DeepStack
Research

The Buildout · Part 18Chips and suppliers

The Fence Stops at the Chip

Washington opened H200 licensing for China in January; Beijing restricted the purchases, and licensed shipments came to less than 1% of Data Center revenue in the quarter to July 26. On the token, Chinese developers drew 47.9% of OpenRouter's text requests in the week of October 5.

By DeepStack · Oct 25, 2026 · 10 min read Thesis book: Scarce parts

Correction, Nov 3, 2026: An earlier version said Kevin Gee described China as building a domestic stack at every layer. He wrote that China is the only country working on a full stack, with compute its bottleneck.

Correction, Nov 3, 2026: An earlier version said Oguz Erkan expected Western firms to avoid Chinese models however close their cost. He wrote that few would prefer them as long as the cost stays close.

Correction, Nov 3, 2026: An earlier version did not say that the account of customers seeking American-made modules came from Applied Optoelectronics, as relayed by Gaetano.

Correction, Nov 3, 2026: An earlier version said Prof G Markets called the Chinese open-weight strategy one of the most effective business models and expected a fight once Washington grasped it. Prof G Markets ranked it second only to IP theft and named President Trump.

Illustration for Part 18

Where this Part stands

The question

Are the US and China building two separate AI markets, or only two markets for chips?

The Call

SplitTwo markets for chips, one for tokens

Both capitals fence the chip, so it trades in two markets. The token is fenced only for national-security buyers and trades as one.

Conviction moderate · 3 of 5

The next test

Nov 9, 2026

The Commerce Department's pause on its affiliates rule ends on paper: whether a Federal Register document extends it by that day. The truce announced on September 23 is not such a document.

The test this Part set is the card beside this one

What would flip the Call

  • Nvidia's 10-Q for the quarter to Oct 25 puts licensed H200 shipments to China at 1% or more of Data Center revenue.
  • A law or rule bars Chinese-developed AI models from agencies beyond the Pentagon and the intelligence community by Mar 31, 2027.
  • Chinese developers' share of OpenRouter text requests falls below 25% in the week of Jan 4, 2027, with no rule behind it.

What to do with it

  1. 1Read the H200 line in Nvidia's 10-Q after the Nov 18 results: under 1% of Data Center revenue means Beijing's gate is still shut.
  2. 2Check the Federal Register on Nov 9 and China's commerce ministry on Nov 10, not the communique: without a document, each pause lapses.
  3. 3Price a Chinese model on one cloud before trusting a tenfold gap: DeepSeek V4 Pro cost $0.20 per million tokens direct, $1.90 on Azure.

Washington opened its gate on January 15, 2026. That day the Bureau of Industry and Security replaced its presumption of denial with case-by-case review for exports of Nvidia's H200, AMD's MI325X and lesser chips to China and Macau, on three conditions: American customers are not shorted, China-bound volume stays at or below half of what American customers get, and a laboratory in the United States tests every shipment.

Beijing kept its own gate shut. Nvidia's 10-Q for the quarter to July 26, 2026 says the American government granted licenses from February to ship small amounts of H200 to specific China-based customers, and the Chinese government restricted those sales. Nvidia took a $0.4 billion charge on H200 inventory and purchase obligations in the first half of fiscal 2027, and the licensed shipments it did make came to less than 1% of Data Center revenue of $89.0 billion. Each paid a 25% tariff on its way through American inspection, and Beijing, the filing adds, has encouraged buyers to purchase from Nvidia's Chinese competitors. Nvidia's CFO commentary of August 26 put all Hopper shipments to China below 1% of that revenue.

The token went the other way. In the week beginning October 5, 2026, models from DeepSeek, Z.ai, Alibaba's Qwen, Xiaomi and Tencent drew 47.9% of text requests on OpenRouter, a service that routes developers to hundreds of models, by DeepStack's reading of its rankings on October 12. OpenAI, Google and Anthropic drew 40.3%, and DeepSeek alone, at 24.5%, led every American developer. Requests are not revenue, and OpenRouter's users chase price. The same Nvidia filing says Washington is considering limits on supporting applications built on Chinese open-source models such as DeepSeek, Qwen and Kimi. What is written reaches only the national-security buyers: the defense authorization act for fiscal 2026, signed December 18, 2025, ordered models from DeepSeek and its backer High Flyer off the Pentagon's systems and out of its contractors' defense work, and had the intelligence agencies remove DeepSeek's application from their national security systems and their contractors'.

Part twelve found on October 9 that the token trades as one market. Part eighteen goes down the stack and asks where the fence is written, and by whom. DeepStack read 243 posts, letters, filings and episodes by 83 voices and found 22 disagreements on the question, 5 of them direct contradictions. That is a thin yield for a question this broad, so most of the argument here rests on the rules both capitals have published.

The fence count

DeepStack counted the written measures since January 2025 at each layer of the AI stack as of October 12, 2026, and whether each carries an end date. On silicon, Washington's side is the January 15 rule, with no end date. Beijing's side has no public document; the restriction shows only in Nvidia's 10-Q. The rule that extends American controls to companies at least 50% owned by listed entities was suspended from November 10, 2025, until November 9, 2026. Treasury Secretary Scott Bessent announced on September 23 that the truce with China would run through January 10, 2027, and a Miller Canfield note of October 1 says the suspension is expected to follow. The bureau's Federal Register documents through October 12 carry no extension.

On minerals, China's Announcement No. 70 of November 7, 2025, suspended its October 2025 controls on rare earths, battery materials and superhard materials until November 10, 2026. Press reports after the September summit put that suspension on the same January 10 date; DeepStack found no ministry document carrying it.

On networks and machines, the FCC's Covered List, updated October 9, 2026, bars new authorizations for foreign-made drones and their critical components, added December 22, 2025, for routers made in any foreign country, added March 23, and for foreign-made power inverters and advanced robots, added July 28. Each entry exempts products the Department of War has granted a conditional approval, and the drone, router and inverter entries accept one from Homeland Security as well. Drones on the Defense Contract Management Agency's Blue UAS list, or built as domestic end products under the Buy American standard, stay exempt until January 1, 2028, and toy drones are exempt outright. A rule effective October 13 reaches devices built with logic-bearing parts from listed producers, and it spares equipment already authorized and applications pending that day. The list has named Huawei, ZTE, Hikvision and China's three state carriers since 2021 and 2022; those entries predate this count. Optics have a reported draft and no entry on the October 9 list. On the model, the defense act carries two bars on DeepSeek, one for the Pentagon and its contractors and one for the intelligence agencies and theirs, and lets the Pentagon extend its bar by guidance to models from any AI company based in China.

The count is seven measures without an end date, all American: the January rule, four Covered List entries, the component rule and the defense act. Six sit below the model. The seventh keeps one developer's models off the systems of the Pentagon, the intelligence agencies and their contractors, and reaches no company, cloud or developer outside those buyers. Washington once reached further. On June 12 the Commerce Department imposed a worldwide license requirement on transfers of two Anthropic models to foreign persons, Representative Sam Liccardo wrote on June 18, and a Cloud Security Alliance note records that the requirement was lifted on June 30. Two suspensions, one in each capital, lapse a day apart in November on paper; the truce they belong to runs to January 10 by announcement, and neither capital has published a document extending its suspension.

Where there is hardware to inspect at a border, the fence is written down.

Where there is hardware to inspect at a border, the fence is written down. Where the product is a download, the only fence in federal law surrounds the national-security buyers.

Two gates for silicon

Kevin Gee wrote in A Letter a Day on September 12 that Washington is betting compute is the linchpin of the stack, which makes it the target of export controls, while China is the only country working on a full stack of its own, with compute its bottleneck. On volume the wager holds. Venkat Somala of Epoch AI estimated on September 4, updated September 24, that Huawei's 2026 chips will be worth 880,000 H100 equivalents against Nvidia's 23 million, under 4% of Nvidia's compute. On Chinese-made memory alone, Epoch puts Huawei near 1.5 million H100 equivalents a year by 2028, against nearly 100 million for Nvidia.

The other side reads the gap as time bought and little more. Doomberg wrote on July 3 that China can invent its way around chip sanctions, citing Huawei's announced framework for 1.4-nanometer-class density by 2031 without extreme ultraviolet lithography. DeepValue Capital wrote on August 24 that Alibaba is ramping its own chip production for a cost advantage.

The written fence also leaks. Isabel Juniewicz of Epoch AI estimated on April 29 that about 660,000 H100 equivalents had been smuggled into China through the end of 2025, within a range of 290,000 to 1.6 million, and put her central figure at roughly a third of China's total compute. The licenses and the restrictions govern the legal market. A gray channel runs beside it, and at the top of Epoch's range it outweighs the legal one.

Gee drew the stakes from Jensen Huang's exchange with Dwarkesh Patel: a DeepSeek model that arrives tuned for Huawei's hardware rather than Nvidia's would cost America ground in the contest for the stack.

DeepStack's reading is that the second chip market is being built from both ends. Washington caps what may cross, Beijing restricts most of what Washington lets through, and Huawei and Alibaba fill the space at a fraction of Nvidia's output. Because both capitals hold the fence, either one can lower its half.

The rack in between

Between chip and model sits the hardware that wires a data center, and there the fence is half drafted. Gaetano warned on August 5 that Chinese suppliers are so large a share of optics that a sudden ban could deepen shortages and push out deployments. Ren answered on August 14 that the FCC's draft ban creates no capacity and hands scarcity pricing to whoever already has supply. Buyers moved first: Gaetano reported on August 11 that Applied Optoelectronics had found customers growing more insistent on American-made modules.

Gaetano also read the proposal as leverage, calling it on August 5 a negotiating tool ahead of a planned meeting with Xi Jinping, since China controls a substantial share of indium phosphide substrate production and of indium exports. Ren located the grip lower down on August 14: firms outside China hold most of the wafers and the equipment, and Beijing's hold is on the indium and phosphorus beneath them. By October 2 Gaetano was relaying a Morgan Stanley note: a Chinese-built transceiver might still enter if roughly 65% of its value came from American companies, with the 3.2-terabit generation as the point the rule would bite. DeepStack reads a content threshold as a fence with a turnstile, one that sorts a module by where its value is made rather than where it is assembled.

Robots show the finished version. Timothy B. Lee of Understanding AI wrote on August 31 that Washington met Unitree's rise with a House bill naming it and FCC rules on foreign-made robots, and warned that a ban could leave America a robotics backwater while Chinese makers dominate elsewhere. Alice Han and James Kynge, on Prof G Media on September 15, put China's 2025 private AI investment at about one twenty-third of America's, while it shipped ten times as many humanoid robots.

The fence runs both ways. Baillie Gifford's third-quarter 2026 letter calls rare earths leverage Beijing holds over other economies, and Doomberg reported on September 1 that the Pentagon's Office of Strategic Capital is structuring deals to move military supply chains away from China.

One market for the token

Oguz Erkan of Capitalist Letters wrote on July 19 that he did not think usage would drain from American labs to Chinese models: as long as the cost per unit of intelligence stays this close, not many Western companies will prefer Chinese models, given the strategic and regulatory risks. The next day Prof G Markets ranked the Chinese open-weight strategy, giving 80% of a product's value at half the price, as the second-best business model in history, after IP theft, and expected it to become "the next geopolitical football once Trump figures out what's happening." DeepStack reads the two as a direct contradiction about whether Western buyers will take the discount. Three months on, OpenRouter's figures for the week of October 5 lean toward Prof G Markets.

American clouds sell across the line. Market Sentiment noted on September 8 that Azure sells the same DeepSeek V4 Pro weights at $1.90 per million tokens that DeepSeek's own service prices at $0.20, and that on one platform Anthropic's Opus 5 costs about twice DeepSeek, not the twentyfold of headline comparisons. Alice Han and James Kynge, in an August 18 episode, described Microsoft closing offices in China while still selling Azure and AI services to ByteDance and Shein.

Michael Spencer of AI Supremacy wrote on July 21 that China holds the cheaper tokens and the more abundant energy. James Bulltard read the same progress on July 17 as a reason Washington will spend and restrict more, not less.

Erkan may be right about the large Western enterprise, which OpenRouter's developer is not, but his fence is a buyer's choice that no document makes outside the Pentagon and the intelligence agencies. The longer risk is the one Gee names: open weights are China's wedge and carry no lock-in, while the silicon, runtime and on-site engineers beneath them are bought from whoever builds the stack, and Huawei sends engineers.

Gee holds that China is the world's supplier and not yet its platform. A day later Stephen Clapham of Behind the Balance Sheet relayed the opposite view from Sir Michael Moritz, the former Sequoia investor, whose conversation at a Financial Times event he had watched: the technological baton is passing to China, as it once passed from Britain to America, on the strength of electric vehicles, batteries and solar. DeepStack reads the two as a direct contradiction. On AI the record favors Gee for now: China's models travel as open weights, and the licensed chips under them stay home.

DeepStack's reading is that one token market is how two stacks could form later. Today a Chinese model runs on an American cloud and an American chip. If the next DeepSeek arrives tuned for Huawei, the token market stays whole only while the weights run on both.

What settles it

Four public documents test the reading by March 31, 2027, each asked before its date.

On November 9, the affiliates-rule pause ends on paper unless a Federal Register document extends it. The truce announced on September 23 runs to January 10, and Miller Canfield expects the suspension to follow, but Washington has not confirmed that it does. Last year's suspension reached public inspection only on November 10, 2025, the day it began, and a document that lands a day late again is a no. China's rare-earth suspension lapses on paper the next day, and the commerce ministry's site is where its extension would appear. The Asia-Pacific leaders' meeting in Shenzhen later in November, the next chance to write the truce down, comes after both lapses.

On November 18, Nvidia reports its quarter to October 25, 2026. The line that matters sits in the 10-Q: whether licensed H200 shipments to China-based customers reach 1% of Data Center revenue, after less than 1% in the quarter to July 26. Below that line, Beijing's gate is still shut.

By December 31, the FCC's Covered List either carries an entry for optical transceivers or it does not. The October 13 component rule does not count. By March 31, 2027, a law or a Federal Register rule either bars Chinese-developed models from federal agencies beyond the Pentagon and the intelligence community or leaves the token market as it stands.

What would prove DeepStack wrong is Nvidia shipping licensed H200 at 1% or more of Data Center revenue, a written bar on Chinese models that reaches past the national-security buyers, or Chinese developers' share of OpenRouter requests falling below 25% with no rule behind it, which would prove Erkan right about the buyer's choice. The chip half describes the licensed market, and smuggled chips run beside it uncounted.

The numbers

Who answers the requests

Share of text requests on OpenRouter by model author, week beginning October 5, 2026, percent; DeepSeek, Z.ai, Qwen, Xiaomi and Tencent are Chinese developers

  1. DeepSeek24.5%
  2. OpenAI18.6%
  3. Google17.8%
  4. Z.ai13.4%
  5. Qwen (Alibaba)5.4%
  6. Anthropic3.9%
  7. Xiaomi2.9%
  8. Mistral AI2.0%
  9. Tencent1.7%
  10. All other authors9.7%

Source: OpenRouter rankings, market share by model author, read Oct 12, 2026; requests, not tokens or dollars

The DeepStack Call

Two markets for chips, one for tokens

The DeepStack Call is that the US and China are building two markets for chips and one for tokens. Both capitals hold the chip fence: Washington licenses the H200 under a rule with no end date, Beijing restricts the purchases without a public document, and Nvidia shipped less than 1% of its $89.0 billion of Data Center revenue under the licenses in the quarter to July 26, 2026. At the model, the written bars keep DeepSeek off the systems of the Pentagon, the intelligence agencies and their contractors, and Chinese developers drew 47.9% of OpenRouter's text requests in the week of October 5. The single most important reason for caution is that the token half rests on usage among price-sensitive developers, which can change without a rule, and that is why conviction stops at three. The chip half describes the licensed market; Epoch AI puts smuggled chips at roughly a third of China's compute. The view changes if Nvidia's November 10-Q puts licensed H200 shipments to China at 1% or more of Data Center revenue, or if a law or rule bars Chinese models from agencies beyond the Pentagon and the intelligence community by March 31, 2027. The first answer arrives on November 9, in the Federal Register.

The strongest published case against

James Bulltard · Jul 17, 2026: China's advances in AI, exemplified by Deepseek, will likely force the US to increase investment and restrictions in AI rather than cut capex, because allowing China to lead would be strategically unacceptable.

Answers Michael Spencer · qualifies

What to watch

The dated tests this Part set

  1. Nov 9, 2026The Commerce Department's pause on its affiliates rule ends on paper: whether a Federal Register document extends it by that day. The truce announced on September 23 is not such a document.pending
  2. Nov 17, 2026Nvidia reports its quarter to October 25, 2026: whether its 10-Q puts licensed H200 shipments to China at 1% or more of Data Center revenue, after less than 1% in the quarter to July 26.pendingQuestion →
  3. Dec 31, 2026Whether the FCC's Covered List carries an entry for optical transceivers or modules, after its October 9, 2026 update left them off.pendingQuestion →
  4. Mar 31, 2027Whether a law or a Federal Register rule bars agencies beyond the Pentagon and the intelligence community from Chinese-developed AI models; the defense act of December 2025 reaches only those buyers and their contractors.pendingQuestion →
23 sources cited
  1. Bureau of Industry and Security - License review policy for advanced computing, Federal Register, Jan 15, 2026
  2. NVIDIA - Form 10-Q for the quarter ended July 26, 2026
  3. NVIDIA - CFO commentary, second quarter of fiscal 2027, Aug 26, 2026
  4. OpenRouter - LLM rankings, market share by model author, read Oct 12, 2026
  5. GovInfo - National Defense Authorization Act for Fiscal Year 2026, Public Law 119-60, sections 1532 and 6604, Dec 18, 2025
  6. Bureau of Industry and Security - One-year suspension of the affiliates rule, Federal Register, Nov 12, 2025
  7. JD Supra - Miller Canfield, BIS affiliates rule suspension extended but uncertainty remains, Oct 1, 2026
  8. China Ministry of Commerce - Announcement No. 70 of 2025, Nov 7, 2025
  9. Federal Communications Commission - Covered List, updated Oct 9, 2026
  10. Federal Communications Commission - Equipment authorization rule, Federal Register, Sep 11, 2026
  11. Rep. Sam Liccardo - Letter to the Commerce Department on frontier model export controls, Jun 18, 2026
  12. Epoch AI - Will Huawei catch up to Nvidia by 2030?, Sep 4, 2026
  13. Epoch AI - Isabel Juniewicz, Diversion and resale: estimating compute smuggling to China, Apr 29, 2026
  14. Understanding AI - Timothy B. Lee on Unitree, Aug 31, 2026
  15. Prof G Media - Alice Han and James Kynge on China, Sep 15, 2026
  16. Baillie Gifford - Q3 2026 letter
  17. Capitalist Letters - Oguz Erkan, Jul 19, 2026
  18. Prof G Markets - Jul 20, 2026
  19. Market Sentiment - Notes on Open Source, Sep 8, 2026
  20. Prof G Media - Alice Han and James Kynge on Microsoft in China, Aug 18, 2026
  21. AI Supremacy - Michael Spencer, Jul 21, 2026
  22. James Bulltard - Jul 17, 2026 recap
  23. DeepStack - Part twelve, The Model Is Not the Price

Text, numbers, Call and tests from the DeepStack engine, data as of Nov 10, 2026. Analysis and opinion, not investment advice.

The thesis this memo supports

Memory spot prices, Korea's chip exports and the revenue of the equipment, storage and network-hardware makers: whether the parts AI runs short of stay scarce.

Explore the thesis and its evidence 25 open events in this thesis →

Thesis discussion

0 contributions · across all memos

Loading discussion…

Earlier memo · Part 17
The Guarantee Has No Date
Back to the question: “Nvidia H200 China sales 1% or more of Data Center?”Open →