There is still much we don’t know about China’s compute capabilities, even though that information matters for decisions the US government is making today. Open questions include how much compute China has, how much it is acquiring, how much it can produce itself, and how it uses what it has. For example:
Epoch AI estimates that 290,000 to 1.6 million H100-equivalents were smuggled into China in 2024 and 2025, with a median of about 660,000. By Epoch AI’s estimate, that median would be roughly a third of China’s total compute.
A May 2026 Substrate analysis estimates that Chinese companies were renting or planning to rent at least 670,000 H100-equivalents through overseas cloud providers (although the estimate is highly uncertain).
A September 2026 Epoch AI forecast projected Huawei’s 2026 AI-chip output at about 880,000 H100-equivalents (with estimates ranging from roughly 600,000 to 1.2 million). This production still relies on manufacturing equipment from the US, the Netherlands, and Japan.
Little is known about how much compute individual Chinese AI developers own or can access, and how much they can use for a single training run.
Although the first three figures measure different things over different periods, each could be a large source of compute for Chinese AI developers. From outside government, it’s hard to know how much of this the US intelligence community (IC) already tracks. Former executive branch officials and congressional staff I interviewed said policymakers want estimates from outside government. In May 2026, Senators Jim Banks and Tom Cotton asked intelligence leaders to prioritize collection on China’s computing power, data-center build-out, and chip smuggling.
Why does getting these estimates right matter? The most obvious reason is that the US is deciding now which chips and semiconductor equipment China should be able to access. If China can already make enough of something itself, or buy it from other suppliers, controlling it may do little good; but if it still can’t get or make something important for advanced AI, controlling it matters much more.
A second reason is avoiding strategic surprise: Being surprised leaves less time to prepare for the military, scientific, or technological capabilities that China’s AI progress could bring. A third reason is international stability. Were the US to have a clearer picture of China’s capabilities, it would be less likely to act on a mistaken belief about how far ahead or behind China is. Overestimating China could increase the pressure to race, or cause the US to plan around capabilities China does not have. Underestimating it could lead to looser controls than needed and a rushed response once it becomes clear the gap is smaller than it seemed.
What the IC should prioritize
The IC is well positioned to collect and analyze this information, but what should it track? I prioritize seven areas.1 They are not ranked. I chose them by asking whether the information could change the US government’s estimate of China’s compute or inform a policy decision, and whether the IC could get a better answer. For each, the IC should look at current stocks and flows and at where they are headed. I don’t treat licensed purchases such as NVIDIA H200s as a separate priority because the government already has license data; where those chips actually end up and who can use them fall under area (1).
(1) Track how much compute China’s leading AI developers and cloud providers have and who can use it. A national chip count does not show how much compute each Chinese developer can use or how much Chinese firms rent abroad. It would be more useful to know how much compute particular Chinese actors use, where they get it, and what they use it for. That includes how much compute each major Chinese AI developer has or can access through Chinese cloud providers, and how much it can use together for a single training run, including limits imposed by networking and interconnect.
Another question is whether Beijing is steering scarce chips towards particular companies, national labs, or its own government programs. In late 2025, China barred foreign chips from new state-funded data centers, according to Reuters. The IC should also track how much new data-center capacity China is building for AI, when it will be ready, and whether those facilities will have enough power to operate at full capacity.
A clearer picture of who has compute would help the Bureau of Industry and Security (BIS) decide which Chinese firms to target for export-control enforcement. A developer that relies more on NVIDIA chips would be hit harder by restrictions on them than one that has already migrated to Huawei’s hardware and software. Even developers that already train frontier models can use more compute to improve those models through further training and run them for more users.
The same information would give the White House and the Department of War more warning of the military, scientific, or technological capabilities Chinese developers could gain. For Treasury, the question is whether covered US investments involve Chinese developers whose AI systems exceed the training-compute thresholds in its outbound investment rules. More precise information about the compute used to train particular models could help Treasury apply those rules and assess whether the thresholds remain appropriate.
(2) Track how many finished, usable advanced AI chips China can produce itself. Production numbers alone may not be enough. It would be useful to know which AI chips China is producing and the yields at fabrication and advanced packaging. Advanced-packaging defects can mean the number of dies produced exceeds the number of usable AI chips.
It also matters how much of China’s most advanced fabrication capacity is going to AI chips. How many AI chips could China make if it shifted more of its leading-edge capacity away from other chips? Ultimately, the question is how many usable chips China produces and how well they perform.
Fabrication may not be the main constraint on how many usable AI chips China can produce. SemiAnalysis argues that high-bandwidth memory is the bigger bottleneck. (Area (5) below looks at what is holding production back.)
Better production estimates would help BIS decide whether to maintain or change restrictions on exports of finished AI chips to China. While China’s domestic production stays limited, US chip exports add a lot to what Chinese developers have. Once China can make large numbers of comparable chips itself, US export restrictions slow it less, though imports still add to what it can make. How much export controls can delay China’s next models depends on this balance.
(3) Track how much overseas compute Chinese AI companies can actually use. There isn’t a very good estimate of how much compute Chinese companies are accessing overseas. The May 2026 estimate of roughly 670,000 H100-equivalents is highly uncertain. Only about 145,000 come from reported arrangements, some involving planned capacity. Most of the remainder is inferred from reporting about ByteDance’s rental budget that ByteDance disputed. More reliable figures are needed.
The IC should identify which companies are using this compute, what they use it for, which chips they can access, and for how long. Accelerator-hours may be a better unit here: Renting 1 million chips for a month is very different from renting them for a year. It would also help to know how much of this compute can be used together for large AI training runs.
In July 2026, White House science adviser Michael Kratsios accused Moonshot AI of accessing NVIDIA GB300 chips in Thailand. The Remote Access Security Act, which would give Commerce authority over remote access to export-controlled chips, passed the House in January 2026 and is pending in the Senate.
Overseas compute matters for a decision facing the Commerce Department: whether to restrict Chinese companies’ access to overseas cloud computing services, rather than relying on controls on physical chip exports alone. If Chinese companies can rent enough compute overseas for large training runs, controls on physical chips are doing less than they appear to. Even compute too fragmented for a large training run can help Chinese companies improve their models and run them for more users. How much they rent and what they use it for would tell Commerce how much a cloud rule could accomplish.
(4) Track how many restricted AI chips and logic dies, and how much high-bandwidth memory, are being smuggled into China, and where they are going. Epoch AI’s estimate above covers chips and servers physically moved into China, not separately smuggled logic dies and high-bandwidth memory. Much less is public about who is getting smuggled hardware, or how. Proxy companies and diversion through third countries make this picture harder to piece together. The IC should also watch for changes in the routes, intermediaries, or methods used by smugglers.
Knowing the smuggling routes would help BIS decide where to focus export-control enforcement, and help the State Department decide which third countries to work with to stop diversion. For example, if there is intelligence suggesting that a few routes or intermediaries account for most of the smuggling, BIS could focus its investigations and enforcement there.
(5) Track the inputs and bottlenecks in China’s production of advanced AI chips. For each major Chinese AI chip, the question is which inputs or production steps limit how many usable chips China can make. The bottleneck could be high-bandwidth memory, advanced packaging, fabrication, or semiconductor manufacturing equipment. Part of the question is how much China still has in stockpiles of foreign-made logic dies, high-bandwidth memory, and equipment, and how long those supplies could last. The IC should also identify where China has already overcome a bottleneck by successfully indigenizing the relevant technology.
If BIS has a deep understanding of what the bottlenecks are, it can better decide which export controls to maintain, tighten, or reconsider. For example, restricting foreign advanced-packaging equipment or materials like photoresists could slow China’s chip production if it still depends on those supplies. Controls may do little in the near term, however, if China can already make enough of an input or has several years’ supply in stock.
(6) Track China’s progress on indigenizing chipmaking equipment and materials, especially lithography. Lithography shows why this is hard to judge from outside. China still lags, but it is making progress. It has reportedly begun producing domestic deep ultraviolet (DUV) immersion lithography machines, although their performance at commercial scale is unproven. It also reportedly completed a prototype extreme ultraviolet (EUV) machine that, as of Reuters’ December 2025 report, had not yet produced working chips. These machines still have to work reliably enough for commercial chip production.
The US government would also want to know which foreign parts are used in Chinese-made equipment. China’s EUV prototype, for example, reportedly used parts from older ASML machines and components obtained through secondary markets. According to the Financial Times, Chinese manufacturer Yuliangsheng uses Zeiss optics in its DUV machines, though how it obtains them is unclear. Which of those parts can China make itself, and do the replacements actually work?
Answers to these questions would help the US and allied governments (especially the Netherlands and Japan) decide which restrictions on semiconductor manufacturing equipment, components, and replacement parts to maintain or change. As Chinese replacements become available at scale, governments should reconsider which restrictions are still worth keeping and where to focus enforcement.
(7) Track illicit foreign fabrication of Chinese AI chip or high-bandwidth memory designs. One example is the TSMC/Sophgo case. TSMC made a chip for Sophgo that was later found in Huawei’s Ascend 910B. Sophgo denies any business relationship with Huawei, and Huawei denies using TSMC-made Sophgo chips. But subsequent reporting indicates Huawei obtained more than 2.9 million TSMC-made dies through Sophgo and other intermediaries before the channel was cut off in late 2024.
SemiAnalysis believes Huawei has since had other chips, including its data-center CPU, made at TSMC through shell companies. The IC needs to determine whether such proxy channels are still operating, how much they supply, and which Chinese firms receive the chips or memory.
This information would help BIS judge whether stricter controls on foreign foundry orders would substantially reduce China’s supply of advanced AI chips. If illicit foreign fabrication remains a major source of AI chips for China, BIS has reason to tighten customer checks and investigate the firms involved. Understanding how these orders get past the foundries’ checks would also help BIS decide whether the existing requirements need to change or whether companies are failing to follow them, and can help BIS design better due diligence requirements for other rules in the future.
Where the IC adds value
Why should the IC specifically do this? There are two reasons. First, the IC has information nobody else has, including classified information. Think tanks and analysts like Epoch AI, SemiAnalysis, ChinaTalk, and the Center for Security and Emerging Technology already do some of this work through open-source reporting, but classified collection can add value when information is hard to find or deliberately hidden. For example, open-source reporting is unlikely to reveal sufficient detail on fabrication yields, covert stockpiles, smuggling networks, and internal Chinese government planning.
Second, the IC has resources and analytical expertise. It can combine classified information, public and commercial data, and what BIS learns from licenses and end-use checks. It can also keep its findings classified so that China doesn’t learn, for example, which smuggling routes and intermediaries the US has identified. Policymakers could also get this analysis on demand, rather than hoping an outside group is already collecting and analyzing the information.
The White House and the National Security Council should ensure these seven areas receive sustained IC attention, whether or not standing intelligence requirements already cover them. The Office of the Director of National Intelligence (ODNI) should assess whether existing collection and analysis adequately answer these questions and identify where more work is needed. IAPS has proposed a dedicated ODNI capability to track adversary compute. The Center for a New American Security has also recommended a joint Commerce–IC analysis team to support chip-smuggling enforcement.
ODNI should also ensure estimates are regularly updated and available to the Departments of Commerce, Treasury, State, and War as well as the White House. Important findings should be shared with Congress and any relevant allies implementing the controls, in unclassified form where possible. Policymakers would then have a better basis for deciding which export controls still work, where to focus enforcement resources, and how much time the US has to prepare for advances in China’s AI capabilities.
I base these recommendations on interviews with 14 people, including former intelligence and BIS officials, congressional staff, and researchers, and on public reporting.
