Khan Capitals | July 2026
Key Takeaways
- The shipments have started. Nvidia H200 China exports have formally begun, the US Commerce Department’s Jeffrey Kessler told Congress on 14 July, describing the volumes shipped so far as “very few” and “trivial”.
- The paper trail is anything but trivial. Roughly $10 billion of export licences have been approved since conditional clearances began in December 2025, covering sales to around ten Chinese firms including Alibaba, Tencent, ByteDance and JD.com.
- Demand dwarfs supply. Chinese technology companies have reportedly ordered more than two million H200 chips for 2026, against total Nvidia inventory of roughly 700,000 units: the licence is no longer the binding constraint, the silicon is.
- The loophole is the political flashpoint. Lawmakers from both parties pressed Commerce over the “Blackwell loophole”, under which offshore subsidiaries of Chinese firms bought top-end chips until the gap was closed on 31 May 2026; chips already delivered remain in use.
- The market bought the signal. Nvidia shares rose 4 per cent to $211.80 on the testimony, a $5.13 trillion market capitalisation, as investors priced the reopening of Chinese revenue that export rules had written down to zero.
Part of: The AI Infrastructure Supercycle – Khan Capital’s hub on the AI infrastructure build-out.
Nvidia H200 China Exports Begin, Barely
The most consequential sentence in markets this week was delivered in a congressional hearing room, not a trading floor. Jeffrey Kessler, under secretary of commerce for industry and security, confirmed to the House Foreign Affairs Committee on Tuesday 14 July that shipments of Nvidia’s H200 artificial intelligence accelerators to Chinese customers have begun under the government’s new case-by-case licensing regime. He was at pains to minimise the scale, calling the number of chips shipped “very few” and later “trivial”, and declining to name buyers or quantities.
Trivial volumes, however, sit atop a non-trivial stack of paper. Approvals under the regime, which began issuing conditional clearances in December 2025, now cover roughly $10 billion of hardware for about ten Chinese firms, a list reported by Bloomberg to include Alibaba, Tencent, ByteDance and JD.com. The Nvidia H200 China exports story is therefore not about what has moved, which is almost nothing, but about what has been authorised, which is a great deal, and about what has been ordered, which is more than the company can physically supply.
“Case-by-case” is doing a great deal of work in that sentence. Under the revised review policy published by the Bureau of Industry and Security, each application is judged against three tests: the export must not reduce the semiconductor capacity available to US customers, the Chinese purchaser must operate demonstrable export-compliance procedures, and the product must pass independent third-party testing. Those conditions give Washington a throttle rather than a switch, and they explain the gap between $10 billion of approvals and a trivial shipment count: a licence is permission to queue, not a delivery date. They also make the regime auditable in public, which is why a routine oversight hearing, rather than a company filing, became the moment the market learned the hardware had moved.
| Measure | Figure |
|---|---|
| Export licences approved | ~$10 billion since December 2025 |
| Approved Chinese buyers | ~10 firms, incl. Alibaba, Tencent, ByteDance, JD.com |
| Chinese H200 orders for 2026 | 2,000,000+ chips (reported) |
| Nvidia H200 inventory | ~700,000 units (reported) |
| Chips actually shipped | “Trivial” per Commerce testimony, 14 July |
| NVDA reaction, 14 July | +4% to $211.80; $5.13trn market cap |
Four Years of Ratchets, One Reversal
Some history clarifies how unusual this week is. Since October 2022, US policy on AI chips for China has moved in only one direction: tighter. The first broad controls barred Nvidia’s top-end accelerators outright. When the company engineered cut-down China-specific parts to fit under the thresholds, the October 2023 revision barred those too, and the H20, the next compliant workaround, spent 2024 and 2025 in and out of regulatory favour. December 2025’s conditional clearances for the H200 were therefore the first structural loosening of the controls era, and this week’s testimony is the first official confirmation that hardware has actually moved under them. One reversal after four years of ratchets does not make a trend, but it does make a precedent, and markets price precedents.
| Date | Policy step | Direction |
|---|---|---|
| October 2022 | First broad controls; top-end AI chips barred from China | Tighter |
| October 2023 | China-specific workaround parts brought under controls | Tighter |
| December 2025 | Case-by-case H200 licences for ~10 Chinese firms begin | Looser |
| 31 May 2026 | Blackwell loophole closed; ultimate-parent rule applied | Tighter |
| 14 July 2026 | Commerce confirms first H200 shipments, “trivial” volumes | Confirmation |
Two Million Orders, Seven Hundred Thousand Chips
The arithmetic beneath the testimony is the striking part. Chinese technology companies have collectively placed orders for more than two million H200 units for 2026. Nvidia’s available inventory of the part is estimated at roughly 700,000. Even if Washington approved every licence application tomorrow, China’s hyperscalers could receive barely a third of what they have asked for, and in practice they will receive far less, because the H200 is last-generation silicon that Nvidia is winding down in favour of newer architectures whose sale to China remains prohibited.
That inversion matters for how investors should read the policy. For three years the constraint on Nvidia’s China revenue was legal: demand existed, sales were barred. As of this week the constraint is physical: sales are permitted in principle, and the products do not exist in sufficient quantity. An order book of two million units against 700,000 of supply is also a live demonstration, for anyone still unsure, that Chinese demand for accelerated computing survived every wave of restriction intact, and simply queued.
The Blackwell Loophole and the Credibility Problem
The hearing’s sharper exchanges concerned the chips that moved without licences. Until 31 May 2026, US export rules keyed on the shipping destination rather than the buyer’s ultimate ownership. Subsidiaries of Chinese AI firms incorporated in Malaysia, Singapore or the United Arab Emirates could therefore buy Blackwell and even Rubin-class hardware, Nvidia’s current and next generations, without triggering the controls at all. Industry estimates cited around the hearing suggest hundreds of thousands of chips may have moved through this route over roughly a year, and a former State Department official described the acquisitions as happening “very likely at scale”.
The Commerce Department’s 31 May guidance closed the gap prospectively, extending licence requirements to any purchaser whose ultimate parent is headquartered in China regardless of shipping address. It did not, and realistically could not, claw back hardware already installed: data centres running loophole-era Blackwells keep running them. The result is a policy with a credibility asymmetry that both parties in Congress find uncomfortable. The controlled channel has shipped a trivial number of last-generation chips under heavy conditions, while the uncontrolled channel delivered frontier hardware at scale before it was shut. Bipartisan criticism at the hearing focused less on the H200 licences themselves than on what the loophole says about enforcement.
What Each Side Is Actually Buying
Strip away the theatre and the regime has a discernible logic. Washington is selling time: the H200 is a generation behind the frontier, every unit sold is revenue for an American champion, and the licence conditions, no diversion of supply from US customers, compliance procedures at the buyer, independent third-party testing, are designed to make each shipment traceable. Beijing is buying capacity: two million units of last-generation silicon still trains and serves models, and every H200 that arrives reduces the urgency gap while domestic alternatives mature. Both sides can describe the same flow as a win, which is usually the condition for a policy to persist.
For Nvidia the equation is simpler. China was once roughly a fifth of its data centre revenue before controls wrote it toward zero, and the reopening converts a written-off market into an option with a government-controlled strike. The 4 per cent move on 14 July, worth roughly $200 billion of market value, is the market marking that option up, not pricing confirmed shipments. It landed in a week when the AI complex needed the reassurance: semiconductors had been under pressure since Samsung’s record quarter was sold, even as TSMC’s results confirmed the build-out’s substance and SK Hynix’s record listing confirmed the capital queuing behind it.
The Precedent Problem
The deeper question is whether case-by-case licensing is a stable equilibrium or a waypoint. The regime’s defenders present it as calibrated leverage: chips as a diplomatic instrument, dialled up or down with the relationship. Its critics, on both sides of the aisle, note that instruments cut both ways. A revenue stream that exists at the discretion of a licensing office is a revenue stream that can be revoked by one, as holders of supply agreements across the AI complex have learned repeatedly this cycle. Investors should resist extrapolating either way from week one: the licences prove the door can open, the “trivial” volumes prove it has barely done so, and the loophole episode proves the walls matter more than the door.
There is also a competitive clock running. The case-by-case policy covers AMD’s MI325X on similar terms, and China’s domestic accelerator programmes are the unpriced variable: every quarter of restricted supply is a quarter of protected demand for home-grown alternatives. The H200 window is, in that sense, perishable inventory for American vendors. Selling a generation-old part into a market that is actively building its replacement is good business exactly as long as the replacement is not ready.
The H20 episode is the instructive precedent. Nvidia’s earlier China-compliant part swung between permitted and restricted as the political weather changed, and each swing moved billions of dollars of guided revenue on and off the company’s books within a single earnings cycle. Analysts who treated any one policy setting as permanent were wrong within quarters, in both directions. The sensible prior for the H200 regime is the same: a setting, not a settlement, with the licence volume as the dial. That argues for watching the administrative record, licence approvals, enforcement actions, hearing schedules, at least as closely as the product roadmap.
Investor Implications
Equities. The 14 July move reprices Nvidia’s China optionality, not its China earnings; the distinction matters for how much disappointment a “trivial” shipment run-rate can inflict on the next results. The read-across is broader than one name: AMD holds parallel licences, and the memory complex, SK Hynix, Micron, Samsung, supplies whichever accelerators ultimately ship. Investors may wish to treat disclosed China revenue in the coming earnings round as the first hard data in a story that has so far run on testimony.
Fixed income. Export policy is now a named risk factor in the AI infrastructure credit stack. The $35 billion of AI chip financing that began trading this month is collateralised, ultimately, by hardware whose addressable market a licensing office can resize; spreads on such paper should, and increasingly do, carry a policy premium.
Cross-asset. The episode is a reminder that the US-China technology relationship is being managed, not resolved. Positioning should reflect a regime of periodic, headline-driven repricing in both directions: licences granted lift the complex, enforcement actions and loophole disclosures cut it, and neither settles the question. The volatility is the policy.
Conclusion
Nvidia H200 China exports have begun, and almost none have shipped. Both halves of that sentence are true, and the space between them is where the next year of the AI trade’s China chapter will be written. Washington has demonstrated it can open a controlled door; the loophole saga demonstrated that the uncontrolled ones mattered more; and two million orders against 700,000 chips demonstrate that demand was never the question. What the market bought on 14 July was not revenue but precedent: the first evidence since the controls era began that Chinese demand for American AI silicon can be monetised at all. Whether that precedent compounds or reverses now depends on a licensing office, a hearing calendar, and a supply chain, roughly in that order.
Frequently Asked Questions
Can Nvidia legally sell AI chips to China now?
Yes, within strict limits. Since December 2025 the US has approved export licences case by case for Nvidia’s H200 and AMD’s MI325X to roughly ten named Chinese firms, subject to conditions including no diversion of supply from US customers, buyer compliance procedures and third-party testing. Nvidia’s newer Blackwell and Rubin generations remain prohibited.
What is the Blackwell loophole?
Until 31 May 2026, US export rules were triggered by the shipping destination rather than the buyer’s ultimate ownership. Offshore subsidiaries of Chinese companies in places like Malaysia, Singapore and the UAE could therefore buy top-end Nvidia chips without licences. Commerce closed the gap by extending licence requirements to any buyer whose ultimate parent is in China, but chips already delivered remain in use.
Why did Nvidia stock rise if shipments are trivial?
The market repriced the possibility of future China revenue rather than current sales. China was once a significant share of Nvidia’s data centre business before export controls effectively removed it, so official confirmation that licensed shipments have begun restores an option the market had valued at close to zero. The 4 per cent rise on 14 July reflected that option value.
Sources: CNBC; Bloomberg; Benzinga; TechTimes; TNW; US Bureau of Industry and Security.
Related Reading: The China licence story is one thread of a broader supply picture traced in NVIDIA’s move to lock up SK hynix HBM supply and confirmed in hard numbers by TSMC’s Q2 results. For the demand-side wobble the testimony interrupted, see the week the supply premise cracked, and for the capital now queuing behind the complex, SK Hynix’s record Nasdaq listing. For how the wider complex repriced, see The Semiconductor Bear Market. The sector’s latest print is analysed in Intel’s Q2 2026 earnings, the fastest growth in 15 years.
Leave a Reply