Trump-Xi summit and Nvidia H200 chip exports to China

The headline contradiction of the Trump-Xi summit is that Washington has done everything it can to enable Nvidia H200 sales to China — and China has done everything it can to refuse them. The U.S. Commerce Department has cleared roughly 10 Chinese firms — Alibaba, Tencent, ByteDance, JD.com, plus distributors Lenovo and Foxconn — to buy up to 75,000 H200 chips each^[1]. Yet as Commerce Secretary Howard Lutnick told a Senate Appropriations subcommittee in April, 'we have not sold them any chips as of yet'^[2]. Trump himself confirmed the standoff on the tarmac in Beijing, saying China 'chose not to' buy and 'they want to develop their own'^[3]. The mechanism is not market drift; it is deliberate. China's State Council has launched a parallel supply-chain security review and quietly told tech firms to pause H200 orders so capex stays inside the country^[1]. That has reduced what was supposed to be a $3.5–$4 billion annual revenue recovery for Nvidia to a paper victory^[4]. The licenses are real; the shipments are not.

The financial architecture behind the H200 reversal is itself unprecedented. When the Bureau of Industry and Security codified the new licensing regime on January 13, 2026, it required Nvidia to remit 25% of every H200 sale to China back to the U.S. Treasury^[4][5]. That fiscal hook gave the White House a politically defensible story — exports as a revenue stream rather than a security concession — and the same template has been floated for AMD and Intel. But the structure also handed Beijing a clean argument for self-reliance: every H200 shipped is a 25% tax flowing to the U.S. government, on top of a chip that Chinese policymakers can plausibly characterize as already a generation behind Nvidia's Blackwell line. That made the political cost of compliance high and the strategic upside low. White House AI czar David Sacks conceded the dynamic publicly back in December, saying China is 'rejecting' H200s and 'outfoxing' U.S. strategy^[6]. The revenue share was designed to make exports easier to defend in Washington; it ended up making them easier to reject in Beijing.

The most underappreciated number from this cycle is that Chinese manufacturers now hold roughly 41% of China's AI accelerator server market^[2]. That share — built largely on Huawei's Ascend line and momentum from DeepSeek-style domestic models — is the engine of Beijing's confidence. Lutnick told senators Beijing is steering investment toward Huawei and other domestic suppliers^[2], and Nvidia's own CEO has acknowledged the collapse of its China business, with Jensen Huang saying 'in China, we have now dropped to zero'^[7]. From Beijing's perspective, allowing a $3.5–$4 billion H200 spend now would not just route cash and 25% of margin to Washington — it would freeze domestic chipmakers out of their best customers right as they hit scale. The State Council's pause directive is therefore less a diplomatic snub than industrial-policy hygiene. Council on Foreign Relations analysts argue the H200 carveout 'risks turbocharging China's development of the most advanced AI models'^[8]— but for now China is choosing the opposite trade: slower frontier compute in exchange for a permanent domestic supply chain. The X.com and Reddit discourse echoes this read; investor and hardware communities openly describe the H200 as a generation-old chip whose license can be 'yanked on a whim,' which makes Huawei look like the safer long-horizon bet.

The optics of the trip captured how thoroughly chip policy has fused with personal diplomacy. Huang was not originally on the delegation; Trump phoned him after media coverage of his absence and invited him to join, with Nvidia's CEO boarding Air Force One during a refueling stop in Alaska^[9][10]. Inside Beijing, Huang's brief was narrow and unmet: secure a Chinese green light on already-approved H200 orders and unlock a reciprocal easing on rare-earth magnets and gallium^[1]. Neither materialized. The summit yielded warm rhetoric — Trump said 'a lot of different problems were settled' — but no concrete chip-export or rare-earth deal^[4]. Markets had front-run a different outcome: Nvidia stock hit an intraday high of $227.84 on May 14 on summit anticipation^[7]. The political fallout in Washington is sharpening: Sen. Elizabeth Warren has framed the trip as a corruption risk, and Sen. Chris Coons has opened a formal investigation into Lutnick's statements on the license process^[11][12]. Social commentary on X has fixated on Trump's personal Nvidia stockholdings as the next angle. Even Nvidia-friendly investor forums note the asymmetry of the readouts — Washington itemized what it wanted China to buy, while Beijing's summary mostly emphasized Taiwan.

Even if H200 shipments stay frozen, the policy framework is reshaping incentives across the AI supply chain. CFR estimates that if Nvidia were to export roughly 3 million H200s annually under the new regime, China's incrementally added AI compute could at least triple and gain a two-to-three-year boost — a meaningful narrowing of the gap with U.S. frontier labs^[8]. The H200 is also more than six times more powerful than the best U.S. AI chip currently legal in China^[8]. CFR's Chris McGuire calls the licensing regime itself 'strategically incoherent and unenforceable,' arguing Commerce cannot credibly verify end-use claims and that several likely buyers, including Tencent, have documented military relationships^[13]. Hardware community discussion on Reddit reinforces the enforcement gap: H200s are reportedly being scalped on grey markets, and Singapore and Malaysia data centers are widely treated as smuggling proxies. Meanwhile rare-earth exports from China are reportedly running ~50% below pre-restriction levels, undermining the reciprocity the H200 deal was supposed to secure^[4]. The net second-order picture: Washington has built a regime that lifts a strategic constraint without delivering the revenue, the reciprocity, or the enforceability it was sold on — while giving Beijing both political cover to consolidate its domestic chip stack and an opportunity to seed a 'second DeepSeek moment' trained on Huawei Ascend silicon.