US government–Nvidia

US licences let Nvidia sell H200s to specific China-based buyers from February 2026 for a 15–25% cut; Beijing held them back, and shipments in the quarter to July were under 1% of data-center revenue.

A deal between US government, Nvidia

Why it matters

Washington charges a 25% tariff when licensed H200 chips enter the US for mandatory inspection. Nvidia absorbs it and says officials named a 15% H20 revenue share that never made it into regulation. Beijing held buyers back, and Nvidia's licensed H200 shipments accounted for under 1% of a quarter's data-center revenue.

Tap a figure for the line in its source.

US government gives Nvidia

  • export licences, from February 2026, to ship small amounts of H200s to specific China-based customers, under a January 2026 rule that reviews each application case by case
  • the rule's volume condition: what goes to China may total no more than half of what the chip has shipped to US customers for use in the US, counted in computing power

Nvidia gives US government

  • tariff on every H200 shipped under the licences, charged when the chip enters the US for the third-party inspection the licences require; Nvidia says it has passed none of it on
  • disputedthe government's cut as stated: the '15% or more' of licensed revenue that officials expected, per Nvidia (whose filings first said 15%, on the H20, in August 2025, and say no regulation was published), or the President's '25%', collected as the tariff

Nov 21, 2025 internal talks reported by Bloomberg; Trump announced it on Dec 8updated Nov 19, 2026licensed but largely blocked by the importer disclosed by the parties

What happened

In December 2025, President Trump said Nvidia could sell its H200 AI chip to approved customers in China, with “25%” paid to the United States. In January, the Commerce Department began allowing those sales under case-by-case licences. China's total is capped at half the computing power of the chips shipped for US use. An independent lab in the US must check every shipment. A tariff signed the same week charges 25% when the chips enter the US for that check. Nvidia says it has been unable to pass on the cost. The government's cut has been described in two ways. Nvidia says officials expected “15% or more” of licensed revenue. Nvidia's filing gave 15% when the H20 was licensed in August 2025. No regulation codified that revenue share. The 25% tariff applies to the chips when imported into the US. The charge is calculated on the import rather than on the sale to China. Nvidia had licences by February, but Beijing held its companies back. Press accounts also described holdups on the US side. As of May 20, Nvidia had no revenue from the licences. As demand for the H200 diminished, Nvidia took a $0.4B charge for excess H200 inventory and purchase obligations. It then shipped a fraction of what the licences allow. Those shipments accounted for under 1% of Nvidia's data center revenue in the quarter to late July.

Sources 47 documents — 15 press reports · 11 government statements · 9 SEC filings · 5 research notes · 4 transcripts · 1 company release · 1 social post · 1 other document

Research trail Found 242 documents and read 226 of them while checking this deal and related ones.

  1. SEC filingsec.gov
    States 5 figures: 25% tariff · tariff incidence · licences · 1% Q2 shipments ceiling · $0.4B h200 charge
  2. GovInfo · government statementgovinfo.gov
    States 3 figures: tpp threshold · 50% volume condition · US testing
  3. White House · government statementJan 2026whitehouse.gov
    States 25% tariff instrument
  4. SEC filingsec.gov
    States no revenue to may
  5. SEC filingsec.gov
    States outlook
  6. CNA (Singapore) · press reportchannelnewsasia.com
    States approved buyers
  7. Yahoo Finance · press reportca.finance.yahoo.com
    States deliveries
  8. Air Freight News · press reportairfreight.news
    States importer block
  9. SEC filingsec.gov
    Cited by the record; no single figure is quoted from it.
  10. company releases201.q4cdn.com
    Cited by the record; no single figure is quoted from it.
  11. SEC filingsec.gov
    Cited by the record; no single figure is quoted from it.
  12. SEC filingsec.gov
    Cited by the record; no single figure is quoted from it.
  13. SEC filingsec.gov
    Cited by the record; no single figure is quoted from it.
  14. SEC filingsec.gov
    Cited by the record; no single figure is quoted from it.
  15. government statementmedia.bis.gov
    Cited by the record; no single figure is quoted from it.
  16. government statementJan 2026whitehouse.gov
    Cited by the record; no single figure is quoted from it.
  17. government statementcontent.govdelivery.com
    Cited by the record; no single figure is quoted from it.
  18. social postpresidency.ucsb.edu
    Cited by the record; no single figure is quoted from it.
  19. government statementgovinfo.gov
    Cited by the record; no single figure is quoted from it.
  20. transcripttranscripts.cnn.com
    Cited by the record; no single figure is quoted from it.
  21. research noteeverycrsreport.com
    Cited by the record; no single figure is quoted from it.
  22. SEC filingJun 27, 2026ir.amd.com
    Cited by the record; no single figure is quoted from it.
  23. transcriptrev.com
    Cited by the record; no single figure is quoted from it.
  24. transcripthearinghighlights.com
    Cited by the record; no single figure is quoted from it.
  25. government statementmofcom.gov.cn
    Cited by the record; no single figure is quoted from it.
  26. government statementmofcom.gov.cn
    Cited by the record; no single figure is quoted from it.
  27. government statementmofcom.gov.cn
    Cited by the record; no single figure is quoted from it.
  28. government statementgzly.mofcom.gov.cn
    Cited by the record; no single figure is quoted from it.
  29. government statementipc.fmprc.gov.cn
    Cited by the record; no single figure is quoted from it.
  30. press reportnews.bloombergtax.com
    Cited by the record; no single figure is quoted from it.
  31. press reportlatimes.com
    Cited by the record; no single figure is quoted from it.
  32. press reportthestandard.com.hk
    Cited by the record; no single figure is quoted from it.
  33. press reportfinance.yahoo.com
    Cited by the record; no single figure is quoted from it.
  34. press reportJan 30, 2026wtvbam.com
    Cited by the record; no single figure is quoted from it.
  35. press reportJul 14, 2026wkzo.com
    Cited by the record; no single figure is quoted from it.
  36. press reportJul 2026nst.com.my
    Cited by the record; no single figure is quoted from it.
  37. press reportAug 18, 2026935thelloyd.com
    Cited by the record; no single figure is quoted from it.
  38. press reportthenextweb.com
    Cited by the record; no single figure is quoted from it.
  39. press reportnews.bloombergtax.com
    Cited by the record; no single figure is quoted from it.
  40. press reportbloombergtechnoz.com
    Cited by the record; no single figure is quoted from it.
  41. other documentbbw.vn
    Cited by the record; no single figure is quoted from it.
  42. press reporttomshardware.com
    Cited by the record; no single figure is quoted from it.
  43. research noteifp.org
    Cited by the record; no single figure is quoted from it.
  44. research notecnas.org
    Cited by the record; no single figure is quoted from it.
  45. research noteiaps.ai
    Cited by the record; no single figure is quoted from it.
  46. research noteeto.tech
    Cited by the record; no single figure is quoted from it.
  47. transcriptrollcall.com
    Cited by the record; no single figure is quoted from it.

The record's own summary of its sourcing: Nvidia FY2026 10-K (Feb 25 2026) and annual report, Q1 FY2027 10-Q (May 20 2026), Q2 FY2027 10-Q (Aug 26 2026), Q2 and Q3 FY2026 10-Qs, Q4 FY2026–Q2 FY2027 CFO commentaries; BIS final rule (Federal Register, Jan 15 2026) and release (Jan 13 2026); Proclamation of Jan 14 2026, White House fact sheet, CBP CSMS #67400472; Truth Social (Dec 8 2025) and Jan 14 2026 remarks (American Presidency Project, govinfo); CNN transcript (Aug 11 2025); CRS LSB11409 (Mar 23 2026); AMD Q2 2026 10-Q; Lutnick (Senate, Apr 2026) and Kessler (House, Jul 14 2026) hearing transcripts; MOFCOM and FMPRC briefings and a MOFCOM message-board reply (Aug 19 2026); Bloomberg (Nov 21 2025; Mar 3 2026 via LA Times; Aug 26 2026, with Bloomberg Technoz's and Bloomberg Businessweek Vietnam's editions), Reuters (Dec 31 2025 via The Standard; Jan 14, Jan 28 and Jan 30, May 14 via CNA, Jul 14 and Aug 18 2026), The Information (Jul 8 2026, via Reuters/NST), FT (Aug 18 2026, via Yahoo Finance), TNW, Tom's Hardware; IFP, CNAS, IAPS; ETO; Roll Call (read 2026-10) · as of Aug 2026

The evidence 14 sourced figures · 5 disputes

The record's headline US government–Nvidia: H200 licences for China-based customers (Feb 2026) for a cut announced as 25% and collected as a 25% US import tariff (officials' earlier '15% or more' uncodified, per Nvidia) — then largely blocked by the importer: Beijing held buyers back, and shipments in the quarter to Jul 26 2026 were under 1% of Nvidia's Data Center revenue

Size
licences for 'small amounts' of H200s · a 25% tariff on each licensed chip (the cut contested at 15–25%) · shipments under 1% of Q2 FY2027 Data Center revenue ($89.0B)
When
Nov 21 2025 internal talks reported (Bloomberg) → Dec 8 2025 announced → Jan 13–15 2026 BIS rule; Jan 14 tariff, effective Jan 15 → Feb 2026 licences → no revenue from them as of May 20 2026 → first shipments in the quarter to Jul 26 2026 → by Aug 18 2026 small batches let into the mainland (FT)
Status
licensed but largely blocked by the importer — US licences since Feb 2026; Beijing restricted the purchases (Nvidia: 'such sales were restricted by the PRC government'; the press also reports holdups on the US side); Nvidia had no revenue from the licences as of May 20 2026 and made 'a fraction of the allowed shipments' in the quarter to Jul 26 2026, under 1% of Data Center revenue; small batches let into the mainland by August (FT, which also reports ~10,000 each received by ByteDance and Tencent, without saying where); no China data-center compute revenue in the Q3 FY2027 outlook · the 25% tariff in force since Jan 15 2026; the 15% revenue expectation uncodified, per Nvidia
Structure
a licence paired with a tariff: BIS moved chips below 21,000 TPP and 6,500 GB/s of DRAM bandwidth (the H200, AMD's MI325X), if commercially available in the US when the rule was published, from a presumption of denial to case-by-case review for exports from the US to China and Macau, on certified conditions, among them China shipments no more than 50% of the chip's cumulative US end-use, by aggregate TPP; enough US supply and no diverted foundry capacity; customer screening; a third-party lab's review of every shipment in the US before export. Exported from the US and reviewed there, each licensed chip is imported into the US first, where a Section 232 proclamation charges 25% ad valorem on chips in two narrow performance bands ('certain advanced computing chips, such as the NVIDIA H200 and AMD MI325X', in the White House fact sheet) unless they are imported for US data centers or other US uses the proclamation lists or the Commerce Secretary determines. ECRA bars fees 'in connection with' a licence application; the 15% of revenue officials expected on the H20 was not put in a regulation, per Nvidia's filings
Parties
US government licensor (Commerce's BIS) + collector of the 25% tariff, the cut the President announced (officials' earlier 15% expectation not codified, per Nvidia) · Nvidia licensee — and the payer of the tariff, which it says it has not passed on and does not expect to
Record
disclosed / primary source · as of Aug 2026

Flows 1 as the record states them

  • Nvidia → US governmenttariff25% ad valorem on each licensed H200 as it is imported into the US for the inspection the licences require — the President's '25%'; charged on the chip's value as imported, which no customs, company or government document shows to equal the China sale price, and borne by Nvidia. The sum paid is undisclosed: the shipments it applied to were under 1% of Q2 FY2027's $89.0B Data Center revenue. Officials' earlier expected '15% or more' of licensed revenue, per Nvidia (its filings first said 15%, on the H20, in Aug 2025), has no published regulation in Nvidia's filings and no recorded payment

Resolved figures 14 each with the verbatim line from its document

  • tpp-threshold21,000
    the rule's line: chips with a TPP below 21,000 and total DRAM bandwidth below 6,500 GB/s — 'such as the NVIDIA H200 or AMD MI325X' — move from a presumption of denial to case-by-case review for exports from the United States to China and Macau, if the applicant meets the rule's conditions (volume-condition, us-testing); every other export, reexport or in-country transfer of these chips to or within Macau or a Country Group D:5 destination stays under a presumption of denial, as do exports to end users outside them that are headquartered, or whose ultimate parent is headquartered, in Macau or a D:5 destination. Published and effective Jan 15 2026; BIS's release of Jan 13 says the rule 'follows President Trump's December 8, 2025 announcement'
    “Specifically, for advanced computing commodities with a TPP less than 21,000 (as defined in Technical Note 2 to 3A090.a and 3A090.b), and a ‘total DRAM bandwidth’ less than 6,500 GB/ s (as defined in the notes to paragraph (dd)(1) in supplement no. 2 to part 748), such as the NVIDIA H200 or AMD MI325X, this final rule specifies certain conditions that, if satisfied, allow for license applicants to move from a presumption of denial to a case-by-case license review policy for exports from the United States destined to China or Macau.”
    BIS final rule (Federal Register, Jan 15 2026) · govinfo.gov
  • volume-condition50%
    the quantity condition, which the applicant certifies like the others and for which it must also 'supply evidence to BIS': the computing power (aggregate TPP) exported to China or Macau may be no more than 50 percent of the computing power shipped to customers for end use in the United States, counted from when the chips first shipped to US customers to the date of the application — a ratio, not a number of chips, per customer or in total (the press figures for both are contested: per-customer-cap, china-ceiling). The rule's preamble puts it as China shipments of 'no more than 50% of the total product shipped to customers for end use in the United States of that product'
    “the aggregate TPP of ‘‘advanced-node integrate circuits’’ exported to China or Macau will be no more than 50 percent of the aggregate TPP shipped to customers in the United States for end-use in the United States for the same advanced computing commodities from when such circuits started shipping to commercial U.S. customers in the United States for end-use in the United States to the time of the license application”
    BIS final rule (Federal Register, Jan 15 2026), supplement no. 2 to part 748, paragraph (dd)(1)(iii) · govinfo.gov
  • us-testing
    the other conditions the exporter certifies include enough US supply of the chip, no diversion of foundry capacity that US customers would use, security procedures at the recipient (customer screening against unauthorized remote access), and independent third-party testing in the United States; the rule's text lists more, among them no military, military-intelligence or weapons end use or end user (and no remote access for military end users), a list of intended remote cloud users in China and other listed countries, no passing of model weights trained on the chips to undisclosed users by cloud providers among the buyers, and a description of the recipient's physical security. In the rule's text, every shipment is reviewed before export by a qualified third-party testing lab whose name and US address the applicant gives BIS; its preamble lets the lab do that review on 'a representative sampling of a batch'. Nvidia's filings describe the result as 'an inspection process in the United States prior to any shipment to the customer', and give it as the reason each licensed chip is imported into the US (tariff)
    “The semiconductors covered by this rule are the Nvidia H200 and its equivalents, as well as less advanced chips—provided that (1) the semiconductors are commercially available in the United States at the time of publication of this rule and (2) the exporter certifies that: there is sufficient supply of this product in the United States; production of this product for exports to China will not divert global foundry capacity for similar or more advanced products for end users in the United States; the recipient has demonstrated sufficient security procedures; and the item undergoes independent, third-party testing in the United States to verify its performance specifications.”
    BIS final rule (Federal Register, Jan 15 2026) · govinfo.gov
  • tariff-instrument25%
    the instrument: a Section 232 proclamation of Jan 14 2026 (Proclamation 11,002, per CRS) — 25 percent ad valorem on 'Covered Products', for goods entered from 12:01 a.m. EST on Jan 15 2026, on top of other duties except other Section 232 tariffs and the reciprocal and border tariffs, which clause (5) displaces. Covered Products are certain advanced chips and derivative products in two performance bands (HTSUS 9903.79.01, per CBP's CSMS #67400472: TPP above 14,000 and below 17,500 with DRAM bandwidth above 4,500 and below 5,000 GB/s, or TPP above 20,800 and below 21,100 with bandwidth above 5,800 and below 6,200 GB/s); the White House fact sheet names 'the NVIDIA H200 and AMD MI325X'. The duty does not apply to imports for US data centers (over 100 MW of new AI load, per CBP), repairs or replacements in the US, US research and development, startups, non-data-center consumer and civil-industrial uses in the US, the US public sector, or other uses the Commerce Secretary determines strengthen the US technology supply chain or domestic manufacturing of chip-derived products. The Staff Secretary's example the same day: 'semiconductors that were transshipped through the United States to other foreign countries would be subject to that 25-percent tariff'
    “Except as otherwise provided in this proclamation, imports of Covered Products will be subject to a 25 percent ad valorem duty rate.”
    Proclamation of Jan 14 2026, clause (3) · whitehouse.gov
  • tariff25%
    every H200 shipped under the licences is subject to it: the rule's case-by-case review covers exports from the United States, and the licences require the H200s to go through an inspection process in the US before any shipment to the customer (the rule's preamble lets the lab review a sample of each batch), so each one is first imported into the US, and 'any H200s shipped under the new licensing program are subject to a 25% tariff upon importation' — the FY2026 10-K (Feb 25 2026) and the Q1 FY2027 10-Q carry the same sentence in the future tense, and AMD's Q2 2026 10-Q says the same of its MI325 licences ('a 25% tariff upon importation into the United States for the inspection'). A duty on the chip as imported, not a levy on China sales revenue: no customs, company or government document gives the dollar value it is assessed on or shows that value to equal the China sale price — press accounts that the US 'would receive 25 per cent of the revenue' (Reuters) assert it without a valuation — and no CBP or Treasury figure for what this tariff has collected is on file. This 10-Q and the proclamation's clause (3) each support this reading; reading it as a share of sales is the President's framing ('25 percent on the sale of those chips') and the press's (usg-cut)
    “The licenses require that the H200s go through an inspection process in the United States prior to any shipment to the customer. As a result, any H200s shipped under the new licensing program are subject to a 25% tariff upon importation into the United States.”
    Nvidia Q2 FY2027 10-Q (Aug 26 2026) · sec.gov
  • tariff-incidence
    Nvidia bears it, and expects to keep bearing it. The 10-K had warned it 'may not be able to pass along all or any of the tariff to our customers'; before the tariff existed, the Q3 FY2026 10-Q (Nov 19 2025) had said the same of 'that fee', if the government were to implement 'a regulation requiring us to pay a percentage of the revenue from such sales'. How much it has paid is not disclosed
    “We have been unable to pass along any of the tariff to our customers, and do not anticipate doing so in the event we are able to sell licensed products into the China market.”
    Nvidia Q2 FY2027 10-Q (Aug 26 2026) · sec.gov
  • licences
    licences from February 2026 for 'small amounts of H200 products to specific China-based customers' — 'a license' in the 10-K, 'licenses' in the FY2027 10-Qs — with the sale held back at the other end: the PRC government restricted the sales, and Nvidia has been unable to sell all it is licensed to. The same filing says the H20 licences of August 2025 ended the same way ('we were unable to sell our H20 inventory'). Which customers, and how much each licence allows, Nvidia does not say (approved-buyers, per-customer-cap)
    “Beginning in February 2026, the USG granted licenses that would allow us to ship small amounts of H200 products to specific China-based customers, but such sales were restricted by the PRC government, and we have been unable to sell all the products for which we have licenses.”
    Nvidia Q2 FY2027 10-Q (Aug 26 2026) · sec.gov
  • no-revenue-to-may
    no revenue under the H200 licences in their first months: the FY2026 10-K (Feb 25 2026) and the Q1 FY2027 10-Q (May 20 2026) say so in the same words, and the Q1 FY2027 CFO commentary reports that 'No shipments of Data Center Hopper products to China occurred during the quarter' to Apr 26 2026, 'compared with $4.6 billion in the first quarter of fiscal year 2026'. The first shipments came in the next quarter (q2-shipments-ceiling)
    “To date, we have not generated any revenue under the H200 licensing program, and do not yet know whether any imports will be allowed into China.”
    Nvidia Q1 FY2027 10-Q (May 20 2026) · sec.gov
  • q2-shipments-ceiling1%
    the first shipments, in the quarter to Jul 26 2026: 'a fraction of the allowed shipments', made after the H200 charge, accounting for less than 1% of Data Center revenue — under about $0.89B of the quarter's $89.0B Data Center revenue (our arithmetic on the bound; Nvidia gives no dollar or unit figure). The CFO commentary says the same of all Data Center Hopper shipments to China that quarter, and Bloomberg's Aug 26 story ('Nvidia Steps Back Into China Market With First H200 Chip Sales') reports the same bound. On Jul 14, Under Secretary Kessler had told the House Foreign Affairs Committee that, to his knowledge, shipments of H200s or equivalent chips to China or Hong Kong had begun but were 'very few'
    “After incurring that charge, we have made a fraction of the allowed shipments under the USG’s H200 licensing program. Those shipments account for less than 1% of Data Center revenue in our most recent quarter.”
    Nvidia Q2 FY2027 10-Q (Aug 26 2026) · sec.gov
  • h200-charge$0.4B
    $0.4B charged in the six months to Jul 26 2026 for excess H200 inventory and purchase obligations, 'as the demand for H200 products diminished' — the only H200 inventory figure Nvidia gives (no units; the press counts are contested in h200-stock)
    “During the first half of fiscal year 2027, we incurred a $0.4 billion charge associated with H200 for excess inventory and purchase obligations, as the demand for H200 products diminished.”
    Nvidia Q2 FY2027 10-Q (Aug 26 2026) · sec.gov
  • outlook
    the outlook counts none: the same sentence stands in the outlooks for the first, second and third quarters of fiscal 2027 (the Q4 FY2026, Q1 FY2027 and Q2 FY2027 CFO commentaries) — the last issued after the first H200 shipments, with third-quarter revenue expected at $108.0B, plus or minus 2%
    “We are not assuming any Data Center compute revenue from China in our outlook.”
    Nvidia Q2 FY2027 CFO commentary (8-K ex-99.2) · sec.gov
  • approved-buyers10
    about 10 Chinese companies cleared by the Commerce Department, 'including Alibaba, Tencent, ByteDance and JD.com', plus distributors including Lenovo and Foxconn — Reuters (May 14 2026, three people familiar), who added that 'not a single delivery has been made so far'. Lenovo confirmed to Reuters that it 'is one of several companies approved to sell H200 in China as part of Nvidia's export license'; a ZTE unit and two other Chinese firms were among the latest approved to buy advanced Nvidia and AMD AI chips by Jul 14 (Reuters). No list is public: asked how many licences BIS had approved, Kessler told the House committee (Jul 14) 'that information is confidential', BIS having given the committee disclosures listing the licence applications for the H200 and its equivalents, with the status of each
    “The US Commerce Department has approved about 10 Chinese companies including Alibaba, Tencent, ByteDance and JD.com to purchase Nvidia's H200 chips, according to the sources, who spoke on condition of anonymity due to the sensitivity of the matter.”
    Reuters via CNA (May 14 2026) · channelnewsasia.com
  • deliveries10,000
    about 10,000 H200s each received by ByteDance and Tencent 'in recent weeks', with a few other Chinese groups that 'could soon obtain approval' for similar amounts — the FT (Aug 18 2026, two people), in a report that small batches had been allowed into the mainland while, one of the people said, Beijing wanted most licensed chips kept outside it; it does not say where these were delivered (TNW's headline: 'though maybe not mainland China'). Reuters and TNW restate it. No company or government count is public: Kessler told the House committee that BIS knows how many approvals have led to shipments and had given it that information in confidence (Jul 14)
    “ByteDance and Tencent have each received about 10,000 H200 processors in recent weeks, while a few other Chinese tech groups could soon obtain approval for shipments of similar size, according to two people with knowledge of the matter.”
    FT via Yahoo Finance (Aug 18 2026) · ca.finance.yahoo.com
  • importer-block
    how Beijing, the importer, held it back, as reported (no PRC document on file says so): Chinese customs told agents in January that H200s were not permitted to enter China, and officials told tech companies not to buy them 'unless necessary' (Reuters, Jan 14 2026, three people); Commerce Secretary Lutnick told an April Senate hearing that, as he understood it, the 'Chinese central government has not let them as of yet buy the chips because they're trying to keep their investment focused on their own domestic industry'; by August, small batches were allowed into the mainland while, one of the FT's sources said, Beijing wanted most of the chips kept outside it — regulators told companies they could ship them to Hong Kong and use them there — and, according to a person with direct knowledge, purchases would still need separate approval through the state planner, the NDRC (FT, Aug 18). Beijing's ministries confirm nothing: asked about H200 restrictions (Jan 22), approvals (Mar 19) and sales (Jul 16 2026), the Commerce Ministry's spokespeople answered '我不了解你提到的情况' each time, saying they were unaware of it; and a company that asked the ministry's public message board in July how to import H200s lawfully, and whether a US licence could still fail at Chinese customs, got an Aug 19 reply about export licensing only. Beijing was not the only brake in the press's accounts: Reuters (May 14) found the path to a completed sale 'obstructed by a tangle of requirements on both sides', and one of its sources said the shift in China 'was partly triggered by changes on the US side', though 'exactly what changed remains unclear'; ETO reports that by early February the State Department 'began stalling the H200 export process'
    “Chinese customs authorities told customs agents this week that Nvidia's H200 artificial intelligence chips are not permitted to enter China, according to three people briefed on the matter.”
    Reuters via Air Freight News (Jan 14 2026) · airfreight.news

Contested figures 10 not reconciled — every position side by side, with its source and tier

usg-cut2 positions, side by side
15% — the cut officials named for the first licences, on the H20 in August 2025: 'USG officials have expressed an expectation that the USG will receive 15% of the revenue generated from licensed H20 sales' (Q2 FY2026 10-Q). From the Q3 FY2026 10-Q the sentence reads '15% or more of the revenue generated from licensed sales of our products', and the FY2026 10-K and annual report print it in the paragraph before the H200 licence; every version adds that no regulation codifying it was published, and the Q2 FY2027 10-Q drops the sentence. The President's account: he asked for 20%, and Huang said 'would you make it 15?' (Aug 11 2025). Nvidia earned about $60M of H20 revenue under those licences; no document shows a payment Nvidia Q2 + Q3 FY2026 10-Qs, FY2026 10-K and annual report, Q1 + Q2 FY2027 10-Qs; CNN transcript (Aug 11 2025)
25% — the President's figure for the H200: '$25% will be paid to the United States of America' (Truth Social, Dec 8 2025); 'the United States is getting 25 percent of the chips in terms of the dollar value' (Jan 14 2026, the day he signed the tariff). What collects it is the 25% tariff on each licensed chip (tariff, tariff-instrument), which CRS calls the arrangement's '25% fee component': a duty on the chip's value as imported, borne by Nvidia. Reuters, naming no tariff, calls the arrangement 'a structure that requires the chips to pass through US territory before being shipped to China, as US law does not permit the direct imposition of export fees' (May 14 2026). No customs, company or government document shows that value to equal what Chinese buyers pay, though the President, Reuters ('25 per cent of the revenue') and ETO speak of a share of the sale. Whether it is lawful is open: ECRA bars any fee 'in connection with' a licence application, and CRS finds it 'debatable' both whether Section 232 authorizes the tariff and whether it is consistent with that bar; two Nvidia shareholders sued over the arrangement in Delaware (note) Truth Social via the American Presidency Project; Jan 14 2026 remarks (DCPD-202600023); CRS LSB11409 (Mar 23 2026); Reuters via CNA (May 14 2026); ETO
per-customer-cap2 positions, side by side
75,000 H200s per approved customer — first as a limit 'the Trump administration has talked about', with AMD's MI325 chips counting toward the same cap (Bloomberg, Mar 3 2026, via the LA Times: enough for 'a midsize data center of roughly 100 megawatts'), then as licence terms: 'each approved customer can purchase up to 75,000 chips under the US licensing terms, two of them said' (Reuters, May 14 2026) Bloomberg via LA Times (Mar 3 2026); Reuters via CNA (May 14 2026)
100,000 H200s apiece — the FT (Aug 18 2026): the Chinese groups 'have been cleared by the US to buy up to 100,000 H200s apiece', in a sentence it credits to one of the people. TNW (Aug 19) sets it against the 75,000 of March and May; nothing on file records a change, and no licence is public. The rule itself sets no per-customer number, only the 50% ratio — but BIS's Kessler told the House on Jul 14 2026 that 'the approach that we've taken is to limit the quantity … of chips to any particular company', giving no number for either figure FT via Yahoo Finance (Aug 18 2026); TNW (Aug 19 2026); House hearing transcript (Jul 14 2026)
china-ceiling2 positions, side by side
1 million H200s — an overall limit for China, an 'upper bound Trump's team set earlier in the regulatory process': 'Total shipments to China could still reach as many as a million units' (Bloomberg, Mar 3 2026), 'more than what many officials would like, though far lower than the figure Nvidia originally proposed'. Not in the rule, which states a ratio Bloomberg via LA Times (Mar 3 2026)
≈787,000–1,000,000 H200s (IFP: '~787,500' if ~75% of past H200 shipments went to US use) · up to 850,000 H200s, or 890,000 H200-equivalents — 'nearly 900,000' — counting AMD's MI325X (CNAS, on an 85% US share of an estimated two million H200s sold) · 900,000 H200-equivalents, initially rising by 150,000 a quarter with US sales (IAPS, whose endnotes base the figure on CNAS's paper) — the rule's 50% applied to estimates of past sales. IFP and CNAS both start from Epoch AI's data, and Nvidia discloses no H200 unit sales, so these are models on one unsettled base, not confirmations; no BIS count is public IFP; CNAS; IAPS
beijing-approvals2 positions, side by side
more than 400,000 H200s in total for ByteDance, Alibaba and Tencent, approved in late January 2026, 'with other firms now joining a queue for subsequent approvals' (Reuters, Jan 28 2026, sources) — but only 'with conditions' that 'were still being decided upon', and a fifth source said 'the licenses were too restrictive' and customers 'were not yet converting the approvals to purchase orders': a permission to buy, not deliveries. Huang told reporters in Taipei the next day that Nvidia had not received such information and that he believed China was still finalising the licence (Reuters, Jan 30); asked to confirm, the Foreign Ministry referred questions to China's competent authorities (Jan 29) Reuters via Yahoo Finance (Jan 28 2026) and WTVB (Jan 30 2026); FMPRC (Jan 29 2026)
fewer than 200,000 in total, still being decided in July: Beijing 'is still determining the exact number of Nvidia chips to approve, and it could amount to fewer than 200,000 in total', 'less than half of what the companies requested earlier this year' (The Information, Jul 8 2026, relayed by Reuters; NST's copy is dated Jul 9) — a report about Alibaba, ByteDance and DeepSeek, whom officials had told in recent weeks that they 'may soon receive permission'. The same relay says Chinese officials 'have withheld approval so far', though Reuters had reported in March that Nvidia had won Beijing's approval and Huang told CNBC it had clearance. Set against January's figure, nothing on file reconciles the two: Alibaba and ByteDance are approved in January's report and still awaiting permission in July's, and so is DeepSeek, whose conditional approval Reuters reported two days after it (Tencent is only in January's). Whether January's figure was a conditional approval later scaled back, what the companies requested, or a premature report, no Chinese approval document says The Information via Reuters / NST (Jul 9 2026)
h200-stock2 positions, side by side
700,000 at the end of 2025 — 'Nvidia currently holds just 700,000 units in stock', against Chinese orders for 'more than 2 million H200 chips for 2026' (Reuters, Dec 31 2025, two people); around 100,000 of the 700,000 are GH200 Grace Hopper superchips (one of the people). Tom's Hardware's 'over 80,000 H200 chips available in existing stock' (Jan 13 2026) names no source in its text (its 'reportedly' links to an earlier Tom's Hardware story that was not read), and IFP's ~950,000 adds ~250,000 it says were made by early March 2026, citing by link an FT story that was not read, and assumes no sales — while a third source in the same Reuters story said new work at TSMC was 'expected to start in the second quarter of 2026' Reuters via The Standard (Dec 31 2025); Tom's Hardware (Jan 13 2026); IFP
about 500,000 in August 2026 — 'Nvidia has about half a million H200 chips in stock mainly for Chinese clients. Sales were held up due to Beijing's restrictions' (FT, Aug 18 2026). Nvidia discloses no unit count; its one figure is the $0.4B charge for excess H200 inventory and purchase obligations (h200-charge). Both are press counts of Nvidia's stock of the chip, months apart, and nothing on file connects them: whether the gap from 700,000 is sales, write-downs or a different count, no document says FT via Yahoo Finance (Aug 18 2026)

Scope note what this record leaves to other records

NOT this record (not carried here): the April 2025 H20 licence requirement and its charge — $4.5B recorded in Q1 FY2026; the 8-K's $5.5B was an 'up to' estimate — Nvidia's China business, as intel-usgov also left it; AMD's parallel case — the same 15% expectation for its MI308, about $390M of MI308 revenue to China in Q4 2025 (AMD's Q4 figure, which Epoch restated as 2025's; AMD's 10-K dates its first licensed MI308 shipments to Q4 2025, and its Q3 10-Q records no MI308 revenue from China-based customers), and its MI325 licences under the same rule and tariff (cited in tariff only as corroboration) — for an AMD record; Nvidia's China geographic revenue (on a customer-headquarters basis since Q3 FY2026, all products, not H200); Chinese buyers' capex and compute prepayments; China market-share and smuggling estimates; H200 prices and Chinese rental tenders; payment terms (Reuters' full-upfront report against Nvidia's denial); H200-versus-H20 performance multiples; BIS enforcement figures. Not a ribbon: the tariff runs from Nvidia, in the money chart's chips column, back to the government in its investors column, a direction the chart (drawn from investors down the chain) cannot place, and its size is undisclosed — so it is a flow here and not a link, as kr-nvidia-260k draws none. Litigation: in January 2026 two Nvidia shareholders sued the company in Delaware state court over the arrangement, 'claiming that ECRA prohibits the government from requiring payments in connection with an export license', and asked for company records rather than to undo it (CRS); no court filing is on file. Corrections from the sources: the 25% is not press alone — Nvidia's 10-K and both FY2027 10-Qs state it as 'a 25% tariff upon importation', naming no instrument (the record ties it to the Jan 14 2026 proclamation through the White House fact sheet, which names the H200), and the December 2025 item was a Truth Social post, not a proclamation; the ~10,000 each for ByteDance and Tencent is the FT's (Aug 18), while Bloomberg's Aug 26 story, as available on file (a Bloomberg Tax teaser and two partner editions, Bloomberg Technoz's in Indonesian and Bloomberg Businessweek Vietnam's, which says Nvidia gave no count of the H200s it sold in China), reports Nvidia's under-1% and no per-company count — and both partner editions date the US licences to January, where Nvidia's filings say February; the wording '15% or more' on the H20 follows the annual report, whose sentence sits in its H20-licence paragraph but covers 'licensed sales of our products' — the H20-only wording was '15% of the revenue generated from licensed H20 sales' (Aug 2025); and the annual report's words are 'the USG will receive', not 'the U.S. government will receive'. Not adopted: ETO's 'the Chinese government has authorized the sale of 400,000 H200 chips' (drops Reuters' 'more than' and states a sourced report as an authorization); a reporter's 'the 50 percent stake will go to the government' (a question's premise at the Aug 12 2025 White House briefing); IAPS's 'around $6.75 billion' fee maximum (a model figure; no collection figure is on file).

More between US government and Nvidia

  1. Jun 15, 2026 updated Nov 18

Everything between US government and Nvidia →