China's AI Silicon Play: How Export Controls Built a Domestic Chip Market

01 The export-control boomerang

The first round of US semiconductor export controls landed in October 2022, cutting Chinese buyers off from the most advanced GPUs then on the market. Washington assumed the result would be a slower Chinese AI sector. Instead, each successive round has functioned less like a brake and more like an industrial policy for Beijing — written by an outside patron, with Chinese industry as the designated beneficiary. October 2023 tightened the rules around performance-density metrics, closing the loophole that had let Nvidia ship A800- and H800-class parts into China. December 2024 pulled high-bandwidth memory inside the perimeter. April 2025 imposed a license regime on the H20, Nvidia's purpose-built China part, freezing the last high-volume channel overnight.

Through 2025 and 2026, procurement pressure did the rest: state guidance steered cloud buildouts, banks, and state-owned enterprises toward domestic stacks regardless of unit economics. The pattern is consistent. Every escalation converted a slice of Chinese industry from customer to competitor. Firms that had budgeted for US accelerators redirected capital toward domestic alternatives, and the state supplied the demand that export controls suddenly made scarce. The boomerang now returns as competition.

02 What the Ascend-class hardware actually does

The flagship domestic alternative is Huawei's Ascend line, designed by HiSilicon — the fabless subsidiary that also builds Kirin smartphone processors — and fabricated domestically by SMIC on 7nm-class processes. The 910C-class parts pair two dies and aim squarely at the inference workloads that dominate deployed AI services: lower-precision math, large memory capacity, and high inter-chip bandwidth. For serving chat, search, and recommendation traffic, that is enough. Chinese cloud operators have been wiring Ascend clusters into production since 2025, and the installed base keeps growing.

The remaining gaps are equally specific. High-bandwidth memory is the sharpest: domestic HBM output is small, and interim packaging workarounds add cost and latency. Extreme-ultraviolet lithography is unavailable, so SMIC multi-patterns DUV tools, which inflates process steps and depresses yield at advanced nodes. EDA software, metrology, and inspection toolchains carry partial foreign dependencies that the controls have narrowed but not closed. The result is silicon credible for inference today and still a generation behind for frontier-scale training.

03 The substitution math

The substitution math starts with demand: China is estimated to account for roughly a fifth to a quarter of global AI accelerator demand, and domestic supply covers a minority of it. Industry estimates put domestically produced ICs at about 16 percent of Chinese IC consumption in 2020, 18 percent in 2022, 21 percent in 2024, and an estimated 24 percent in 2026 — directionally steady, roughly a point and a half of import substitution per year. Applied to AI silicon specifically, the trajectory is sharper than the all-chip average suggests.

Nvidia's China data-center revenue share has contracted from around a fifth of its total a few years ago toward a low single-digit slice as license requirements and procurement directives bite. Procurement is the lever that makes substitution sticky. Central and provincial guidance favors domestic chips for state-linked cloud buildouts, and telecoms, banks, and state-owned enterprises have shifted tenders toward Ascend-class hardware even where imported parts remain faster per dollar. Demand, once redirected by policy, tends to stay redirected.

04 Manufacturing reality check

Yield is where the domestic story meets physics. Multi-patterning on DUV lithography multiplies process steps, and every additional step multiplies the probability of a killing defect. SMIC's 7nm-class lines are widely believed to run at yields well below what TSMC achieves on comparable designs, which raises effective per-die cost before packaging even begins. Metrology and inspection — much of it historically sourced from US-aligned suppliers — form a second dependency the controls have only partially substituted.

The economics still close because the target product differs. Inference silicon tolerates older nodes, lower clock speeds, and chiplet workarounds; what it needs is memory bandwidth, interconnect, and volume manufacturing. That is what 910C-class parts optimize. Training frontier models is less forgiving: it punishes chip-level inefficiency at cluster scale, and there the domestic stack still leans on stockpiled high-end parts, dense clusters of weaker chips, and algorithms that trade compute for wall-clock time. Good enough for serving is not yet good enough for the frontier.

05 Market mechanics

Hardware is the smaller half of substitution. Nvidia's moat is CUDA: two decades of libraries, tuned kernels, and developer muscle memory. Huawei's answer, CANN, has narrowed the gap but still demands porting effort, and every framework release must be validated against it. Switching costs are real, which is why Beijing pushes framework maintainers and model toolchains to treat domestic accelerators as first-class targets rather than ports. Developer time is the hidden subsidy every ecosystem transfer requires, and state coordination lowers it.

Those frictions produced a strange market structure. Gray-market H100s circulate at steep premiums; H20s sold briskly until the license regime froze the channel; and domestic chips, shielded from direct comparison, enjoy a price umbrella — often neither cheaper nor faster per unit of work, but available, sanction-proof, and procurement-approved. In a market where the state sets the buying criteria, that combination is enough to move serious volume.

06 Limits and open questions

The open questions are about scale and stamina. Whether domestic capacity grows fast enough to meet training-scale demand turns on HBM supply, advanced-node wafer output, and cluster networking — each improving, none solved. Financing is the second question: state-guided lending can carry capital-hungry fabs and design houses for years, but intensity on this scale eventually demands real returns. Power is the third; data-center electricity demand is compounding faster than grid additions in several provinces.

The trajectory also depends on the controls themselves. A loosened US regime would tempt some Chinese buyers back to US silicon and compress the domestic price umbrella. Tighter controls — or Chinese counter-restrictions on materials, tools, and critical inputs — would accelerate substitution by decree. Policy, not engineering, remains the most volatile variable in the forecast.

07 What it means for the global AI buildout

The likeliest 2027 picture is a bifurcated accelerator market: a US-centered ecosystem around Nvidia and its suppliers, and a Chinese ecosystem around Ascend-class hardware and its own toolchain, overlapping only at the gray-market edges. Pricing effects reach both sides — China-specific parts were designed down to a license envelope, and losing that volume removes a subsidy from the global cost curve while handing domestic vendors guaranteed scale at home.

Watch three indicators: SMIC advanced-node wafer output and yield, domestic HBM qualification at volume, and whether CANN-based toolchains reach parity on the frameworks that matter. Whichever moves first will tell you whether China's silicon play is a hedge that matured into an industry — or a market that simply closed.