Trump Approved H200 Chip Sales to China in January. China Refused to Buy Them. NVIDIA Stopped Making Them. Zero Chips Crossed the Border. Call the Policy Whatever You Want.

On January 15, 2026, the US Bureau of Industry and Security published Federal Register notice 91 Fed. Reg. 00789, revising its license review policy for NVIDIA H200 and AMD MI325X chips to "case-by-case review" for exports to China. Coverage called it a major reversal of Biden-era AI export controls. It was, on paper.

Five months later, zero H200 chips have been delivered to China.

The standard framing of this outcome — "export control saga continues," "chip war uncertainty," "policy in limbo" — misses what actually happened. The H200 approval exists on paper precisely because both sides had reasons to keep it there while doing something different in practice.

What Beijing decided.

When the BIS rule was published January 15, Beijing did not immediately embrace the newly permitted imports. Within days, China's State Council convened meetings to assess H200 demand. The conclusion: the BIS conditions attached to H200 exports were "too restrictive." The specific requirements — third-party US-based certification of each shipment, buyer KYC procedures, military-use certifications, model-weight transfer restrictions — created what Beijing assessed as a surveillance apparatus that would be built into China's AI infrastructure.

There was a second calculation. In April 2025, the US government cut off NVIDIA H20 chips to China mid-stream, forcing a $4.5 billion NVIDIA charge on outstanding inventory and stranding Chinese AI labs that had built deployment plans around H20 supply. The H20 cut-off was a demonstration of supply-chain vulnerability that Beijing processed carefully. A country that builds its AI infrastructure on NVIDIA hardware can have that infrastructure terminated by a Washington policy decision. China watched itself experience that outcome in real time.

The H200 reversal offered chips with conditions. China's answer — still officially deliberating as of June 2026 — is functionally: we will tolerate a near-term capability gap on Huawei Ascend to avoid recreating an H20-style dependency that a future US administration can sever.

This is not a negotiating position or a temporary hold. Huawei Ascend 910C production is targeting 600,000 units in 2026, up from near-zero two years earlier. Chinese tech companies that once ran NVIDIA-first inference pipelines have rebuilt for Ascend at scale. ByteDance, Alibaba, and Baidu all deploy Ascend at production depth. The H200 approval window — if China ever accepts the BIS terms — may find an installed base that has already migrated away.

What NVIDIA decided.

NVIDIA also made a unilateral production decision that rendered the H200 approval moot in practice. In March 2026, NVIDIA halted China-bound H200 production and redirected TSMC N3-node wafer capacity to Vera Rubin, its next-generation chip (H2 2026 commercial rollout).

The business logic is straightforward. H200 and Vera Rubin compete for the same TSMC N3 capacity. Vera Rubin is blocked from China under the same BIS thresholds that cleared H200; Vera Rubin's specs far exceed both the TPP (< 21,000) and bandwidth (< 6,500 GB/s) ceilings. Redirecting to Vera Rubin captures the US hyperscaler market — OpenAI, Google, Microsoft, Meta, all of which have co-designed infrastructure around Vera Rubin — rather than waiting for Beijing to complete a deliberation with uncertain outcome.

The result: NVIDIA's usable inventory was approximately 700,000 H200 chips at the time of the policy announcement. NVIDIA planned an initial 40,000–82,000 chip first tranche from that stock. The Lunar New Year delivery target was missed because China paused purchases. Production was then halted and repurposed. The ceiling on what China could receive under this rule is now, practically speaking, the 700,000-chip inventory — and shipments from that inventory never began.

Jensen Huang's public earnings call statement in May 2026 was direct: "Today in China, we have now dropped to zero." CFO Colette Kress excluded China from forward guidance entirely. NVIDIA's response to lost China revenue is a $20 billion Vera revenue plan from new markets — not a China recovery thesis.

The chips that actually crossed the border.

While zero H200s moved under the new approval regime, a different category of chips appears to have flowed to China throughout this period.

NVIDIA Blackwell-class chips — B100/B200/B300, which exceed BIS technical thresholds and remain on presumption-of-denial — were acquired by Chinese firms through overseas subsidiaries registered in Malaysia, Singapore, and the UAE. These entities are formally Chinese-owned but are not Chinese entities under EAR definitions; US chip suppliers sold to them legally.

On May 31, 2026, BIS issued guidance specifically closing this loophole. Industry sources cited in reporting estimated "hundreds of thousands" of advanced chips had flowed through the oversight gap. The Trump administration disputed whether the loophole was intentional. The guidance's existence — and its timing — confirms the scale was material enough to warrant formal action.

The practical implication: if hundreds of thousands of Blackwell chips entered Chinese AI infrastructure through the loophole, the approved-but-unshipped H200s are less significant than they appear. China may have acquired, through back channels, precisely the next-generation compute the official policy blocked through front channels. The H200 approval is a story about 700,000 chips that never moved. The subsidiary loophole is a story about hundreds of thousands of more powerful chips that did.

The enforcement problem the rule always had.

The BIS rule's nine certification requirements were analyzed within days of publication by CFR, CNAS, and the Institute for AI Policy and Strategy. The consensus: unenforceable.

The problems are structural, not operational:

Certification requirement #2 mandates that exporters certify that China shipments will not delay US customer orders — "during a period of severe global AI chip shortage," as CFR noted. NVIDIA operates in an environment where US hyperscaler demand exceeds supply for every advanced chip category. Certifying no delay is commercially absurd to do honestly, and BIS has no independent verification mechanism beyond the exporter's own attestation.

Certification requirements #5 and #6 prohibit military end-use and require KYC from Chinese buyers. Alibaba, Tencent, and ByteDance — the primary approved buyers — have documented relationships with Chinese military and security services. The CFR analysis was specific: "certifying no military use for chips sold to these entities is contradictory to publicly available facts." The military-civil fusion structure of China's tech sector makes buyer-level certification of end-use a theoretical exercise.

Independent third-party testing infrastructure — requirement #9 — does not exist at scale. No entity capable of certifying H200 exports at volume, independent of buyer and seller, has been identified or stood up. As of June 2026, this certification infrastructure is notional.

BIS, characterized in multiple expert analyses as "operating on Google searches and Microsoft Excel," lacks the investigative capacity to verify complex supply chain compliance for 750,000 chips across 10 Chinese buyers. No enforcement actions have been taken under the new rule — largely because zero chips have shipped under it, so there is nothing to enforce against.

AMD MI325X: eligible, irrelevant.

AMD's MI325X was included in the BIS rule because it falls within both technical thresholds (TPP ~20,800 < 21,000 ceiling; bandwidth 6.0 TB/s < 6,500 GB/s ceiling) — just clearing the compliance bar on bandwidth.

In hardware specifications, MI325X is formidable: 256GB HBM3E memory versus H200's 141GB, and 1.8x the memory capacity at higher raw compute. It should, on paper, be attractive for AI inference.

In practice, it is not. China's AI infrastructure runs on CUDA-native codebases — years of development optimized for NVIDIA's software ecosystem. AMD ROCm's practical deployment for LLM inference at Chinese scale carries substantial migration friction. MI325X's inclusion in the rule created the appearance of a dual-vendor policy; it did not create a realistic alternative adoption path for Chinese AI labs on the timescale the rule contemplates.

The market share outcome the policy produced.

The combined effect of the export control saga — H20 cut-off in April 2025, H200 reversal with no deliveries in January 2026, Blackwell blocked throughout, Blackwell loophole active until May 2026 — is a Chinese AI chip market that looks nothing like it did 24 months ago.

NVIDIA held approximately 95% of China's AI chip market in 2023. As of mid-2026, NVIDIA is in single digits — primarily legacy H20 residual and non-data-center applications. Huawei Ascend has risen from near-zero to approximately 55% of China's domestic AI chip market. The remaining share is distributed across domestic alternatives and legacy NVIDIA products.

Huawei Ascend 910C delivers approximately 60% of NVIDIA H100 inference performance — roughly 42–46% of H200 equivalence, with acknowledged deficiencies in chip-to-chip interconnect and software ecosystem maturity. For full-scale frontier model training from scratch, Chinese AI labs describe Ascend as inadequate. For inference and fine-tuning workloads, it is now their primary infrastructure.

NVIDIA estimates Huawei will not produce a chip more powerful than H200 until approximately Q4 2027 (Ascend 960). The US is assessed as maintaining roughly a three-year lead at the chip frontier. But the lead is measured against what China is trying to build — and China's stated goal has shifted from "acquire NVIDIA equivalents" to "eliminate NVIDIA dependency entirely."

The export control saga, measured by its stated objective of preventing advanced AI compute from reaching China, produced mixed results: it blocked the front door while hundreds of thousands of Blackwell chips entered through the side. Measured by its unstated effect — accelerating Chinese domestic chip development and pushing Chinese labs off NVIDIA infrastructure permanently — it has been remarkably successful, in a direction no one explicitly intended.

The January 2026 reversal changed nothing on the ground. It clarified, in the most concrete possible terms, that the chip war's primary casualty is not Chinese AI development. It is NVIDIA's China revenue.