Key takeaways

• Substantial quantities of AI chips have been sent to China in violation of US export controls. Evidence of diverted or missing chips, drawn from indictments and investigative reporting, points to nearly 300,000 Nvidia H100-equivalents by the end of 2025. This would equal roughly a quarter of the compute China acquired through legal channels or domestic production. Because much smuggling goes undetected, the true total is likely higher.

• We estimate, with 90% confidence, that between 290,000 and 1.6 million H100-equivalents of compute were smuggled through the end of 2025. Our median estimate of 660,000 represents roughly 3% of the global compute stockpile, comparable to what xAI, a leading US AI lab, had at the time. The upper bound of our estimate would mean that, by the end of 2025, the majority of China’s AI compute had been smuggled.

• We are uncertain about many variables, notably the magnitude of undetected smuggling and the proportion of chips allegedly diverted or missing that ultimately reached China.

Subscribe now

Overview

Over the past several years, the United States has applied multiple rounds of compute-related export controls to China. The earliest rounds of export controls applied to semiconductor manufacturing equipment or to specific entities (e.g., Huawei’s addition to the BIS entity list in 2019). In October 2022, the Commerce Department instituted export controls on all advanced chips with certain characteristics. Commerce tightened these export controls and expanded their scope in October 2023, increasing the incentives for smuggling chips to China. A late 2023 Center for New American Security (CNAS) analysis suggested that the volume of smuggled chips at the time was relatively low, but expected that there would be larger-scale smuggling efforts in the future.

This prediction appears to have been borne out. Investigative reporting by the New York Times, Wall Street Journal, and The Information in 2024 found widespread evidence of banned chips being smuggled into China. In April 2025, Commerce imposed an additional licensing requirement on H20 chips, a type of chip popular in China that Nvidia had developed to comply with prior export controls.1 This further restricted exports and incentivized smuggling. The Financial Times reported that over $1 billion in AI chips were smuggled between April and July.2

Evidence of large-scale diversion by cloud companies has also accumulated. The DOJ indicted several Supermicro employees and arrested a company executive. The employees were accused of diverting $2.5 billion in Nvidia chips to a pass-through entity, which was redirecting those servers to China. Bloomberg investigative reporting insinuates that Megaspeed, a Singaporean cloud provider, has tens of thousands of fewer chips in its Malaysian data centers than it imported from Nvidia, and its CEO is under scrutiny due to close ties to China.

We estimate the total compute smuggled to China using Monte Carlo simulations3 based on evidence from indictments and investigative reporting. We classify this evidence into two types. Diversion evidence traces how chips leave legitimate supply chains and reach smugglers. Resale evidence focuses on the marketplace for smuggled chips within China: the number of vendors or brokers and volumes they transact. Some sources cover both diversion and resale of the same set of chips.

Because every successfully smuggled chip must be both diverted and resold4, we can build two parallel estimates of the same underlying quantity that should, in theory, approximately agree. While our resulting distributions for diversion and resale aren’t identical, they do substantially overlap.

The diversion estimate is created by building a database of allegations from 2024 and 2025, and then adjusting for false or overstated allegations, and for cases missed by reporting and enforcement5. The database does not include cases where diversion was attempted but known to have failed. The resale estimate approximates total volume in the market for smuggled compute by aggregating investigative reporting on vendor counts and the scale of individual transactions. The resale estimate draws on portions of a 2024 CNAS estimate of smuggled compute by Grunewald and Fist.

Cumulative allegations of diverted or missing chips total almost 300,000 H100e6 by the end of 2025 — a quarter of the compute that China is estimated to have legally imported or domestically produced. This figure is neither a ceiling nor a floor on actual smuggling. This figure may overstate actual smuggling for two primary reasons: the allegations are largely not yet proven in a court of law, and some of the chips redirected to China may not have arrived there. Conversely, the 300,000 estimate captures only publicly alleged cases — some, probably most, smuggling goes undetected by journalists and authorities. Thus, we believe it is almost certain that the true total quantity of compute smuggled into China is higher than 300,000. More details and a cumulative summary of reported and alleged smuggling are here.

Our model estimates 660,000 (90% CI: 290,000–1.6 million) H100e of cumulative compute smuggled through the end of 2025. This is somewhat higher than other estimates, including those by CNAS and ChinaTalk, though the median is well within their confidence intervals and vice versa. A comparison against these other estimates, including a discussion of the differences, is available here. Further details on the model and our estimates are here. Our estimates cover quantities of chips and servers physically moved into China. They do not cover cloud compute located outside of China but used by Chinese customers, an arrangement that, subject to end-use restrictions and sanctions-list screening, is generally permitted under current US export controls. These estimates can be viewed in a global context by checking out Epoch’s AI Chip Ownership Hub.

Read the full report for interactive charts, a breakdown of our methodology and evidence, and a comparison to other estimates.

Subscribe now

Updated April 23, 2026

Introduction

The United States is in the middle of an unprecedented build-out of AI infrastructure. No project illustrates the scale of that effort more than Stargate, a $500 billion endeavor involving AI developer OpenAI, cloud provider Oracle, and investment company SoftBank.

Stargate has seven locations across the US, all of which are now showing active development. The most advanced—in Abilene, Texas—is already operating at an estimated capacity of 0.3 gigawatts (GW).1 The six other sites include two more in Texas, as well as facilities in New Mexico, Wisconsin, Michigan, and Ohio. Together, the seven sites add up to over 9 GW of planned capacity, which is comparable to the peak power demand of New York City.2 This will be enough to power the equivalent of 20 million Nvidia H100 GPUs, which was the total amount of AI compute in the world by the end of 2025.3

Stargate’s design choices reveal how builders are navigating the key challenges of gigawatt-scale AI data centers in the US. To sidestep lengthy queues for connecting to energy grids, at least three of the seven sites will make use of on-site natural gas plants. To address public concerns about water usage, at least six sites will use closed-loop liquid cooling systems, which do not evaporate water.4 These decisions will likely save the project time but raise the cost of the facilities.

Based on announcements from 2025, SoftBank will own the hardware at the Milam County and Ohio sites, while Oracle will own the hardware at the remaining sites. All sites will serve OpenAI’s workloads.

The sites

Abilene, Texas

Current capacity: 0.3 GW | 250,000 H100-equivalents5

Projected capacity: 1.2 GW | 1.0 million H100-equivalents

Projected completion: Q4 2026

The Stargate project’s flagship location is in Abilene, Texas. Built by AI infrastructure company Crusoe, Abilene is the most complete Stargate site to date, with an estimated four of the eight buildings already operational. These buildings house state-of-the-art Nvidia Blackwell chips.

Power is currently supplied by a mix of on-site natural gas and grid power, which includes local wind power.

OpenAI had planned to expand this site to 2.1 GW, but recently reversed course, deciding to direct that capacity to other locations. Microsoft has since partnered with Crusoe for the adjacent 900 MW site.

View with satellite explorer

Shackelford County, Texas

Current capacity: 0 GW

Projected capacity: 2 GW | 4.2 million H100-equivalents

Projected completion: Q4 2028

Just across the county line from the Abilene site, data center developer Vantage is constructing a massive 1,200-acre (4.9-square-kilometer) campus with 10 buildings.

The campus will be powered by an onsite natural gas microgrid.

Vantage has given a delivery date for the site’s first building of late 2026.6 Satellite imagery shows that roofing is underway for this building (visible in bright white).

View with satellite explorer

Doña Ana County, New Mexico

Current capacity: 0 GW

Projected capacity: 2.2 GW | 4.6 million H100-equivalents

Projected completion: Q4 2028

In New Mexico, STACK Infrastructure is developing Project Jupiter, which consists of four large buildings. Satellite imagery shows that foundation work is underway.

This site will be powered by two natural gas microgrids designed to limit impact on the local grid.

View with satellite explorer

Milam County, Texas

Current capacity: 0 GW

Projected capacity: 1.2 GW | 2.5 million H100-equivalents

Projected completion: Q4 2028

SoftBank subsidiary SB Energy is building and operating what is described as a “fast-build” site in Milam County, Texas, around 70 miles (110 kilometers) northeast of Austin. A satellite image from March shows steel framing and roofing for the first building (visible as a blue rectangle). Regulatory filings indicate this building will be delivered by October.

SB plans to fund and build new energy generation and storage to supply the majority of the campus’s power.

View with satellite explorer

Port Washington, Wisconsin

Current capacity: 0 GW

Projected capacity: 1.3 GW | 2.6 million H100-equivalents

Projected completion: Q4 2028

Vantage, which is also the developer behind the Shackelford County site, has broken ground on a campus named “Lighthouse” in Port Washington, just north of Milwaukee. Foundation work can be seen in satellite imagery.

The site is described as “sustainable-by-design,” with 70% of power drawn from solar, wind, and battery storage.

View with satellite explorer

Saline Township, Michigan

Current capacity: 0 GW

Projected capacity: 1.4 GW | 2.9 million H100-equivalents

Projected completion: Q4 2028

Related Digital is developing a campus dubbed “The Barn” in Saline Township, southwest of Detroit. Satellite imagery shows foundation work underway for the first building.

DTE Energy will provide 100% of the power, augmented by a battery storage system financed by the project.

View with satellite explorer

Lordstown, Ohio

Current capacity: 0 GW

Projected capacity: <0.3 GW | <0.3 million H100-equivalents

Projected completion: Unknown

The seventh site is in Ohio, where some land has been cleared, but no large-scale data center construction is visible. The site is primarily a manufacturing facility for AI servers and data center equipment, operated as a joint venture between SoftBank and Foxconn. The capacity of the data center will likely be no more than 0.3 GW, with OpenAI announcing that the Milam County and Lordstown sites could scale to a combined 1.5 GW by 2027.

The Lordstown data center will likely draw power from the grid, as the Foxconn plant already has a substation connected.

View with satellite explorer

The road ahead

At this point, the full $500 billion Stargate project is more than pure ambition. The build-out has started all over the US, leaving enough time to finish by 2029. However, there is a long road ahead for all seven sites. Plans can change even after construction begins, as shown by OpenAI pulling out of the Abilene expansion. Financing and procuring equipment will also be challenging at this unprecedented scale. Finally, political opposition is a real factor, as evidenced by a ban on future data centers in Lordstown and local opposition to the Michigan site. Epoch AI will be following the Stargate project and the broader data center build-out closely to see how this all pans out.

AI tools have moved from a niche technology to a part of everyday life. In a new Epoch AI/Ipsos survey of over 2,000 U.S. adults, half reported using AI tools in the past week.

But adoption rate alone does not capture the full picture of how AI is used. Among employed users, it has become a work tool that is already changing the tasks they perform, with substantially higher workplace use among paid subscribers than free-tier users.

AI is now a workplace tool, not just a personal one

Among employed Americans who used AI in the past week, half reported using AI at least as much for work as for personal tasks, while the other half reported primarily personal use.

This aligns with a broader shift toward workplace adoption. Pew Research Center reported in October 2025 that roughly one in five U.S. workers used AI on the job. Though not directly comparable with our survey, our results support the idea that AI is becoming part of routine workflows.

AI both creates and replaces tasks at work

We asked employed AI users whether AI had changed the tasks they perform at work. Among those who use AI at least as much for work as for personal tasks:

27% said AI has replaced some of their existing tasks (task automation). An example of task automation could be an employee using AI to summarize a document they would ordinarily read themselves. Most in this group reported this happened without them taking on new AI-enabled tasks.
21% said they had started doing new tasks because of AI (task augmentation). An example of task augmentation could be a data analysis tasks that would ordinarily require the worker to know how to code. Nearly half in this group reported this happened without any existing tasks being automated away.

We did not collect detailed examples of specific tasks, but these results provide an early, nationally representative snapshot of how AI is reshaping work at the task level.

Work use is higher among paid subscribers

The largest differences in workplace use appear between paid-tier users (such as subscribers to ChatGPT Plus) and free-tier users.

Among employed AI users, 38% of free-tier users reported using AI at least as much for work as for personal tasks. The share rises to 58% among self-paying subscribers and 76% among users with employer-provided subscriptions. As we would expect, paid access, especially when provided by employers, is associated with more intensive workplace use.

It’s unclear whether this is driven by employers or by workers. One possibility is that employers drive adoption by providing access to paid tools, making workplace integration dependent on organizational decisions. Microsoft Copilot, which leads paid AI usage among both work-oriented and personal-oriented users, illustrates this dynamic: its prevalence likely reflects bundling with Microsoft 365, a product widely deployed in workplaces through enterprise licensing.

Another possibility is that workers adopt paid tools in response t work demands after encountering the limits of free-tier access, including cases where frequent users request subscriptions from their employers.

Conclusion

Taken together, our findings show that AI use is now widespread and that its role in work is already substantial for many. While most people still use AI mainly for personal tasks, about half of employed users use it at least as much for work. This share is even higher among those with paid tools, particularly when provided by employers. Broader workforce impacts will ultimately be driven by accumulations of individual changes in how people work, making usage patterns like these an important signal to track.

To learn more about AI usage in the workplace and beyond, visit the Polling on AI Usage explorer to find the full survey results, methodology, and interactive visualizations!

Can AI do real digital work? We reviewed three benchmarks to find out: Remote Labor Index (RLI), GDPval, and APEX-Agents.

Our take: progress here will indicate substantial economic value, but tasks are too self-contained to tell us about wholesale automation.

• Remote Labor Index (RLI) uses real projects from Upwork. It’s mostly multimedia, from interior design renderings to custom musical arrangements. It’s also the most complex of the three, with average human completion time at 29 hours. Models score very low, partly due to the evaluation scaffold.

• APEX-Agents focuses on high-end white collar work. Agents must navigate messy file systems to complete banking, consulting, and legal tasks. Expert humans take 2 hours on average. Current top scores are 30%, and again the scaffold could be improved.

• GDPval has the greatest breadth, drawing from 9 industries and 44 occupations. Humans take 7 hours on average. Models have come the farthest here, with top scores at least in the mid-70s. Tasks here are the most unrealistically “clean” in their presentation.

Think of these benchmarks like SWE-bench Verified. Progress means models can do well-defined digital work tasks mostly autonomously, not that they can function as drop-in replacements.

See the full report for much more!

Subscribe now

OSWorld consists of 361 tasks sourced from forums and tutorials. Models get an Ubuntu VM and task instructions, and write code to interact with the mouse and keyboard.

Here is one task’s starting state. Instructions: “Make a duplicate of the last two slides for me, please.”

Most tasks are realistic but relatively simple, requiring fewer than ten steps (clicks, text inputs, etc.). These tasks take humans only a few minutes to complete.

Contrary to standard practice, OSWorld is updated continuously. A major update in July changed most tasks, and 10% of task instructions have been updated since then. Furthermore, 10% of tasks rely on live data from the web. This makes it hard to compare results over time.

OSWorld is about computer use, but many tasks require little use of graphical user interfaces. About 15% can be solved with only the terminal and a further 30% can rely heavily on Python scripts. We even found cases of models downloading packages to manipulate spreadsheets.

We also noticed ambiguity in several task instructions. While this *is* realistic, it makes interpreting progress on OSWorld harder. If a model improves, is that because it got better at computer use, or because it got better at guessing the intent of the tasks?

Finally, we found about 10% of tasks to have serious errors (though this isn’t an unusual rate). Some tasks have incorrect answer keys, some evaluation functions are too strict or too lax, and some instructions are fatally ambiguous.

Read the full report on our website.

Using underutilized generation as a proxy of power availability, we identify a 4,800 km network of 23 sites in the U.S. that could theoretically support a 10GW distributed AI cluster, helping to alleviate power bottlenecks.

To use such a distributed cluster for training, we consider fully synchronous data parallelism. In this setup, each site will process a part of each training batch, and then synchronize to update model weights in unison. This strategy has been tested before by NVIDIA to train a Nemotron-4 340B model between two datacenters 1,000km apart.

Such a setup will only be practical if the time spent synchronizing is sufficiently small compared to the time it takes to process each batch. The best public work on determining the critical batch size we are aware of is from Bersgma et al. (2025), who posit that the critical batch size increases with roughly the square root of the training dataset size.

For synchronization, we will consider a bidirectional ring all-reduce algorithm. This allows us to complete a synchronization with a single round trip around the network. The synchronization time is then determine by the point-to-point network bandwidth, and bounded by the network latency.

Achieving sufficient bandwidth to keep synchronization time under control sounds daunting. We estimate that to train 72T-parameter models we would need over 25x the bandwidth of the MAREA transoceanic fiber cable — the highest-capacity internet cable crossing the Atlantic. However, the main cost of fiber deployment is installation, so increasing bandwidth is cheap compared to the overall cost of datacenter construction.

The bottom line

Conducting large decentralized training runs is feasible without a large increase in either training time or budget.

However, distributed clusters have many downsides:

• more complex permitting processes

• additional engineering to manage a long-range network connection and reliability

• constraints on communication-heavy paradigms

We expect that AI companies will prefer to scale AI campuses as much as they can, and only resort to distributed clusters to go beyond the scale that utilities are willing to provide through the grid.

Evidence for this is Microsoft’s Fairwater datacenter in Wisconsin, a planned multiple GW site that will allegedly become “part of a global network of Azure AI datacenters”— meant to “enable large-scale distributed training across multiple, geographically diverse Azure regions”.

Find detailed analysis, calculations and sources in the full article on our website.

Note that this is the publicly available version of Deep Think, not the version that achieved a gold medal-equivalent score on the IMO. Google has described the publicly available Deep Think model as a “variation” of the IMO gold model.

Good performance on FrontierMath requires deep background knowledge and precise execution of computations. Deep Think has made progress but hasn’t yet mastered these skills, still scoring lower on the harder tiers of the benchmark.

This version of Deep Think got a bronze medal-equivalent score on the 2025 IMO. We challenged it with two problems from the 2024 IMO that are a bit harder than the hardest problem it solved on the 2025 IMO. It failed to solve either problem even when given ten attempts.

Deep Think approaches geometry problems differently than other LLMs: rather than casting everything in coordinate systems, it works with higher-level concepts. This is how humans prefer to solve geometry problems as well.

We noticed Deep Think making several bibliographic errors, referencing works that either did not exist or did not contain the claimed results. Anecdotally, this was the model’s main weakness compared to other leading models.

We also worked with some professional mathematicians to get their anecdotal impressions. They characterized Deep Think as a broadly helpful research assistant.

Our new report zooms in on two things: (1) whether scaling continues (compute, data, power, capital), and (2) the capabilities this enables—especially for scientific R&D.

Scaling is likely to continue to 2030

Building on our past work, we forecast that:

• Training clusters would cost hundreds of billions of dollars

• Compute scaling is likely "not hitting a wall"

• Synthetic & multimodal data may be needed to ease bottlenecks

• Power demands will increase but be manageable in principle

AI will accelerate scientific R&D across several domains

Given that we expect scaling to continue, what does this mean for the resulting AI capabilities? We focus on scientific R&D—a stated priority for leading labs—and assess both benchmarks and real-world use, allowing us to forecast the kinds of tasks AI will be able to automate.

Despite benchmarks’ weaknesses (e.g., contamination and overfitting), progress has tracked real-world improvements, and usage already supports productivity gains. People already spend billions to use AI for coding, writing, and research.

AI may be a transformative tool well before it can work autonomously.

We explore future capabilities across four domains:

- Software engineering: AI is already transforming software engineering through code assistants and question-answering. On current trends, 2030 AI will autonomously fix issues, implement features, and solve hours-long defined problems.

- Mathematics: AI for mathematics may soon act as a research assistant, fleshing out proof sketches or intuitions. Mathematicians already offer examples of AI being helpful in their work, mostly for studying.

- Biology: Tools like AlphaFold are already revolutionising biology, and will expand to predict more properties for more complex structures. AI assistants for desk research are at an early stage, but offer great promise.

- Weather prediction: AI weather prediction already outperforms traditional methods from hours to weeks ahead. The next challenges lie in further improving predictions, especially rare events, and integrating new data sources.

In summary

At minimum, we expect that AI for scientific R&D will follow in the footsteps of coding assistants for software engineers today, boosting productivity for desk-based research by 10-20%.

Impacts may differ greatly across domains. We might see a flourishing of lab research in molecular biology for years before producing new medicines, because clinical trials and approvals take years.

For more details, check out the full report at our website.

Power demands for frontier AI training have been growing at 2.2x per year, with frontier runs now exceeding 100 MW. The primary factor driving this growth is the scaling of the compute used to train models, at a rate of 4-5x per year.

We examined three key factors that affect power requirements:

• Training compute scaling (likely to continue at 4-5x/year)

• GPU energy efficiency improvements (~40% annually)

• Training run duration growth (likely to slow as runs already exceed 100 days)

Based on these trends, we forecast power demand for frontier training will grow by 2.2x to 2.9x per year. At this rate, individual training runs could reach 4-16 GW by 2030, enough to power millions of home. The lower end aligns with already-announced data center plans. Beyond individual runs, we estimate total AI power capacity could reach >100 GW worldwide and >50 GW in the US by 2030, approaching 5% of the US’s total power generation capacity!

Currently, AI power demand splits roughly equally between training, experiments, and inference. Training will likely retain a large share of demand even if AI usage surges, as long as growth in inference continues to motivate further investments in scaling.

Frontier training is of special interest to the energy sector because it drives the scale of the largest data centers, with concentrated power demands. This could be mitigated by geographically distributed training, but for now companies are still pursuing massive data centers.

To learn more, you can find a summary on our website, or read the full paper at EPRI.

Grok 4 set a new record on high school math competitions with short-form numerical answers. Here is one problem Grok 4 solved that no other model has solved.

Whereas the human reference solution is quite conceptual, Grok 4 approaches it via coordinate geometry and calculus. This “grind it out” style is typical of LLMs.

Such benchmarks are very nearly saturated. Going forward, the most interesting thing we can learn from them is whether models start to solve these problems in more conceptual ways. Grok 4 doesn’t indicate any change on this front.

Grok 4 also improved at proofs. Below is a competition problem which, among all models, Grok 4’s “Heavy” mode comes closest to solving. Its solution is not “from scratch”, however, but relies on extensive background knowledge.

Much headroom remains here. Grok 4 did not shine on the IMO, a more recent proof-based competition. It also seems to be less capable when deep background knowledge is less relevant. Like all LLMs, its attempted proofs can contain hard-to-spot logical flaws, hampering utility.

We also worked with two professional mathematicians to evaluate Grok 4’s capabilities. Both were impressed with Grok 4’s ability to search the mathematical literature, although they emphasized that it should be used with care.

Overall, Grok 4’s progress is more incremental than revolutionary. This suggests that "RL at pretraining scale" does not immediately deliver AI systems capable of doing a mathematician’s job.

See the full report on our website for more insights.

SWE-bench Verified is a set of 500 issues and patches from real python repositories. Each benchmark sample contains:

• The repository before the PR was merged

• The issue description(s)

For evaluation it also contains:

• The original, human-written solution

• A set of tests

The model is put in the position of a maintainer: given the issue description, it has to patch the repository to resolve the issue and make tests pass. This is much more realistic than assessing models on LeetCode-style tasks.

We think a relatively low fraction (~5-10%) of SWE-bench Verified’s issues are incorrect and unsolvable. The issues are all from live codebases Human annotators filtered the issues from a larger subset, SWE-bench. We manually analyzed 40 issues

People sometimes equate SWE-bench Verified scores with general software engineering ability. Instead, SWE-bench Verified predicts whether an AI can fix simple issues (taking at most a couple hours for an SWE to solve) in a codebase.

Furthermore, the low diversity of codebases limits external validity. Django comprises nearly half of all issues and five repositories account for over 80% of the benchmark.

Contamination is a major concern for SWE-bench Verified: since all the issues are sourced from famous repositories, and date back from 2023 and earlier, it’s likely that current models have been trained on these repositories.

Recently, many benchmarks have been developed that build on SWE-bench Verified’s foundation, such as SWE-bench Multimodal, SWE-bench Multilingual, Multi-SWE-bench, and Terminal-Bench. We may review them in future issues.

Read the full report on our website.