GPT-5.5 Pro also set new records on FrontierMath, scoring 52% on Tiers 1-3 (up from 50%) and 40% on Tier 4 (up from 38%). Across runs, it and GPT-5.5 solved two Tier 4 problems that no model had solved before, one by Hailong Dao and one by Ahsan Khan (@ahsanxrr). They had this to say:

For all these and more, check out our website!

Subscribe now

On FrontierMath, Opus 4.7 scored 44% on Tiers 1-3 and 23% on Tier 4, ahead of every model except GPT-5.4 and GPT-5.4 Pro.

On SWE-bench Verified, Opus 4.7 scored 83%, a new record for our evaluations on this benchmark. SWE-bench Verified may be increasingly contaminated, and we hope to evaluate Opus 4.7 on other coding benchmarks soon.

Opus 4.7 scored 30% on our Chess Puzzles benchmark. This is a substantial jump over previous Anthropic models, though it remains far from the frontier. Opus 4.7 hit its max output limit before answering for 31% of problems. We plan to improve elicitation here.

Opus 4.7 also solved the last problem on our older math benchmark, Mock AIME, that had not been solved by any model before. This problem involves interpreting the diagram shown below when given as Asymptote vector graphics code.

Opus 4.7 was a bit behind the top scores on GPQA Diamond (90% vs. 95%). Anecdotally, this may be due to refusing to answer some questions for safety reasons, even though GPQA isn’t meant to cover dangerous topics. See, for instance, the cut-off sample below.

ECI also incorporates scores from third-party benchmarks. So far, for Opus 4.7, we have ARC-AGI-1, ARC-AGI-2, WeirdML v2, and SimpleBench. Its ECI score will evolve as we collect additional benchmark scores.

For all these and more, check out our website!

Subscribe now

Our AI Chip Owners explorer contains interactive visualizations of our analysis of the number of leading AI chips owned by the largest US hyperscalers and cloud companies, one frontier AI developer (xAI), and Chinese customers — with breakdowns by chip family, chip model, and shifts in ownership over time. We build upon our estimates the total volumes of Nvidia, Google TPU, Amazon Trainium, AMD, and Huawei chips from the AI Chip Sales, and distribute these chips among major owners using estimates from analysts and industry researchers, company financial disclosures, capital spending, and our analysis of frontier-scale AI data centers.

The AI Chip Owners explorer is intended as a resource for researchers, policymakers, and anyone tracking the strategic landscape of AI compute. You can visit the explorer here, and read on for highlights!

Hyperscalers own the majority of global AI compute

Most of the world’s AI computing power is owned by hyperscalers and large cloud companies. “Hyperscalers” are the leading companies in data center deployments; among US companies, this refers to Amazon, Google (Alphabet), Meta, Microsoft, and Oracle. We estimate that over 60% of global AI compute (in terms of total computing power) is owned by the five US hyperscalers, led by Google.

Google holds the equivalent of around 5 million Nvidia H100 GPUs in compute capacity, roughly 25% of the world’s total! This capacity is led by the large scale of its custom TPU chips, which we estimate have a total compute capacity of almost 4 million H100-equivalents [confidence interval: 3.1M to 4.5M]. For the other hyperscalers, Nvidia is responsible for the majority of their AI compute acquired to date.

Besides Meta, the hyperscalers are all major cloud companies, meaning that they rent out part of their compute to other companies. Many frontier AI developers, including Anthropic and OpenAI, acquire almost all of their compute from hyperscalers and other cloud providers. OpenAI’s compute primarily comes from Microsoft, Oracle, and CoreWeave, and Anthropic’s from Google and Amazon, though neither should be assumed to be renting out the entire capacity of these clouds. We will follow up soon with an analysis of how much compute is used by frontier AI developers like OpenAI and Anthropic, as well as the frontier model divisions housed within Google and Meta.

Chinese customers own just 5% of global AI compute

We estimate that as of the end of 2025, Chinese companies collectively own just over 5% of the cumulative computing power of the leading AI chips sold in recent years — less than any single top US hyperscaler, and a share that has decreased over time. We bucket Nvidia, AMD, and Huawei chips purchased by mainland Chinese customers into a single “China” category, with specific customer breakdowns left for future work.

Our estimates of Chinese compute do not include chips smuggled in contravention of US export controls. Prior research and recent reporting suggest these volumes may be significant. Grunewald and Fist estimate that over 100,000 Nvidia A100s and H100s were shipped to China in 2024, though with significant uncertainty. These imports continued through 2025, according to multiple reports of Nvidia shipments upwards of several billion dollars, or at least tens of thousands of chips. However, chip smuggling does not seem likely to add up to the millions of chips required to significantly close the balance of compute between Chinese companies and US hyperscalers.

Like our other ownership figures, these China totals do not include any offshore cloud compute rented by Chinese companies.

Notably, in the past year Huawei has overtaken Nvidia as the leading source of AI computing power in China, at least in terms of rated FLOP/s, which may not reflect real-world performance. This is due to a pause in official Nvidia exports to China following tightened controls on the Nvidia H20 chip. Nvidia is now preparing to export the more advanced H200 to China.

We plan to follow up this work with more coverage and analysis of how AI compute is allocated and used by major players, complementing our coverage of overall AI chip sales and frontier AI data centers to give a detailed overview of the world’s AI compute capacity.

To learn more, visit the AI Chip Owners explorer to find the methodology, full dataset, interactive visualizations, and more analysis!

Thanks to Brendan Halstead, Erich Grunewald, Konstantin Pilz, and Theo Bearman for their feedback.

The newly-solved problem came from Will Brian, who had placed it in the Moderately Interesting category. It is a conjecture from a paper he wrote with Paul Larson in 2019. They were unable to solve it at the time, or in several attempts since. Brian had this to say.

Brian plans to write up the solution for publication, possibly including follow-on work spurred by the AI’s ideas. This matches his prospective assessment: a solution would be publishable in a standard speciality journal, and would be fairly likely to generate new questions.

Congratulations to Kevin Barreto and Liam Price, who first elicited a solution from GPT-5.4 Pro! They have the option to be coauthors, with Brian, on any resulting paper. Congratulations also to Geby Jaff who elicited a solution shortly thereafter.

We have replicated this elicitation in our scaffold for testing models on the problems. In this scaffold, Gemini 3.1 Pro, GPT-5.4 (xhigh), and Opus 4.6 (max) are all capable of solving the problem at least some of the time. For more about the problem, including a full chat transcript showing GPT-5.4 Pro’s original solution, and other models’ solutions in our harness, see the problem page on our website.

And check out the main FrontierMath: Open Problems page to learn more about the benchmark. So far one Moderately Interesting problem has been solved. Which problem will be solved next, and when?

We had pre-release access to evaluate the model. On Tiers 1–3, GPT-5.4 Pro scored 50%. On Tier 4 it scored 38%.

See below for commentary and additional experiments.

FrontierMath was funded by OpenAI, who has exclusive access to: all 290 problems in Tiers 1–3; solutions to 237 of these problems; 28 of the 48 problems in Tier 4; solutions to these 28 problems. Epoch holds out the rest.

On Tiers 1–3 GPT-5.4 Pro solved 52% of the non-held-out set and 42% of the held-out set. On Tier 4, GPT-5.4 Pro solved 25% of the non-held-out set and 55% of the held-out set. Neither of these differences is statistically significant.

GPT-5.4 Pro solved one Tier 4 problem that no model had solved before. In a preliminary analysis, it appeared to have found a preprint from 2011 which let it shortcut much of the intended work. The problem author was unaware of this preprint.

We ran GPT-5.4 (xhigh) an additional ten times on Tier 4 to get a pass@10 score. This was 38%. In one of these runs, it solved another problem no model had solved before. This problem was by Bartosz Naskręcki, who responded as follows:

Across all runs ever, 42% (20/48) of the problems in Tier 4 have now been solved at least once.

We also evaluated GPT-5.4 Pro on FrontierMath: Open Problems. It did not solve any problems. It made some novel observations on one problem, but of a form that the author had anticipated and characterized as relatively uninteresting. More here.

Check out our website for more results and commentary about FrontierMath overall!

Subscribe now

IEEE Spectrum just published a piece featuring our past & present work on FrontierMath — as well as Aletheia and First Proof.

Epoch researcher Greg Burnham called this “a more-the-merrier situation.”

Our own new contribution is FrontierMath: Open Problems, which consists of open problems from research mathematics that professional mathematicians have tried and failed to solve.

Subscribe now

We accidentally ran Gemini 3.1 Pro on Tier 4 a second time. The score above reflects the first, official run. But we noticed in the second run that it had solved a problem no model had solved before. The newly-solved problem is by Emmanuel Breuillard. See his commentary below.

We are waiting for API access to evaluate Gemini 3 Deep Think

Check out our benchmarking hub for more, including Gemini 3.1 Pro scores on other benchmarks!

Subscribe now

Catch up with an event recorded Feb 6th exploring recently-published research: Can AI companies become profitable?

In particular, explore whether OpenAI is profitable, as Azeem Azhar sets up the event, then Matt Robinson of AI Street interviews both Epoch AI’s Jaime Savilla and Exponential View’s Hannah Petrovic & Azeem Azhar to unpack their collaborative analysis.

• Can sign up via this Luma page, which will display the live event link (& also send an email when the event link is live).

• Or can follow Exponential View or Azeem Azhar on Substack to be notified within Substack when the event is live.

Note that Substack live events require a (free) Substack account.

Kimi K2.5 set a new record among open-weight models on the Epoch Capabilities Index (ECI), which combines multiple benchmarks onto a single scale.

Kimi K2.5 also took the lead among open models that we have benchmarked on FrontierMath. Its score of 28% on Tiers 1–3 is on par with GPT-5 (medium) and Gemini 2.5 Deep Think. On Tier 4 it scored only 4% (2/48 problems solved).

We evaluated Kimi K2.5 on Fireworks, the developer’s launch partner. When using third-party providers, performance differences are always a concern. We can note, at least, that our score for GPQA Diamond matches the score in the developer's release materials (87.6%).

Check out our website for more benchmark scores, on Kimi K2.5 as well as other models!

This work was supported by a grant from Schmidt Sciences.

AI math capabilities have come far and fast. In mid-2024, high school math was still a challenge. By the end of 2025, AI systems were solving extremely hard problems designed to be solvable only by top human experts. As we write this, it seems likely that AI systems will soon be able to solve problems that no human has solved before.

Indeed, there are already glimmers of this, e.g. AI systems solving several previously-unsolved Erdos problems. However, these results have been difficult to contextualize. Are these problems mathematically significant? How hard had humans previously tried to solve them? Does this say anything new about AI capabilities?

Today we are releasing a pilot version of a new benchmark, FrontierMath: Open Problems, which we hope will shed light on this topic. The benchmark consists of open problems from research mathematics that professional mathematicians have tried and failed to solve. To facilitate evaluation at scale, we include only problems for which proposed solutions can be verified programmatically.

At the time of release, to the best of our knowledge, none of these problems have been solved by humans or AI systems. If an AI system solves any problem, it will be a meaningful advance in the frontier of human knowledge. Moreover, we will be able to say something about just how meaningful: contributing mathematicians have characterized the significance of the problems, ranging from moderately interesting results to major breakthroughs.

The pilot release consists of 14 problems with:

• Write-ups explaining the problem’s significance and difficulty

• Precise prompts that can be given to AI systems – try them yourself!

• Initial attempts by AI systems to solve the problems

We will add more problems in the coming months and are actively commissioning contributions. See our problem proposal form if you are interested in contributing.

As for the verifiers — the programs that evaluate candidate solutions — we offer access for a fee. We structure access this way to help defray costs of expanding the benchmark. Problems are labor intensive to create and each solution effectively reduces the number of problems in the benchmark, so we ask that those who wish to use the verifiers partner with us to help fund further expansion. We commit to grant access uniformly, and not exclusively to any entity. Inquire at math@epoch.ai.

Example problems

Problems are mathematically meaningful, diverse, and hard

Problems are contributed by professional mathematicians. Contributors suggest problems they are familiar with from their own research, and which they would be interested to see solved. Contributors rate the meaningfulness (notability) of solutions, ranging from results of moderate interest within a subfield all the way up to major breakthroughs. We aim for problems to be evenly distributed across these tiers.

Our goal is to find problems that are meaningful to mathematicians on their own terms. We do not select problems to be difficult for AI per se. Unlike problems contrived for a test, these problems are core to doing research math.1 We want to know whether AI systems can solve them: if they can, so be it.

That said, at least for humans, the problems are hard. Contributors rate problems with estimates of how many mathematicians have made serious attempts to solve them, with responses ranging from 2–4 mathematicians to 50–100.

Contributors also take a guess at human time-to-solve, specifically how long it would take the mathematician most capable of solving the problem to have a 50% chance of solving it, if they worked full time. Responses range from 1–4 weeks to 3–10 years.2 In other words, the human baseline is high.

Problems cover a range of mathematical topic areas. The pilot set has a tilt toward combinatorics and number theory, where we happen to have found the most problems amenable to automatic verification. We aim to keep up the diversity of topic areas as we expand the benchmark.

See the problem proposal form for more details on the problem sourcing process.

Solutions can be verified automatically

Evaluating AI solutions to unsolved math problems is a major logistical challenge. Math research typically proceeds via natural-language papers. Evaluating such papers is labor intensive and error-prone even for humans. While AI systems have made progress at evaluating natural-language mathematics, we cannot rely on the accuracy of their evaluations for advanced material.3

Our approach is to find problems where, even though no solution is currently known, a proposed solution can be checked by a relatively straightforward computer program running on a typical computer. It is not obvious that such verifiable problems exist, but they do.

For example, some problems ask for a very concrete mathematical object. One such case asks for a polynomial with a certain property.4 It is quick to check if a given polynomial has the desired property, but finding one is beyond the reach of any known technique, including highly optimized, large-scale search. The problem’s meaningfulness stems from the fact that a conceptual approach appears to be required to construct the desired object.

In other cases, we want a construction that works for all positive integers. We can’t verify this in general, but we can ask for an algorithm that takes an integer and returns a construction for that integer. We can verify the algorithm on a challenge set of integers where no constructions are currently known and where the integers are large enough that search is intractable. Success here gives strong evidence that the algorithm implements a general solution.

The downside is that this approach limits what we can ask about. Our ideal would be to take a random sample of all unsolved math problems, but this constraint introduces a bias. The benchmark problems tend to be relatively concrete, and may not require fuzzier mathematical pursuits like “theory building”. Still, we have been pleasantly surprised by how readily mathematicians have been able to come up with a diversity of mathematically meaningful problems that satisfy this constraint.5

Some benchmark problems may not be solvable

One risk inherent to this benchmark is that problems might not have solutions as stated. This issue takes two forms: either the desired mathematical object does not exist, or it does exist but is too large for the verifier to certify it as valid.

We don’t think such cases undermine interpretation of the benchmark overall. Successful solves are clearly meaningful. Failure to solve all problems at a given difficulty level is also probably meaningful, and will become more so as we grow the problem set beyond the current pilot size. We thus encourage interpretations of the form, “We have seen several examples of AI solving moderately interesting open math problems, but not yet any examples of major advances.”

That said, we strove to include only problems where a solution is at least likely to exist. For some problems there are heuristic reasons to believe that a solution of the desired form exists, and, in all cases, there is at least an absence of any reason to believe that no solution exists.6

Our target was an assessment from the problem contributor of at least an 80% probability that the problem was solvable as stated.7 However, mathematicians often expressed high uncertainty when assigning a probability in the 50–80% range.

Solved problems will be removed from the benchmark

If a problem is solved — whether by humans or AI — the result will be published.8 Thus, any future AI system attempting the same problem would simply be able to look up the solution. For this reason, we will remove solved problems from the benchmark.

While this “first-to-solve” set-up is atypical, we don’t think it undermines the value of the Open Problems benchmark as a whole. The benchmark won’t give a score that can be used to compare models’ ability to solve open problems, but it will tell us whether AI systems are capable of solving problems of a given difficulty and significance.

The benchmark helps track AI “research taste”

Most immediately, this benchmark asks whether AI can solve unsolved math problems. We think this also helps us track fuzzier concepts like “research taste” — how good AI systems are at choosing the right direction to pursue, noticing the right patterns, etc.

Such capabilities seem likely to be relevant for theoretical math, where the right ideas can be especially difficult to find. If AI systems are able to solve math problems that have resisted significant human effort, then they might be moving toward superhuman research taste in general.

It’s by no means a guarantee. Perhaps, like chess or Go, it will turn out that math’s formal nature makes it unusually “easy” for AI systems. Or perhaps AI systems will solve these problems in ways that we regard as relatively inelegant.9 Still, we are glad to add this benchmark to our toolbox for tracking harder-to-quantify capabilities.

We hope to see strong attempts to get AI systems to solve these problems

Our primary goal is to understand the frontier of AI math capabilities. However, it is not clear how best to elicit AI capabilities on the benchmark problems.

So far we have tried simply prompting GPT-5.2 Pro and Gemini 3 Deep Think in their web apps.10 The results of this are shown on the individual problem pages. In this setting the models are generally capable of solving “warm-up” problems: variants of the open problems where solutions are known. This tells us that they understand the instructions and are familiar with the subject area. It also helps test the verifiers.

But, when given the actual open problems, the models don’t show much promise. Sometimes they seem set on trying optimization approaches instead of the more conceptual approaches that are likely necessary. Other times they recognize the problem as open and simply give up.

It seems likely that more “thinking” will be an important ingredient. Models already plan, execute, revise, and iterate — but they may need a lot more time and resources to do so if they are going to have a chance at these problems. Exactly how to enable them in this way is an open research question.11

We’re working on a scaffold that can facilitate this sort of extended attempt, and we hope others will try as well. Contact us at math@epoch.ai with any questions.

Appendix: caveats on future AI solutions

This benchmark is essentially a preregistration of the interestingness of a set of math problems. That said, if these problems end up being solved in certain ways, some caveats may be warranted. Here we try to preregister these caveats as well, to help mitigate concerns about moving the goalposts later.12

Collaborations. There have already been productive mathematical collaborations between humans and AI systems. A typical pattern is for the AI system to work out some examples, and for the human to generalize to a complete solution. Indeed, using computers to search for useful examples long predates LLM-based AI systems. We will have to assess the division of labor behind any such solution. The more an AI system is responsible for the conceptual part, the more that will be indicative of advancing capabilities.

Prior art. AI systems’ breadth of mathematical knowledge may already outstrip that of top humans. It is possible that an existing result has already done most of the work to solve a problem and the mathematicians who attempted the problem were simply unaware of it. This is unlikely for better-known problems, but, in any case, an AI solution relying on such a result would be significantly less indicative of advancing capabilities. Of course, if an AI system adapts known results in new ways, that needs no caveat: much of math research has this character.

Conventional compute. If an AI system suggests an optimized, parallelizable search algorithm and an AI company devotes a supercomputer-month to executing the search, then a problem could be solved with less mathematical insight than expected. While we selected problems where brute-force alone is unlikely to be successful, it is hard to be certain. Most problems have not had industrial-scale resources brought to bear on them.

Verifier misspecification. A verifier may accept an AI solution even though the solution does not represent the broader conceptual achievement that the contributing mathematician intended the verifier to identify. In simple cases these misspecifications will be no more than bugs. In more complex cases, it may be that a solution was harder to verify than the mathematician initially believed. We will report on any such cases and fix the verifiers when possible.

Sample bias. We must contend with the bias introduced by the verifiability criterion. Previous caveats aside, progress on this benchmark really should amount to solving meaningful open math problems. But it may turn out that AI systems are uniquely suited to solving the sort of meaningful open math problems that are amenable to automatic verification. If so, progress on the benchmark may not generalize well to progress in other areas of math.

Originally published on Epoch AI’s website.

Do you know:

• How fast LLM inference prices are falling?

• How fast compute stocks are growing?

• How long it takes to build a GW scale data center?

Find out on our Trends page!

We evaluated GPT-5.2 Pro manually on the ChatGPT website. We did this after encountering timeout issues with the API in our scaffold. We’re working to resolve these issues, but a manual evaluation seemed worthwhile in the meantime.

Prior to this run, 13 problems from Tier 4 had been solved by any model ever. GPT-5.2 Pro solved 11 of those, and 4 more besides. Its total for this run is thus 15/48 (31%), and the pass@the-kitchen-sink for Tier 4 (all problems solved ever) is now 17/48 (35%).

OpenAI has exclusive access to 28 Tier 4 problems and their solutions, with Epoch holding out the other 20 problems. GPT-5.2 Pro solved 5 (18%) of the non-held-out set and 10 (50%) of the held-out set. In other words: no evidence of over-fitting.

During evaluation we found issues with two problems. GPT-5.2 Pro and GPT-5.2 (high) should have been credited with solving both, and GPT-5.2 (xhigh), (medium), and GPT-5 Pro should have been credited with solving one. We’ve fixed the issues and updated the scores on our hub.

One of the newly-solved problems was from Joel Hass, whose research is in low-dimensional topology and geometry. Afterward, he suggested we try a different, more challenging formulation of the same problem. GPT-5.2 Pro solved that too.

Another problem, by number theorist Ken Ono was initially solved by GPT-5.2 (xhigh), and also by GPT-5.2 Pro. Ken gave the solution a generally favorable review, though noted that the rigor of its prose explanation was somewhat lacking.

Another newly-solved problem was by number theorist Dan Romik. He was impressed.

A pair of problems by Jay Pantone, who works in analytic combinatorics, were solved earlier: one by GPT-5 and one by GPT-5.1. GPT-5.2 Pro solved these as well. The solutions were both valid, but Jay noted that both used numerical shortcuts that he didn’t intend.

What remains unsolved? One author thinks models get his problem wrong because they make a plausible assumption without trying to prove it. If they tried to prove it — as he had to, when he encountered the problem in his own research — they might realize the truth is more subtle.

Check our website for more info on FrontierMath and analysis of AI math capabilities!

We'll be posting more of our results over the next few weeks, but you can also view everything on our website right now.

The AI industry is scaling exponentially — investment, compute, data center buildouts. So, it turns out, is demand for making sense of it all.

See how we’ve kept up!

Have a look at our 2025 Impact Report, which includes our plans for 2026!

We are fundraising $3M to execute a more ambitious version of our plans. Interested in supporting us? Donations of any size are appreciated.

If you have questions, our DM’s are open. You can learn more and donate through our Donate page or email donate@epoch.ai.

Discussions about AI progress increasingly hinge on computing capacity – aka compute – which is essential in order to develop, train, and deploy AI systems. But public data on the total capacity of AI computing hardware can be fragmented and incomplete.

To address this, we are releasing a new AI Chip Sales data explorer, estimating and visualizing both the number and capacity of AI accelerators that have been sold or delivered in recent years. We leverage data and evidence from earnings reports, company disclosures, analyst coverage, and media reporting to produce estimates of AI chip counts across major vendors: Nvidia, Google, Amazon, AMD, and Huawei, broken down by AI chip model.

We believe this release provides the most complete publicly available picture to date on the global stock of AI compute.

Compute

We find that cumulative global AI compute capacity has reached the equivalent of more than 15 million Nvidia H100 GPUs, measured using each chip’s respective peak specifications in 8-bit operations per second.

Last year also saw major transitions due to new chip generations. Notably, Nvidia’s new Blackwell generation has largely displaced the H100 and H200, with the new B300 alone now accounting for the majority of AI compute capacity sold by Nvidia.

Costs and power

Acquiring and deploying these chips requires massive resources.

Overall, the cost to purchase these chips, even before auxiliary capital costs such as networking and data center construction, has rapidly escalated to tens of billions of dollars per quarter.

AI chips also demand a lot of electrical power. Even before accounting for the power overheads of servers and data centers, the total quantity of chips we track would draw over 10 GW of power. This is around twice the average power consumption of New York City.

We hope the AI Chip Sales data explorer serves as a useful tool to understand trends in global compute. To see our full datasets and interactive visualizations, visit the explorer here!

Our Data Insights offer short, visual, self-contained investigations of key trends and metrics in AI.

Gradient Updates is our outlet for leading-edge commentary by specific authors (also offered as a newsletter on Substack), without necessarily representing the views of Epoch AI as a whole.

Below we bring you our top 10 most popular Data Insights and Gradient Updates in 2025.1

Most popular Data Insights

LLM inference prices have fallen rapidly but unequally across tasks

🠊 In short: Between April 2023 and March 2025, we saw a >10x and larger drop in the price per token at an equivalent performance level.

🠊 Why this matters: API cost reductions indicate a more competitive market and large gains in efficiency, making AI more affordable to customers. If these trends hold up, any AI capability that exists today will soon be available for very cheap!

Learn more

Frontier AI performance becomes accessible on consumer hardware within a year

🠊 In short: The best open models that can run on a consumer GPU lag the frontier of AI by only a year or less, as measured by several metrics of performance including GPQA, MMLU, AA Intelligence, and LMArena.

🠊 Why this matters: Open models that can run on a personal computer are accessible for billions of people. The relatively small lead held by frontier models suggests it would be hard to maintain a market advantage with a fixed level of model capabilities. Instead, companies face pressure to continue developing better models or excel at other services such as better integrations. This trend has direct implications for AI policy. Any capabilities appearing at the frontier are likely to be widely available and unrestricted in less than a year, complicating regulatory options.

Learn more

Most of OpenAI’s 2024 compute went to experiments

🠊 In short: Media reporting indicates that most of OpenAI’s compute fleet in 2024 wasn’t used for inference or training runs, but rather to run experiments that enable further development.

🠊 Why this matters: AI development is capital-intensive. So we should expect leaders in the field to have access to large amounts of compute that they use for experiments. This finding also suggests that most of the costs associated with AI currently are associated with experiments, rather than their direct training and deployment.

Learn more

The stock of computing power from NVIDIA chips is doubling every 10 months

🠊 In short: The amount of installed AI compute from NVIDIA chips has more than doubled annually since 2020. New flagship chips account for most of the existing compute within three years of their release.

🠊 Why this matters: Compute is a fundamental input for AI development and deployment. Exponentially more computational resources are needed to maintain the current pace of AI development, which has driven sustained demand for chips from NVIDIA and other manufacturers.

Learn more

GPT-5 and GPT-4 were both major leaps in benchmarks from the previous generation

🠊 In short: Both GPT-4 and GPT-5 greatly exceeded the performance of their direct predecessors.

🠊 Why this matters: When GPT-5 released, some were disappointed that the performance improvements were only marginal compared to existing models. But this is better explained by a more frequent cadence of releases in the last two years, rather than a slowdown in capabilities.

Learn more

Most popular Gradient Updates

How much energy does ChatGPT use?

🠊 In short: Josh estimated the average energy cost of a GPT-4o query, finding it was less than running a lightbulb for five minutes. His estimate was later corroborated by Sam Altman, and it is also similar to the energy cost per prompt for Gemini reported by Google.

🠊 Why this matters: Many are concerned about AI’s energy use. This piece helped quantify the costs and put these concerns in context, showing that AI energy use at the time was not very significant compared to other household activities. AI energy use has continued to grow exponentially since then, though, and it might become a larger issue in the future.

Read the post

How has DeepSeek improved the Transformer architecture?

🠊 In short: This post covered three techniques introduced by the team behind theDeepSeek v3 paper that allowed them to release the best open-source pretrained model at the time, while using 10× less compute than the next best open model Llama 3. The techniques are multi-head latent attention (MLA), innovations on the mixture-of-experts (MoE) architecture, and multi-token prediction.

🠊 Why this matters: Three days after we published this post, DeepSeek attracted far wider attention by releasing a reasoning model, R1. This model matched the performance of OpenAI’s o1 while using what we presume is a fraction of the development cost. The innovations they introduced illustrate the patterns of training compute efficiency, whereby year-to-year models become 3× cheaper to develop because of new training techniques and data improvements.

Read the post

How far can reasoning models scale?

🠊 In short: Josh discussed the growth in compute for RL reasoning training. Labs like OpenAI and Anthropic claimed in early 2025 that their rate of RL scaling couldn’t be sustained for more than 1-2 years, since it would quickly run into the limits of their compute infrastructure.

🠊 Why this matters: Reasoning has become a highly important axis for scaling model training, leading to excellent results in math, software engineering, and elsewhere. The limits to its growth suggest that the exceptional growth in capabilities during 2024 and 2025 could soon slow down.

Read the post

How big could an “AI Manhattan Project” get?

🠊 In short: Arden and Anson estimated how large a national US AI project could get, comparing it to the relative spending during the Manhattan Project and the Apollo program. They conclude it could result in a training run 10,000x larger than GPT-4.

🠊 Why this matters: In November 2024, the US-China Economic and Security Review Commission’s top recommendation to Congress was to “establish and fund a Manhattan Project-like program dedicated to racing to and acquiring an Artificial General Intelligence capability.” This exercise puts into context the potential size of such a national project.

Read the post

Most AI value will come from broad automation, not from R&D

🠊 In short: Ege and Matthew argued that most of the value AI will create will be mediated by its ability to automate many tasks across the economy, not its ability to speed up R&D. Relatedly, R&D activities have arguably contributed only a modest amount to productivity growth in the 1988-2020 period.

🠊 Why this matters: Many stories about explosive growth from AI, such as those posed by Sam Altman, Denis Hassabis and Dario Amodei, posit automation of R&D as a key lever of growth. This suggests that the impact of AI might be rapid, salient and localized.AI might suddenly automate the last hurdles to R&D automation and quickly make great advances within AI companies. But AI might instead primarily affect society through a diffuse and gradual process that lasts several years or decades as different orgs adopt AI to improve efficiency.

Read the post

We started our Data Insights and Gradient Updates programs so we could offer timely input related to ongoing developments.

The rapid uptake has been gratifying. On our website alone, these new formats amassed nearly half a million views and a total engagement time of nearly 5,000 hours while related posts on Twitter had over 6 million impressions with over a quarter million engagements. They have been used widely to inform the discourse on AI.

If you found any of our outputs helpful this past year, please take our 2025 Epoch AI Impact Survey! Next year we will bring the same passion to improving how we pursue our mission of informing the world about AI trends.

Happy holidays!

Simple QA

SimpleQA Verified consists of 1,000 fact-seeking questions, developed by Google and built on prior work from OpenAI. Here’s a question that all models we benchmarked got wrong.

“What was the name of the man who served as Reeve of North Red Deer, Alberta, between 1924 and 1925?”

Google maintains a nice leaderboard, but with ours we can expand the set of models covered. For instance, we find Qwen3-Max-Instruct to have held the high score (67%), until Gemini 3 Pro recently surpassed it (73%). Also notable: GPT-5/5.1 didn’t improve on o3’s score.

Qwen3-Max-Instruct’s score may be due in part to data contamination. We consider this possible because Qwen3-Max-Instruct ranks much lower on AA-Omniscience, a closed benchmark broadly similar to SimpleQA Verified.

Chess Puzzles

The Chess Puzzles benchmark consists of 100 novel chess puzzles we generated programmatically using a chess engine. We hope it will function as a “lite” version of more involved gaming benchmarks, measuring aspects of spatial reasoning and planning.

All puzzles have a single best next move which models must find. Our subjective assessment is that the puzzles range in difficulty from somewhat straightforward to somewhat difficult for a strong amateur. Here is one example of each: easier left, harder right; white to move.

Models are far from saturating the benchmark: the current high score is GPT-5 (high) with 37%, with other frontier models scoring similarly.

Check out our benchmarking hub for all of this and more!

Watch the webinar here:

Transcript

Introduction [00:00:00]

Maria

All right. Thank you everyone for joining this webinar. We’re going to be speaking about the Frontier Data Centers hub that we released recently at Epoch. My name is Maria and I want to introduce our speakers today. Yafah Edelman is head of data at Epoch. She led this project, and she’ll share a little bit about what motivated it and her plans for it. We’ll also hear from Ben Cottier, who’s a researcher at Epoch. He will be going over some insights from the research and doing a walkthrough of the hub and our methodology. I’ll hand it over to Yafah. You can get started.

Yafah

Thanks, everyone, for coming. We’ll be talking a bit about our Frontier Data Centers hub. I’m excited to share this with you. I think this is one of a very exciting product that Epoch has been working on, and we’re very happy to be able to explain everything about it to the public, just like all of our methodology and everything, because we think that’s important for people to understand.

Our Frontier Data Hub is tracking these frontier data centers. These are the very largest data centers that are being constructed right now. Many of them cost $10 billion or more. We think that these are incredibly important to track, because these infrastructure projects are some of the largest in history, with the largest rivaling the Manhattan Project, similar in size to the Manhattan Project or other very large outlays of R&D expenditure. The largest we’re tracking right now, we expect to cost $100 billion when it comes online in 2028.

Tracking these data centers lets us understand the distribution of compute among companies and nations, as well as the trajectory of AI, how much investment there is in it, the extent to which the investment is increasing or decreasing, and the extent to which companies are able to provide the compute that they plan on and continue scaling at the historic rates.

Our database is entirely free and open, as I mentioned earlier. There are a few data center trackers out there. Ours being free and open means that the public can actually understand. And hopefully all of you can understand a bit about how we do this. We’re committed to transparency. We also share our methodology in even more detail on our website.

In addition to this commitment to transparency, we also plan to expand our coverage beyond the US to China, the Middle East, and Europe, something that we believe many other databases aren’t doing. This will allow people to get a sense of the global politics and status of AI, rather than just understanding what’s happening in the US. For more details on this, I’ll hand the presentation over to my colleague Ben, who will tell you a bit more about our results and methodology.

Satellite explorer demo [00:02:59]

Ben

Thank you. So I’m going to jump into the data hub on our website and show you some of that. I’m going to follow this link. Feel free to hop on to our website yourself and play around. I’m just going to switch what I’m presenting here.

All right. So here we are on the website. First thing I’m going to do, which you can see at the top here, is go to our satellite explorer. So I’ll follow this link. You can see this opens up a map where there are some points marked all over the map of the US. I can click on them and I can see different data centers like OpenAI Stargate, Abilene. These data centers are also listed on the left-hand side. You can see some key details about them, like the current power capacity, who owns the data center, who’s likely using the data centers, and so on.

I can navigate like Google Maps here and zoom in and out. I’m going to pick one example here. Let’s say I’m interested in this Google Omaha data center. I’m going to click here. This opens up a satellite view where we have a high-resolution satellite image of this data center. You can see that this image is annotated. So we have buildings which are numbered 1, 2, 3 and outlined in green. We’ve got grid power annotated here. These are substations. We also have backup power in purple. I can zoom in and see there’s a row of backup generators here next to the building.

Finally, we have cooling equipment which is outlined in blue. If I zoom in there, I can see a row of cooling units with some large fans on them. I’ll come back to that in a moment. But I want to explore this a bit more. So let’s look at the timeline at the top here. This gives you a timeline of satellite imagery so I can go back in time. Here we have another image. You can see that building three is a little less constructed than it was in the other image. I can go further back, all the way to when there wasn’t any building there. So this gives you a little insight into how we figure out the timeline of data center construction and when different parts of it go online.

You can also see this quantitatively. If I open up the graph view on the left here. This shows the power capacity for this data center which is operational at each point in time. We also show the compute capacity and the capital cost of the construction and GPUs and everything involved in this data center project.

Methodology: estimating power capacity [00:06:30]

Ben

So how do we do this? How do we estimate things like power capacity? Well, in this case we rely a lot on the cooling equipment. Coming back to that, if I zoom in here I have a nice clear image of these cooling towers. We developed a model where we can take some feature, some visual feature, like the diameter of these fans and actually predict quite accurately how much cooling capacity, how much heat is being taken out of the building by these cooling towers.

In this case these fans are roughly five meters in diameter. We can plug that into a formula that we’ve developed based on empirical data about cooling towers. We can estimate that each one of these fan units is cooling about eight megawatts worth of heat. Eight megawatts, an average US household is about one kilowatt. So eight megawatts is enough energy to power thousands of US homes just in this one cooling tower.

That’s how we get the capacity of a cooling tower, and then we can just add up the number of fans here to estimate a total cooling capacity for this building. Then we need to make a couple of adjustments to account for redundancy. For the sake of reliability, they don’t necessarily run all of these cooling towers at once. They keep a couple spare. Also there’s some overhead involved in cooling the GPUs or TPUs inside the building. After accounting for that, we can get a pretty good estimate of the amount of power in this building which is powering the GPUs or TPUs.

In this case, we actually have a ground truth for this. We found a permit. This is one of our key sources which is shown here on the left in the notes. In this permit it actually stated the power capacity of these cooling towers and how many cooling towers there were. This capacity that they stated actually lined up almost perfectly with what we estimated using our model. It was within 2%.

We’re not always quite this accurate, but we’re fairly confident that our estimates are within plus or minus 50% of the actual value at any point in time.

Graph view and key insights [00:09:25]

Ben

That’s a look at the satellite explorer and a bit of our methodology. At this point, I might want to not just look at one data center but compare multiple data centers on one graph. So that’s where the main page comes in. I’m going to head back there. If I scroll down we have this graph view which is showing the timeline of power capacity that’s operational for a few different data centers. I can play around with this. I can select Google Omaha, which we were just looking at on the satellite view, and other data centers like Microsoft Fairwater. I can also change the y-axis to look at not just power but compute capacity and cost. So that’s the graph view. We use that to do some of our analysis.

That’s a quick demo of the data hub. Now I’m going to head back to the slides and talk about some of the key insights that we’ve gained from this research.

The first thing I want to give you an intuition for is just the sheer scale of the data centers that are getting built right now. The largest one that we found is Meta Hyperion in Louisiana, and this is due to be operational in 2028, but it’s already under construction and we’ve seen it on the satellite images. The total amount of land for this data center campus would be enough to cover a large fraction of Manhattan, about one fifth. It’s several times bigger than Central Park, which you can see in green in this satellite image here.

We also found that several gigawatt-scale data centers are due to come online next year, in 2026, and several of these are on track to be built in two years or less, going from starting construction to reaching one gigawatt operational. Particularly notable here is xAI Colossus 2, which is expected to go from construction to one gigawatt in just one year. So xAI is particularly fast-paced here.

Then looking further out into the future, looking at the future plans, we expect this growth to just keep continuing. So going from one gigawatt in 2026, up to around three gigawatts towards the end of 2027 and 2028. You can see here Microsoft Fairwater, Wisconsin. That one is reaching around three gigawatts in 2027. We estimate that this would give it a capacity of around 5 million H100 equivalent GPUs in terms of computing power.

Those are a few key insights. Now I’m going to hand it back to Yafah to talk about our future plans a little bit.

Future plans [00:12:57]

Yafah

We have a lot to do. I’m excited to talk a little bit about what we’re going to be adding and what additional things the hub is going to cover in the future. Our most exciting, or one of the most exciting first things we’re going to be doing is increasing coverage. This means trying to get up from our current coverage. Currently, we cover about 15% of AI compute by our estimates. We want to get that up quite a lot, as high as 90% by covering more of the largest data centers in the US and also expanding around the world. This means covering data centers in the US and China, in the Middle East and in Europe. Here we have a picture of a data center in the UAE. We’re working on this and should hopefully have some updates for everyone pretty soon. I’m really optimistic about how we’re going to expand our coverage.

Another major update that’s coming is we’re working on tracking more metrics about data centers. An example here is the number of construction workers, which are often reported in permits. We also might be able to track similar numbers based off the number of cars in parking lots, which we have in our satellite photographs. We are also hoping to track which data centers are networked together. This will let us understand how big the full scope of the potential clusters that might exist beyond just individual campuses.

These steps and more are coming in the next few months. There are other things we’ll also be excited to share with everyone beyond these, but for now, I’ll pass it over to Maria to set up our Q&A.

Q&A session [00:14:44]

Maria

Sure. So now we’ll be moving on to the Q&A. Thank you, Ben and Yafah. All right, so we have this one by Marco De DiDonato. Speaking about power generation, it seems that many large data centers you are mapping still rely on local grid to power themselves. Others like Colossus and Stargate 1 also rely on onsite gas turbines. What trend can we expect to see in the coming years? Can we realistically think that grid operators will be able to keep up with multi-gigawatt scale projects? Or will hyperscalers focus on local power generation going forward? That’s a lot. Maybe let’s start with what trend can we expect to see in the coming years.

Yafah

What we’ve seen a lot of is data centers providing onsite power as a bridge before they get grid power. I believe this happened at Stargate Abilene. This also happened with Colossus. I expect similar techniques to be used. As of now, my current impression is that all data centers plan to eventually add grid power. I have seen grids keeping up more with power than I think most people are under the impression. I’ve been seeing grids keep up more with adding capacity than I think most people are under the impression. That’s my answer there.

Maria

Now with Anton Trubnikov. Maybe I missed this, but how are you identifying sites in the first place? Many data centers and campuses are not publicly announced.

Ben

I can take that. We identify sites through multiple methods. One thing is we just look at news or rumors about data centers. There’s often local news, which will talk about some new construction of a data center which may not be widely publicized, but you can find it in local news. But there are also websites that are quite good at tracking lots of news about data centers, which we follow, and then we go and investigate further. So we might have just a city location for a data center. We don’t know the exact location, but then we might go into Google Earth and look around. Maybe we happen to find some new construction, and then we end up determining that that’s actually the data center in question.

We also look at permit documents. We use ChatGPT. Actually, this is quite helpful as a tool to search quite deeply for permit databases and find permits, which often mention the address of the site. So we can narrow down the location that way.

Going forward, we’re interested in doing more automated methods of discovery using satellite imagery. But so far we’ve just been using a combination of news and permit documents and satellite imagery, manually searching satellite imagery.

Attendee

I saw that it looks like there’s going to be more data on China, which is exciting. What are the challenges to gathering that data and how are you going to get around them?

Yafah

There are a few challenges here. One is that we are not nearly as optimistic about the potential for permits and similar filings to cover data centers as well. I also expect there to be a lot more noise with people claiming buildouts that don’t exist. However, for our first steps in this, we’ve looked through some sources, located a data center, and it turned out to be half a gigawatt, which is very large. It is an older data center, but there is some amount of new buildout. So to some extent, I think some of the similar sources, looking for new sources, people talking online, is at least promising for first steps.

For additional steps, we are excited to hopefully use low-resolution satellite imagery to automatically search for very large construction sites, which we can then do more research on to discover whether there are data centers. We’re also excited to investigate other avenues for similar plans. Additionally, our satellite methodology we expect to largely work in China, allowing us to, once we have found a potential site for a data center, do an analysis very similar to our current one.

Attendee

First of all, thank you for doing this, all of you. How will you go about determining which data centers communicate for decentralized computing?

Yafah

We’re still at early stages of this. One thing I’m optimistic about is seeing the infrastructure being built for fibers, looking at permits and looking at prior coverage of this sort of thing. This is something that we are eager to do but have not currently done.

Ben

Just to add to that, sometimes the companies do just say that the data centers are connected. For instance, the recent announcement from Microsoft of the second Fairwater Data Center. One’s in Atlanta, one is in Mount Pleasant, Wisconsin. They say that these data centers are connected by fiber. I think some claims may be more vague than others. We’d want to verify it as best we can. But in the absence of other information, we would just take a company’s word for it that data centers are in fact connected or planned to be connected. We indicate our confidence level in these—whether or not they are connected or other things, like who is using the data center. We have a confidence level that we indicate just to give people a sense of, are we confident in this? Or is this more like a rumor?

Maria

Yes. I see a question here from Yan Riviere. Earlier you talked about the gas turbines used before going to the grid. Do you think the turbines may be moved from one data center to another while waiting for the grid to be linked? I’m not sure if that was what you were implying.

Yafah

There are two instances of this I specifically refer to, one of which is xAI Colossus, which I believe rented portable gas turbines to power their data center before it was connected to the grid. Those would have been moved somewhere else afterward. Those are, I believe, typically used for disaster areas and providing backup power for them.

The other example is Abilene. Abilene is near a large amount of wind power, which is intermittent and thus needs a lot of backup power regardless. So I believe their plan, as stated, is to connect to the grid, which I believe they have already done. Ben can correct me if I’m wrong on that. And use the gas turbines as backup for the grid. The gas turbines there are permanent.

Ben

There’s I believe 200 megawatts worth of grid power already supplying Stargate in Abilene, and they’re working on a one gigawatt additional substation to fully power the final eight buildings.

Maria

Great. We have another question here from Ashby Field. They say I missed most of the presentation, so I apologize if it was covered. How do you derive the speculative users of the center based on map data? Will you track policy, tax incentivization, any location-based grid intensity datasets?

Ben

In terms of the speculative uses, this isn’t just based on map data. We look at announcements or permit documents for mentions of this. Sometimes permit documents will be filed under a shell company, but we’re able to track that back to the actual company. Sometimes if it’s speculative, that usually means we have some reason to believe that a certain company is using the data center, but this hasn’t been confirmed in any way.

We actually updated our estimate for this. There’s a data center called Goodnight, where we initially thought it was speculatively for OpenAI because the buildings had the same design as Stargate. But we’ve since heard that Google is actually investing in a data center in that location. So we switched to thinking that it’s more likely to be used by Google. So we’re updating our database based on the best evidence we have.

Yafah

We use tax abatements in some cases for learning about data centers. We expect to continue using this as one of many data sources. Some of those documents will be linked in the notes section of the satellite view of our data hub.

Any location-based grid intensity datasets—we do not have any location-based grid intensity datasets that we are releasing, that we have. We use some grid information that’s similar to that sometimes, but nothing substantial that we’ve created on our own, beyond the amount of power data centers use, which is obviously location-based, but we do track.

Maria

Okay. Here’s another question by Edward Kant. Are there any legal challenges in collecting and publishing such information? Do you want to take this or should I?

Yafah

I’ll take this. The answer is largely no. There is one particular thing which we had to deal with, which was getting proper licensing for some of the images. We did this, but there are no other legal challenges involved. Satellite data is—I know it surprises some people—satellite data is publishable.

Maria

There’s this new one by Marco DiDonato. I think it could be interesting to track the overall supply, global supply of gas turbines that could be used on data centers. There are actually only three major gas turbine suppliers in the world. It could help to understand what is the actual hard cap on the amount of power that can be deployed.

Yafah

This is definitely an interesting methodology. There’s one question you can ask, which is, to what extent are there enough traditional gas turbines to power these data centers? However, there are other alternative sources which we believe will provide power should these data centers not be able to get gas turbines. Gas turbines are often the most cost-effective power source, especially combined-cycle ones, but there are several alternatives.

In particular, a data center can have a flexible load where they go down during peak times, which makes it much easier for grids to supply them. A data center can also use solar and batteries, which we believe is very scalable with fairly small lead times, although it is more expensive than traditional power, but still not nearly as expensive as the GPUs which will go into that data center. So we expect that if necessary, AI companies will start using solar plus batteries.

Additionally, there are alternative types of gas turbines which you might not expect. For instance, some places are taking gas turbines out of jets, out of planes, and using them to power data centers. Something which I think is very cool. I’m not exactly sure how common that is yet. It’s probably more expensive because those types of gas turbines are simple-cycle or single-cycle turbines, which are less efficient.

Additional discussion on power [00:27:41]

Yafah

One thing I want to repeat, because I think it’s a thing that a lot of people are wrong about. There are challenges with power. Infrastructure in the US is designed for demand for power which has largely plateaued, which happens in developed countries as they start using power more efficiently rather than adding additional power. This means that it has designed for a situation where you can plan very far in advance. Loads aren’t increasing by a lot, and you thus care a lot about keeping costs down as much as possible. This sort of slow planning out years in advance, providing a large amount to keep costs down strategy is—

I think a lot of the worry that people have about data centers not being able to get enough power is really not going to pan out, because data centers can absolutely afford to pay more and will afford to pay, will pay more if necessary, just because of how much GPUs are still going to be a much larger cost than power. At the end of the day, if they have to pay twice as much to get the power quicker, they just will.

Maria

I see a new question. You mentioned jobs, which is obviously super important to local jurisdictions and governments considering data centers as economic stimulus. Are you collecting data for buildouts and perpetual operations separately? Nevertheless, this is great work. Do you want to take this?

Yafah

Right now we are not publishing job data, I believe. But when we add jobs to the hub, I expect to separate out these two sources. If anyone’s interested in my current impression, somewhere in the vicinity of perpetual operations requiring maybe a tenth to a fifth of the amount of people as the actual construction. Although this is a very preliminary, I’m sort of eyeballing it number, and I hope to have a lot more detail to share in the future.

Identifying sites and construction companies [00:30:46]

Yafah

While we’re waiting for anyone else to offer questions, I’m going to return to one. I think there’s some interesting additional information. Anton asked, how are we identifying sites? One thing I’m excited to dig more into is construction companies will often discuss the sites they are working on in their earnings reports or on their website. We’ve found some pretty interesting sources using this, and I’m excited to continue using this to discover data centers which have not been formally announced.

Very often, “formally announced” is not a very clear definition. You will often have a data center where the local cities all know about it and it’s in their newspapers. But the company, like Microsoft, has not announced it in a big PR campaign. So there are these intermediate things that are also relevant.

Maria

I think one question that is on the website and that I found interesting is whether an entire data center can be used to train one AI model. Could you explain that?

Yafah

Typically, we are under the impression that most of the new data centers could, in theory, be used to train one AI model. However, contrary to what some people are implying, these companies imply, we believe that this happens pretty rarely, and that in fact, the largest AI models that are currently being trained are being trained on a fraction of the data center.

This is also part of why, when people talk about being able to connect these data centers and do distributed training, I mean, I’m interested in this and I think it’s important to track, but I think it is not actually necessary for scaling up in the present. There’s a lot of room just to use a larger portion of the data centers that already exist.

So when Google or when Microsoft talks about how we have a data center in this state and this state and they’re connected to a super whatever, I’m like, well, you’ve never used even a fifth of one of those data centers to train a model. So you’re getting ahead of yourselves. People should maybe on the margin pay a little bit less attention to this, although it’s still a thing we’ll be tracking because we think it could be important. It’s interesting and important to keep track of, if they wanted to, how big can they get? But this might not, in practice, be what they’re actually using.

Maria

Thank you. Okay, I think we’ve run out of questions so we can probably leave it here. Thanks so much, everyone for joining, for asking questions, and Ben for presenting. You can email us if you have more questions. We’ll be answering, or feedback, for example, things you want to see in the future. You can explore the data center hub at epoch.ai. You’ll see several databases. This is one of them. Thank you so much. Have a great rest of your day.

Ben

Thanks everyone.

You can now examine annotated, recent, high-resolution satellite imagery of the world’s largest compute clusters directly from your phone at https://epoch.ai/data/data-centers.

Here’s a look at the updated Satellite Viewer: