Start with the numbers. Morgan Stanley's estimates show China's AI chip self-sufficiency rate climbing from roughly 10 percent in 2021 to 41 percent in 2026 — a more than fourfold increase in five years — with about 86 percent projected by 2030. Bernstein's forecast is even more specific: Nvidia's share of China's AI semiconductor market will plunge to 8 percent by 2026, while Chinese firms' combined share will reach four-fifths.
Five Numbers and One Story
But behind these impressive figures runs a less impressive undercurrent: it is hard to find a complete plan, openly articulated at the policy level, that engineered this transition. The more common narrative is the sweeping verdict that "sanctions forced self-reliance." That is true — but it is not enough.
On July 14, 2026, Jeffrey Kessler, an official at the U.S. Department of Commerce's Bureau of Industry and Security (BIS), disclosed an unassuming but powerful fact at a House Foreign Affairs Committee hearing: after receiving U.S. government approval, only a "very small quantity" of Nvidia H200 AI chips had been shipped to Chinese customers — in his own words, "Very small quantity of chips, so it's trivial."
That testimony pulls the tension of the whole affair into focus. The United States did, last December, clear Nvidia's H200 for sale to China for the first time — a notable easing of AI chip export controls, formally institutionalized in January, albeit with a 25 percent U.S. levy attached. Yet in April, U.S. Secretary of Commerce Howard Lutnick revealed under congressional questioning that China has not purchased a single H200 chip, because Beijing wants to center its investment on the autonomous development of its domestic industry.
On one side: "I'm done selling." On the other: "Even with the door open, I'm not buying." It is the superposition of the two moves that forms this story's true starting point.
The sale of Nvidia's H200 chips to China is one of the signature controversies of the US-China tech rivalry. Since late 2022, the U.S. government has progressively escalated AI chip export controls on China — from the A100 to the H100 to the H200 — with exports of each generation subject to renewed review. In December 2025, Trump approved the export of the H200 to China, widely read at the time as an "easing." Kessler's testimony now shows that this easing produced almost no actual shipments. Lutnick's supplementary remark goes further in revealing why: the power of decision does not rest entirely on the American side.
From 10% to 86% — The Machinery Behind the Numbers
The climb in China's AI chip self-sufficiency is not a smooth curve. Its rhythm is the superposition of four independent forces.
First, the supply-side effect of sanctions. U.S. export controls on AI chips to China were ratcheted up layer by layer between 2022 and 2025, and each escalation forced downstream Chinese customers to reassess supply-chain security. When the H800 — Nvidia's China-specific chip — was banned in 2024, substitution demand erupted in concentrated fashion. The role sanctions played here was not "blockade" but "accelerator": they turned a substitution process that would otherwise have advanced steadily, driven by cost and technological inertia, into an urgent reprioritization. The difference lies in the timetable, not the direction.
Second, the substantive mass production of domestic chips. China's homegrown AI chipmakers (Huawei's Ascend series, Hygon's DCUs, and others) have seen verified gains in both the scale and stability of output over the past 24 months. The Huawei Ascend 910B now delivers roughly 60–80 percent of the H100's performance on some inference workloads, with a stable supply channel. That gives downstream cloud providers and large enterprises a realistic option: swapping a fully unavailable high-end chip for a domestic one whose performance is uncertain.
Third, the strategic divergence of downstream customers. The leading cloud providers (Alibaba Cloud, Tencent Cloud, Baidu Cloud) are not pursuing "full domestication" but "hybrid chip pools" — splitting training and inference workloads across different chips. High-precision training still leans on stockpiled Nvidia inventory and a small number of H200s, while inference and fine-tuning tasks have migrated to domestic chips at scale. This strategy cuts the trial-and-error cost of migration and lets domestic chips iterate faster under real workloads.
Fourth, the ground-level push of policy. From "new quality productive forces" to the Big Fund for the chip industry, the public capital and guided investment that the Chinese government has poured into AI chips over the past five years add up to a figure large enough to deserve its own line item — but that is not all. What matters more is policy certainty: firms know export controls will not be lifted in the medium term, and are therefore willing to commit to long-horizon R&D. A clear negative expectation can, in some cases, drive a transition more powerfully than positive industrial incentives.
Nvidia at 8% — A Gradualist Endgame of "Decoupling"
Bernstein's projection of Nvidia's China share "falling to 8 percent" would, if realized, be a structural endgame signal. Nvidia still holds more than 80 percent of China's AI chip market (though below its peak). Contraction from 80 percent to 8 would demote Nvidia from the dominant player in the Chinese market to a marginal supplier.
But this is not a simple "decoupling" story. In fact, several structural factors are shaping the process into something more like "gradual stratification" — rather than a sudden cutoff:
- The installed-base effect: Chinese cloud providers have accumulated vast inventories of Nvidia GPUs over the past several years (including A100s and H100s hoarded before the bans), and these chips still run. Replacement of an installed base is gradual, not instantaneous.
- Hybrid architectures: Most major players' strategy is to mix domestic and imported chips, not to replace them across the board. That means even as its share falls, Nvidia retains a presence in segments such as high-end training.
- The stickiness of the ecosystem: The CUDA ecosystem remains Nvidia's moat. Domestic chipmakers (Huawei's CANN ecosystem for Ascend, for instance) are catching up, but ecosystem migration is a process measured in years.
These three factors mean Nvidia's "8 percent" will not arrive as a sudden drop but as a slow compression across successive annual iterations. That hands China's domestic chip industry a rare "soft landing" window — not a plunge into vacuum, but a gradual expansion of share with installed-base support, hybrid transition, and parallel ecosystems underneath.
The Compounding Effect of the "Chokepoint"
The most noteworthy narrative thread in this whole affair is the compounding effect of being "choked by the neck" (卡脖子). It took China's chip self-sufficiency five years to move from 10 percent to 41 percent — faster than most independent analyses expected. And Nvidia's share is projected to collapse from 80 percent to 8 percent — steeper than any analyst in 2022 would have dared to forecast.
Laid side by side, the two speeds show that the marginal effect of sanctions is decaying — and decaying at an accelerating rate: the stricter the export controls, the greater Chinese industry's adaptation to the sanctions pathway; the more concentrated the substitution investment spawned by being "choked," the smaller the incremental damage the next round of sanctions can inflict. This is the curve of "diminishing marginal returns to sanctions" — nothing new in economics, but in the specific domain of AI chips, it has now been corroborated by real industrial data.
The original intent of sanctions is to slow the other side's technological catch-up. But if the intensity of sanctions lands precisely in the zone that is "enough to trigger large-scale substitution investment, yet not enough to destroy the industrial base," they instead become the legitimation and the capital-mobilization engine for the other side's indigenous R&D. That is the compounding effect of the chokepoint: every attempt to tighten the grip accelerates the arrival of the next moment when you are "no longer needed."
Two Uncertainties — Time or Direction
Of course, the road from 41 percent to 86 percent is harder than the one from 10 to 41. Three reasons:
First, the bottleneck in advanced process nodes. China's capacity in leading-edge nodes (below 7nm) is constrained by the EUV lithography embargo; even where design capability exists, competitive physical capacity is limited. SMIC's N+2 process has reached mass production, but still lags TSMC's 5nm in yield and performance. AI chips are especially dependent on advanced nodes — compute density and power efficiency are tightly coupled to the process node.
Second, the ecosystem gap. The CANN-versus-CUDA divide is not a technological one; it is a matter of user habit and toolchain maturity. "The war over AI chips is fought off the chip" — the completeness of software development tools, operator libraries, debugging utilities, and deployment frameworks determines whether a chip gets widely adopted. On this dimension, domestic chips still trail Nvidia by a full generation.
Third, the asymmetric game of geopolitical risk. The rise in China's AI chip self-sufficiency will itself trigger harsher U.S. export-control countermeasures, producing spiral escalation. Kessler's "trivial" characterization of H200 shipments at the congressional hearing is itself a political signal: the executive branch may face tougher congressional scrutiny, further squeezing Chinese chips' access to overseas components.
1. When China's AI chip self-sufficiency reaches 86 percent, what will the remaining 14 percent — the high-end chips still dependent on imports — turn out to be? Nvidia's next-generation flagships, or specific chip types on specialty processes? That 14 percent is precisely the critical band that decides whether "self-reliance" equals "good enough."
2. If domestic AI chips reach 86 percent self-sufficiency in the home market, can they be exported outward? In world markets, what Chinese AI chips face is not the sanctions problem — it is the problem of standards, ecosystems, and trust. An impressive domestic self-sufficiency figure does not automatically translate into competitiveness.
Conclusion
Return to the opening line: the climb from 10 percent to 41 percent self-sufficiency was driven by three forces released in sequence — sanctions, "not buying," and indigenous R&D. On the road from 41 to 86 percent, all three will keep operating, but the diminishing marginal returns of export controls will grow ever more visible, the technical climb of domestic chips will approach its limits, and external geopolitical pressure will not ease.
What most deserves a place in the Wiki about this story is not the numbers themselves but the logic they reveal: when an economy under external pressure forms a "directed investment preference" for technological substitution, export controls themselves become the speed of the forcing mechanism — provided that economy has enough market scale and capital depth to sustain the transition.