Nvidia's H200 Paradox: U.S. Clears China Sales, But Not a Single Chip Has Been Delivered

The paperwork was done. The export licenses were approved. The company names were listed. And yet, not a single Nvidia H200 artificial intelligence chip has crossed into Chinese hands.

That is the extraordinary paradox at the heart of Reuters’ exclusive reporting this week: the United States government has formally cleared approximately ten of China’s largest technology companies to purchase Nvidia’s H200 AI processors under a controlled licensing framework — and not one delivery has been completed. The companies cleared include some of China’s most powerful technology giants: Alibaba, Tencent, ByteDance, and JD.com, according to three sources familiar with the matter.

The terms of the deal were significant. Each approved Chinese customer would be permitted to purchase up to 75,000 H200 chips per year, subject to a 25 percent surcharge — a Trump administration innovation that treats premium semiconductor access as a revenue stream rather than purely a security firewall. On paper, it was a historic concession: these chips are more capable than the H100, which was restricted for export to China in 2023, and the administration’s willingness to approve them represented a meaningful shift in posture.

Beijing, it turns out, has other ideas about what its companies should be buying.

The Counter-Directive

Three people familiar with the situation told Reuters that the H200 deals have stalled because Chinese firms received informal guidance from Beijing discouraging them from placing orders. The Information reported further that Beijing has explicitly instructed several major technology companies to halt H200 purchase discussions entirely.

The Chinese government’s rationale flows directly from its industrial policy priorities. State Council directives require audits of foreign dependencies in critical technology infrastructure, with explicit concerns flagged about potential backdoors or supply chain vulnerabilities embedded in U.S.-sourced chips. The fear — whether strategically real or politically convenient — is that hardware sourced through American manufacturers could be compromised or subjected to future disruption in the next escalation of trade tensions.

That security argument is compounded by a commercial reality Beijing can now credibly point to: China’s domestic AI chip industry has made remarkable strides. Huawei Technologies and DeepSeek-affiliated hardware partners have collectively pushed Chinese manufacturers to 41 percent of China’s AI accelerator server market — up from near-zero just three years ago. With domestic alternatives now commercially viable, even if not yet at H200 performance levels for the most demanding workloads, Beijing sees little strategic benefit in deepening its tech giants’ dependency on U.S. silicon.

U.S. Commerce Secretary Howard Lutnick acknowledged the root cause at a Senate hearing last month. Beijing is “directing investment toward its own domestic industry,” he said — a candid admission that the stalled H200 sales are less about technical objections than deliberate industrial policy.

The Summit That Changed Nothing — Yet

The timing of the U.S. approval made the chip diplomacy stakes unmistakable. Jensen Huang, Nvidia’s chief executive officer, was not initially included in the White House’s business delegation traveling to Beijing for the Trump-Xi summit. He was added at the last minute, personally invited by President Trump, who reportedly picked Huang up in Alaska during the trans-Pacific flight — a signal that unlocking H200 sales had become a diplomatic priority at the highest level.

The summit produced optimistic language. Both governments described constructive conversations on trade and technology cooperation. Jensen Huang was photographed alongside Chinese officials, and there was genuine expectation in the semiconductor industry that a breakthrough on H200 deliveries was imminent.

Three weeks on, the deals remain frozen. Chinese companies have received U.S. approval but not Chinese permission to proceed. No chips have been delivered.

A Market That Already Left

For Nvidia, the H200 standoff is the latest chapter in a story of accelerating market loss in China that predates this specific episode. The company’s China revenue fell 21.2 percent year-over-year in fiscal 2025 — the first full fiscal year in which comprehensive U.S. chip export restrictions were in force. Huang has publicly acknowledged that Nvidia’s share of China’s high-end AI chip market has “effectively collapsed to zero.”

The scale of that collapse is remarkable against Nvidia’s historical position. At its peak, the company controlled approximately 95 percent of China’s market for high-performance AI chips, with China accounting for roughly 13 percent of total global revenue. That Chinese revenue has been largely replaced by surging demand from American hyperscalers — Microsoft, Amazon, Google, and Meta are consuming H100 and H200 output at rates that keep production backordered for months — but the strategic loss of the Chinese market carries long-term implications for competitive positioning.

The H200’s Unusual Regulatory Position

The H200 sits in a carefully engineered regulatory middle ground. It is Nvidia’s second-most-powerful AI chip currently in production — more capable than the H100 that was restricted in 2023, but below the H100 NVL and the newer Blackwell-architecture B200 and B100 chips that remain fully restricted for export to China. The Trump administration’s decision to approve H200 exports represents a calculated calibration: enough flexibility to demonstrate commercial goodwill, but not enough to meaningfully close the frontier AI compute gap between the two countries.

The 25 percent surcharge adds another layer of complexity to the commercial calculation. Chinese companies would be paying a significant premium for hardware they are simultaneously being discouraged from acquiring by their own government. For companies like Alibaba and Tencent, which have invested tens of billions of dollars in Huawei’s Ascend chip ecosystem and in internally developed AI accelerators, the arithmetic does not obviously favor taking the deal — even if Beijing were to lift its informal guidance.

What the Industry Is Watching

Analysts are tracking two timelines simultaneously. The first is whether any H200 deliveries materialize before export licenses expire — typically 12 to 24 months after issuance. If the licenses lapse without deliveries, the political window for this specific arrangement closes.

The second and more consequential question is whether Chinese firms’ capital expenditure plans for 2027 shift if domestic AI chip alternatives fail to keep pace with Nvidia’s next-generation Rubin architecture, which is already in early production. Huawei’s Ascend 910C has narrowed the gap with the H100, but the H200’s successor-class performance remains significantly out of reach for domestic Chinese alternatives.

The broader geopolitical question is whether the Trump administration’s licensing framework — approve exports with per-company quotas, surcharges, and supply chain monitoring requirements — can be sustained as a model for managing advanced semiconductor access. It represents an attempt to keep U.S. chipmakers commercially viable in China while preserving strategic technology advantage. Whether China’s tech giants choose to participate in that framework, or continue building around it with domestic alternatives and strategic stockpiles, will shape the AI chip market for the rest of the decade.

For Nvidia, the paradox is a familiar one in the current era of technology statecraft: the U.S. government finally gave them permission to sell. China just isn’t buying.