What DeepSeek’s $5.6 million estimate says about China’s AI race

The AI Power Dossier · Evidence checked through 11 November 2026

DeepSeek reported that the official training of V3 used 2.788 million Nvidia H800 chip-hours. At an assumed rental price of $2 an hour, that comes to $5.576 million. Its table includes pre-training, an extension of the model’s working context and post-training. DeepSeek expressly excluded earlier research and experiments. The figure is Verified as the company’s estimate for that run. How much it spent developing the model altogether remains Unresolved. DeepSeek V3 technical report, Table 1

That accounting choice matters in a debate about how much Washington’s chip restrictions can slow Chinese AI. A laboratory can improve its software or spend less on a training run while still relying on foreign-made chips. This investigation looks separately at what the models can do, what Chinese chip suppliers can deliver, and what export rules and court records establish about access.

In an April 2026 evaluation published on 1 May, the US Center for AI Standards and Innovation estimated that DeepSeek V4 Pro lagged leading publicly released US models by about eight months on its aggregate benchmark measure. The comparison used a suite covering five domains. Results varied by task, and the estimate leaves open how the models would perform across other uses, at different prices or under different deployment constraints. NIST’s May evaluation

Stanford’s AI Index gives a different comparison. In its March 2026 Arena snapshot, the leading US model scored 1,503 and the leading Chinese model scored 1,464. Arena uses human preference ratings. The CAISI evaluation used other models, dates and tasks. Neither result establishes an overall lead across uses. Stanford AI Index, technical performance

Epoch AI asked a different question. Its median 2026 projection puts Huawei’s annual accelerator output below 4% of Nvidia’s when converted to a common measure of theoretical performance. The projection depends on estimates of units produced and the performance assigned to each design. It says little by itself about how much hardware either country already has, where it runs, or how well it handles a particular workload. The two measures can move at different speeds. A laboratory’s model can improve on a test while domestic chip production remains constrained. Epoch AI’s Huawei output analysis

There is a case for expecting the hardware constraint to ease. Huawei says it expects its Ascend 960DT in the first quarter of 2027, ahead of the timetable it previously gave. More efficient software could also allow researchers to get better results from the equipment they already have. The company’s timetable is Verified as an announcement. Buyers would need to see shipments, available capacity and independently measured performance to know how much it changes their options. Huawei Connect announcement, 17 September

A Shenzhen research team reports testing post-training infrastructure with DeepSeek V4 Pro on Huawei Ascend 910C chips and separately running task-specific post-training and evaluation with V4 Flash. This is evidence about the work the team describes. It leaves the hardware used for DeepSeek’s original V4 pre-training and general performance at scale unresolved. Shenzhen team’s Ascend report, 22 July 2026

A US rule effective 15 January 2026 moved qualifying shipments of certain chips to China or Macau, including the H200, to conditional case-by-case licensing. In its filing for the quarter ended 26 July, Nvidia said it made a fraction of the shipments allowed under the US H200 licensing programme. Nvidia said those shipments accounted for less than 1% of its Data Center revenue in that quarter. In the same filing it reported $7.880 billion in revenue attributed to direct customers headquartered in China, including Hong Kong, across its products. The first number describes a slice of data-centre sales. The second uses a customer-headquarters measure across products, so it cannot tell a reader where particular accelerators ended up. US licensing rule, 15 January · Nvidia quarterly filing, revenue geography and licensing discussion

Prosecutions offer a different kind of evidence. In March, federal prosecutors in New York charged three people in a case concerning servers with advanced chips. The government alleged an intermediary bought about $2.5 billion worth of servers and that about $510 million worth was diverted to China over three weeks. The purchase figure covers a wider set of transactions. An October court order records a September guilty-plea proceeding for one defendant and directs that related materials be unsealed. The order alone leaves the accepted counts and any sentence Unresolved. The charges against the other defendants remain allegations in the records examined here. Super Micro said it was cooperating and that the company was not charged. Neither the charges nor the plea order identifies DeepSeek as an end user. US Attorney’s charging announcement · October unsealing order

This is the enforcement question that aggregate chip estimates leave out. A company can make a lawful purchase, while a later transfer raises a separate question about licences, the parties and end use. The evidence needed to connect those transactions to a particular model would include shipping and deployment records identifying the hardware and the customer. The records reviewed here do not make that connection.

In September the NSA, FBI and CISA issued a joint assessment alleging that China-based AI companies systematically extracted capabilities from US frontier models. China’s commerce ministry called the allegations groundless and said Washington was politicising a normal industry technique. Contested. The agencies’ public release and China’s response establish the two positions. The full advisory was unavailable to this review, and the underlying account and training evidence needed to test the allegations about individual companies remains Unresolved. NSA announcement · Chinese commerce ministry response

An October US notice describes a Homeland Security order excluding covered DeepSeek products and services from civilian federal procurement and requiring the removal of covered products and implementations from civilian federal IT. That is a specific federal action with a stated scope. It does not answer how frequently businesses or individuals elsewhere use the models. Federal Register notice · Energy Department implementation notice

Separate rules effective 15 July 2026 cover public AI services in China offering sustained emotional interaction. They require users to be told they are interacting with AI and prohibit virtual intimate-relationship services for minors. Ordinary work assistants and question-answering services without sustained emotional interaction fall outside these companion-service rules. Enforcement remains unverified in the records reviewed. CAC companion-service rules, Arts. 2, 14, 18 and 32

For commercial users, the availability of downloadable model weights gives another option. A developer can run those weights on its own equipment or choose a hosting provider. That creates a route for Chinese models to gain users beyond China. A download leaves open where the model runs and whose chips it uses. Companies can also benefit from demand for hosted services. Alibaba’s June-quarter filing reported 48.44 billion yuan of sales in AI Cloud and Compute Services, a newly combined segment that includes its chip business, T-Head. The sales figure cannot tell us how much came from any one model. DeepSeek V3 technical report and model access · Alibaba June-quarter results

The investigation compares the technical paper with government testing, supplier estimates, issuer disclosures and legal records. It attributes projected output to the modeller and roadmap dates to the issuer. The full US advisory, several procurement and ownership records and the latest dockets were unavailable. The scope of each source is stated where it changes what the evidence can establish.

The records establish how DeepSeek calculated one training estimate, how the US assessed a later model on selected tasks, and what particular rules and court orders cover. They give less certainty about installed chip capacity, the equipment used for future models and how procurement rules affect actual deployments. Verified shipments of Huawei’s next generation, comparable independent tests of released models, and court and trade records linking equipment to end users would change that assessment. Until then, a cheaper run or a strong result on a benchmark tells us about one part of the race.