forest, an area in the Pacific Northwest named for the trees that dominate the landscape and the IT industry that has more recently sprouted up there. The local grid needed more capacity to support more data centers. “Data centers wanted ‘speed to power,’” says Isaac Barrow, a manager of data-center relations at Portland General Electric, or PGE, the local power generator and distributor, “but transmission buildout is a long process that’s very costly.”
In 2024, Aligned Data Centers came to PGE wanting to expand its operation in Hillsboro, Oregon, and PGE followed a recommendation from GridCare. Aligned will install a 31-megawatt battery, set to be in service in May 2027, and decrease its draw by up to that amount when the grid becomes congested. Bundled with other flexibility measures, that battery has allowed PGE to increase the capacity it can offer Aligned and other nearby operators by 80 megawatts without any new power plants. Though the buildout of data centers in Hillsboro has faced plenty of pushback from locals, Barrow points out that it could have the knock-on effect of lowering costs for ratepayers, because it spreads out the tab.
Other companies are promoting different flavors of flexibility. Google has been moving processing loads from facilities in areas experiencing demand spikes to those in less stressed spots since 2023. It’s signed agreements with five utilities, including the Tennessee Valley Authority and Indiana Michigan Power, that add as much as a gigawatt of flexibility.
Voltus, a major VPP provider across the US and Canada, markets a “bring your own capacity” program in which a data-center company can fund a VPP nearby. The grid operator can use the VPP to decrease demand at busy times, and participants get a financial thank-you. “We can spin up new VPPs on the order of months,” says Emily Orvis, Voltus’s vice president of energy markets. In June, the company signed their first such data-center deal: a three-year plan in which Google will bankroll a VPP in the PJM interconnection.
Of all the approaches to flexibility, Emerald AI’s may be the most ambitious: asking data centers to dial into the grid’s needs. The company’s Conductor software, which can run on premises or in the cloud, builds on the research of chief scientist Coskun. Her group at Boston University showed in a pair of 2013 papers that a data center could watch the grid and help balance big power fluctuations, such as the intermittent effects of solar and wind power. By 2022, she and her colleagues had tested their methods on a cluster of 36 research servers and shown that the system could respect power limits without breaking the processes it was running.
One of the most important questions for Conductor is deciding which AI processes can be slowed down to save energy without kneecapping performance. A lot of companies label their jobs by priority—a real-time chatbot query, for instance, might outrank something like a web search that’s part of a deep research project. When they don’t, Emerald AI tries to infer priority from the nature of the job. Conductor then analyzes the AI workload to determine how tweaking the power to a given processor will affect the performance and help meet the usage limits set by the grid operator.
“The performance curve changes for different kinds of workloads,” says Coskun. “Each AI job is going to have a different location on that curve. Our intelligence is figuring out where you are on that curve.”
PETRA PÉTERFFY
Last year, Emerald AI began assessing the technology’s readiness for real-world use in a series of tests, raising the difficulty each time. The trials were carried out in partnership with the Data Center Flexible Load Initiative—a collaboration among tech companies like Google and Nvidia, utilities like Duke Energy, and grid operators like PJM that aims to help establish a repeatable framework for power-flexible data centers.
The first challenge was in Phoenix, a fast-growing computing hub. For the test, Conductor took control of a group of server racks laden with 256 Nvidia A100 GPUs—hardware that can use about as much power as around 170 US homes. When presented with a simulation of a busy grid, Conductor reduced the power to the chips by 25% for three hours, while maintaining acceptable computing performance. Emerald AI and its partners reported the results in a paper in Nature Energy in December 2025.
The next trial forced the system to juggle surprise grid fluctuations without advance warning and redirect AI jobs from a data center in Virginia to a less busy one in Chicago. Then, in London, Conductor took the reins of equipment beyond the main GPU processors and faced a more complicated mix of fluctuations, including very short and long bouts of congestion—plus the notorious teakettle effect.
The progress so far shows that flexibility can work, at least in some situations, but only a small fraction of operators have pursued it as yet. “We’re just in the beginning innings of the game,” says Jesse Jenkins, one of the authors of the 2025 Princeton study and cofounder of Firma, a startup that works on data-center flexibility. “People are recognizing that this is a potential solution. The motivation is there; there are some bespoke examples. But there’s no uniform solution set that’s the default option, which is where we need to get.”
While data centers are going up across the US, no place on Earth comes close to the accumulated computing muscle in Northern Virginia’s Data Center Alley. The region is home to around 500 compute-crunching facilities, which represent 13% of the entire world’s capacity; the next two hot spots, Beijing and Oregon, contain 6% each.
There are proposals to build hundreds more facilities in Virginia, but a government study found that the state’s electricity demand will increase 183% (around 26 gigawatts) by 2040 if they all go forward, and supporting even half would be difficult. The power-flexible data center that Emerald AI, Nvidia, Digital Realty, and their partners are building in the suburb of Manassas could demonstrate how data centers can squeeze the power they need out of existing capacity. The facility, slated to come online later this year, is intended to give Conductor the chance to manage power at the largest scale yet and to respond to conditions on a live grid for the first time. In the UK demonstration, Conductor managed a 130-kilowatt AI cluster; in Manassas, it will pull the strings of a 96-megawatt hyperscale AI factory.
Some degree of flex will play a key role as we transition away from fossil fuels and toward a future that has to juggle technologies like solar and wind power, batteries, and electric cars.
For PJM, the Manassas facility points to a potential path through the current power crunch. “We think data-center flexibility, in different forms, will be essential for the reliable integration of data-center load over the short to mid term,” says Scott Baker, who manages demand-side markets at PJM.
But not all grid experts are so sanguine. PJM’s market monitor, which oversees the grid operator, says there are no workarounds when it comes to adding capacity. “The notion that large amounts of data-center load can be added without adding new generation is magical thinking,” says Joseph Bowring, an economist and the head of PJM’s market monitor since 1999.
One problem, he says, is that there’s no way to guarantee that a data center will actually take less power when demand is high. That is, absent any legal or regulatory push for flexibility or compliance, the utility won’t be able to step in to help prevent, say, a blackout. Utilities can rely on resources like power plants, but they can’t control or rely on data centers. “They do not want to be fully interruptible,” Bowring says of the facilities.
Stephen Empedocles, an advisor for technology companies, views flexibility as more of a tool than a silver bullet. “These approaches are excellent for improving grid reliability and getting more out of the infrastructure we already have,” he says, “but they are optimization tools.” They’re not substitutes for the “generation, transmission, and distribution expansion that will still be required,” he continues.
Flexibility advocates agree that over the long term, whether or not AI continues to boom, electrification will drive a need for more generation and transmission. Some degree of flex will play a key role in using grid infrastructure better as we transition away from fossil fuels and toward a future that has to juggle technologies like solar and wind power, batteries, and electric cars. A report published by the International Renewable Energy Agency in January 2026 found that grids around the world will need three times as much flexibility in 2030 as they had in 2019—and 10 times as much by 2050—to balance increasing demand with fluctuating supplies of renewable energy.
The challenge of powering AI could provide just the spark we need to do the work of designing and building smarter, more flexible grids, says Coskun. “I think with a crisis like this, there’s no quick solution,” she says. “Sometimes a crisis like this creates an opportunity to do something differently.”
Amos Zeeberg is a freelance science and technology journalist based in Bucharest. He’s developing a book about technology networks, including electric grids.
This story was updated on June 20, 2026 to clarify details about Emerald AI’s test in London in 2025.
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.
When I landed in Seoul after a grueling 12-hour flight from San Francisco, I walked through an unmanned immigration checkpoint, where a machine scanned my face and passport. On the subway home, people were glued to their phones (powered by flawless 5G even underground), as we raced past platforms lined with LED screens of ads celebrating K-pop idols’ birthdays. When I got off the station in Gangnam, a cartoon-eyed robot on wheels was waiting patiently at a crosswalk to deliver someone’s dinner. Internet cafés dotted the sidewalks, crammed with teenagers playing computer games, maybe hoping to become the next legendary pro gamer.
I stood at a bus stop with interactive touch screens showing real-time bus schedule updates. It will soon become an “AI bus stop,” the Gangnam district announced in June, with a kiosk that answers riders’ questions in multiple languages. The news didn’t surprise me. Having grown up in the city, I’ve watched Seoul transform from a scrappy boomtown into the gleaming tech capital it is today.
South Korea loves AI.
While a public backlash against AI is brewing across the US, South Koreans are optimistic. Only 16% say they are more concerned than excited about AI—the lowest of any of the 25 countries surveyed by the Pew Research Center—while 50% of Americans were more worried than excited. A majority of Koreans use AI every day, either as a sort of personal assistant or to do tasks at work, according to surveys by the Ministry of Culture, Sports, and Tourism and Korea Chamber of Commerce and Industry.
One of the most wired countries in the world, South Korea loves to street-test every new technology on the block—AI webcomics, virtual K-pop idols, and humanoid monks. And the appetite for experimentation doesn’t stop with ordinary citizens. Government agencies are early adopters too, deploying AI textbooks in schools and AI eldercare robots in welfare centers. South Koreans share a deep conviction that embracing technology is integral to modernizing the country and cementing its place in the global order. Their fascination with AI is just the latest incarnation of that ethos—and it’s making them anxious to stay ahead.
Engineered enthusiasm
All this techno-optimism has largely been engineered by South Korea’s national agenda to make AI a motor of economic growth. “The South Korean government has designated an AI-powered Fourth Industrial Revolution as the country’s path forward and aggressively promoted and invested in it,” says Chihyung Jeon, a professor of science and technology policy at the Korea Advanced Institute of Science and Technology. “South Koreans have consistently and relentlessly been told by the government about AI’s potential to create a better future.”
As South Korea rose from the ashes of the Korean War, technology lifted the nation from poverty into an economic powerhouse. In the 1970s, South Korea manufactured steel and ships, then semiconductors in the 1980s, broadband in the 1990s, and smartphones in the 2000s. Today, Samsung and SK Hynix supply most of the world’s high-bandwidth memory chips, which power the cutting-edge Nvidia hardware used to train AI models. South Korea’s economy now orbits these two semiconductor giants: The country’s main equity index, Kospi, surged to record highs in 2026, powered by the soaring share prices of both companies, each valued above $1 trillion.
Lee Jae-myung, president of South Korea, has pledged to vault the country into the ranks of the “top three AI powers” alongside the US and China. After taking office in 2025, he launched the Presidential Council on National AI Strategy to help buy massive amounts of computing power and a sovereign AI foundation model project that funds Korean companies to develop homegrown AI models. The government has also supported semiconductor titans, including Samsung and SK Hynix, through generous tax credits and low-interest financing.
South Korea’s policy posture also prioritizes accelerating AI development over safety considerations. In 2024, South Korea’s legislature passed the AI Basic Act, one of the world’s first comprehensive AI laws, to promote AI development and establish light-touch regulatory guardrails. Seventy percent of South Koreans say advancing science and medicine through AI innovation is a bigger priority than protecting industries through regulation, according to the 2026 Stanford AI Index.
All of that effort might be paying off. The same index ranked South Korea as having the third largest number of notable AI models in the world, based on criteria such as state-of-the-art advancements or high citation rates. For many small countries like South Korea, AI is a chance to punch above their weight.
The blind spots
But that single-mindedness can crowd out critical reflection on AI’s broader societal impacts. “Because the national agenda on AI prioritizes economic development,” says Jeon, the professor of science and technology policy, “there isn’t much reflection on the social, political, ethical dimensions of the technology.” In 2025, the South Korean government faced a fierce backlash for rolling out AI textbooks riddled with factual inaccuracies and data privacy risks without testing them first in a pilot program to evaluate how they affect student learning.
And despite their optimism, South Koreans are still worried that AI could displace them from their jobs. After Hyundai announced in January that it will deploy Atlas humanoid robots across its car factories, the Hyundai Motor Group union protested vehemently. “Without labor-management agreement, not a single robot using new technology will be allowed to enter the workplace,” the union said. Sixty-four percent of South Koreans fear AI could displace human labor and exacerbate inequality, although 52% believe it could also increase productivity.
On a recent Friday night in the Seoul Central Market, I went out with my cousins to a pocha, a late-night restaurant that serves fish cakes stacked in neat pyramids. As we clinked our cups of soju cut with beer—the scrappy staple cocktail of every Korean night out—one cousin asked me if I’d asked ChatGPT about my saju, a traditional Korean fortune-telling practice.
A 29-year-old insurance agent in Seoul praying for a new job and a boyfriend, she said asking ChatGPT about work and dating was her favorite pastime. She pulled up her phone and punched my birth date into the chatbot.
Addicted to their screens, trapped between unemployment and dead-end jobs, and priced out of marriage and homeownership, 46% of South Koreans in their 20s have used a chatbot to read their fortunes, according to a survey by Korea Gallup.
My cousin said she also asks ChatGPT for tips on trading stocks, dreaming big about making bank on her investment accounts into which she’s been pouring her salary. ChatGPT, she believes, is her portal out of reality into a better future.
Despite how fond she is of the chatbot as her shaman and financial advisor, she fears losing her job to AI. She still uses ChatGPT feverishly at work, as all her coworkers do, afraid of falling behind.
“I sometimes fear AI, but for now, it’s just so useful,” she said.
Login
