Research Brief

AI Chip Export Controls

Are US semiconductor export restrictions actually slowing China's AI progress?
Brandon Huey · January 2026

Rather than slowing China's AI model development, US export controls have accelerated a $50B domestic chip ecosystem that did not exist three years ago.

NVIDIA's China market share collapsed from 95% to 0% between 2022 and 2025.[1] In that same window, Huawei shipped the Ascend 910C at 60% of H100 inference performance,[2] SMIC scaled 7nm capacity to 45,000 wafers per month,[3] and DeepSeek trained R1, a frontier reasoning model, on just 2.8M H800 GPU-hours.[4] The controls succeeded at preventing China from deploying AI at hyperscaler scale, but failed at what policymakers said mattered more: stopping China from building the capability to do it domestically.

"Export control was a failure. We went from 95% market share to zero. If they don't have enough Nvidia, they will use their own."
Jensen Huang, CEO of NVIDIA, Computex Taipei, May 2025

From Containment to Monetization

The original rationale for semiconductor export controls was national security: deny China access to the compute needed to train frontier AI systems and develop advanced weapons. The Bureau of Industry and Security banned A100 and H100 exports in October 2022, then closed the H800/A800 workaround loophole a year later. In January 2025, the Biden administration issued the AI Diffusion Rule, a three-tier country framework restricting chip flows globally.

That framework lasted four months. The Trump administration rescinded it in May 2025, calling it too burdensome on American companies.[5] What replaced containment was something new: revenue extraction. In August 2025, NVIDIA and AMD agreed to give the US government 15% of H20 chip sales to China. By December, the rate jumped to 25% for H200 sales.[6] Washington has shifted from trying to stop the flow of chips to simply taxing it.

Meanwhile, enforcement tells a parallel story. DOJ's Operation Gatekeeper dismantled a $160M smuggling network in December 2025, indicting four defendants who used straw purchasers and domestic warehouses to move restricted H100 and H200 GPUs out of the country.[7] CNAS estimates between 10,000 and several hundred thousand restricted chips reached China in 2024, with a median estimate of approximately 140,000.[8]

Timeline of Controls vs. Chinese Countermoves
Date US Action Chinese Response Net Effect Conf.
Oct 2022 BIS bans A100/H100 exports to China Huawei accelerates Ascend 910B development China domestic GPU investment surges 3x YoY High
Oct 2023 BIS closes H800/A800 loophole SMIC ramps 7nm N+2 process for Ascend 910C Huawei begins sampling 910C to Baidu, ByteDance High
Jan 2025 Biden AI Diffusion Rule: 3-tier country system DeepSeek publishes R1 trained on 2.8M H800 hours Proves frontier AI trainable on restricted hardware High
Apr 2025 H20 exports require license; NVIDIA takes $4.5B charge[13] Biren, Moore Threads, Enflame IPOs raise $8B+ combined Capital flood into domestic GPU startups High
May 2025 Trump rescinds AI Diffusion Rule SMIC hits 45,000 wspm on 7nm; plans to double in 2026 Policy whiplash undermines allied coordination Med
Aug 2025 15% revenue-sharing deal for H20 sales China discourages local firms from buying NVIDIA Controls become de facto export tax High
Dec 2025 25% revenue-sharing deal for H200 sales Huawei targets 200,000 Ascend chips for 2025 US monetizes access; China builds parallel supply chain Med
Feb 2026 BIS reviews posture; case-by-case for H200/MI325X China targets 100K wspm at 7nm/5nm within 2 years 5x capacity expansion plan announced Med

The Domestic Response

What Jensen Huang calls a "national mobilization" is visible in the numbers. China's semiconductor investment has produced a cohort of GPU startups, dubbed the "Four Little Dragons," that collectively raised over $8B in 2025 and 2026 IPOs. These are not paper companies. Moore Threads demonstrated 1,000 tokens per second decoding DeepSeek V3 on its upcoming S5000 GPU.[9] MetaX's Xiyun C600 delivers performance between the A100 and H100, with mass production slated for the first half of 2026. Enflame's L600 is rumored to be the first Chinese GPU with HBM3 memory.

At the fabrication layer, SMIC's 7nm N+2 process, built without EUV lithography using DUV multi-patterning, is the backbone of this ecosystem. Capacity stood at roughly 45,000 wafer starts per month at the end of 2025. SMIC plans to double that in 2026, and the broader Chinese government target is to reach 100,000 wspm at 7nm and 5nm within two years.[10] That is a 5x increase from current levels.

The flagship product is Huawei's Ascend 910C. Manufactured on SMIC's N+2 process, it delivers 2.4 petaFLOPS of FP16 compute with 128GB of HBM3 memory, which is about 60% of the H100's inference throughput according to benchmarks published by DeepSeek researchers.[2] While it cannot compete at training scale, where NVIDIA maintains an undisputed lead, the inference gap is narrower than the export control regime assumed.

Who Is Building China's GPU Stack?
Huawei Ascend 910C: 2.4 PFLOPS FP16, 128GB HBM3. 200,000 chips targeted for 2025. Powers DeepSeek R1 inference.
SMIC 7nm N+2 process. 45,000 wspm in 2025, doubling to ~90,000 in 2026. Net profit up 35.6% in H1 2025.
Moore Threads S5000 GPU: 1,000 tok/s on DeepSeek V3. STAR Market IPO late 2025. Huagang arch targets H100 parity.
Biren Tech Raised HK$5.58B in Hong Kong IPO (Jan 2026).[11] BR200 successor in development on 7nm.
Enflame STAR Market listing accepted Jan 2026. L600 chip rumored first Chinese GPU with HBM3.
MetaX Xiyun C600: A100-to-H100 class performance. Mass production slated for H1 2026.
Factors Accelerating China's Chip Independence

The Comparison

The clearest way to evaluate the export control regime is to compare its intended effects against China's actual domestic response across every relevant dimension, from chip design and fabrication to capital formation and policy consistency.

US Containment Strategy vs. China's Domestic Response
Dimension US Export Controls China Domestic Ecosystem
Leading chip NVIDIA H200 (restricted); B200/Rubin withheld entirely Huawei Ascend 910C: 60% of H100 inference, 128GB HBM3
Fabrication EUV equipment blocked via ASML/Dutch agreement SMIC 7nm DUV multi-patterning; 45K wspm, targeting 100K
Frontier models Controls intended to prevent training at scale DeepSeek R1/V3 trained on 2.8M H800 hours; GPT-4 class
Deploy scale Limits on inference chip volume (H20 licensing, then ban) 1-4% of US production in 2025; 200K Ascend chips targeted
Capital NVIDIA lost $4.5B on H20 inventory write-down[13] "Four Little Dragons" raised $8B+ in 2025/2026 IPOs
Talent Visa restrictions on Chinese STEM students debated 50% of world's AI researchers are Chinese nationals[12]
Enforcement Operation Gatekeeper: $160M network dismantled Est. 140,000 restricted chips smuggled in 2024
Policy stability AI Diffusion Rule issued Jan 2025, rescinded May 2025 15th Five-Year Plan: semiconductors designated "strategic"
Revenue model 15% (H20) to 25% (H200) export tax on chip sales Beijing discourages local firms from buying NVIDIA
Trade-Off Short-term bottleneck for China; long-term loss of US market access and allied credibility 2-3 year fab lag; accelerating closure via massive state investment

Assessment

The export control regime aimed to achieve three things: constrain China's ability to train frontier AI models, limit deployment at scale, and prevent the emergence of a self-sufficient domestic chip supply chain. It has partially succeeded at one of these objectives.

YES: Deployment Scale

Controls constrain China's ability to deploy AI at hyperscaler scale — Huawei's 200,000-chip target for 2025 is roughly 1-4% of US production, and inference at the volume needed for consumer products remains gated by fabrication capacity rather than model quality.

NO: Model Development

DeepSeek proved that algorithmic efficiency can substitute for raw compute: R1 was trained on 2.8M H800 hours, a fraction of what US labs spend, yet matches GPT-4 class performance — demonstrating that controls did not prevent frontier AI capability development.

NO: Ecosystem Independence

Controls triggered what Huang calls a "national mobilization": four domestic GPU startups raised $8B+ in IPOs and SMIC plans to 5x leading-edge output by 2028, meaning the controls ultimately created the economic conditions for the very supply chain they aimed to prevent.

Conclusions

1
Export controls bought time without lasting advantage. The 2-3 year fabrication lag is real but shrinking. SMIC's plan to reach 100,000 wspm at 7nm/5nm within two years, combined with China's 5x capacity expansion target, suggests the manufacturing gap closes by 2028 or 2029. The window of US compute superiority is narrower than policymakers assume.
2
Algorithmic efficiency is the great equalizer. DeepSeek's R1, trained at a fraction of US compute budgets, proved that model quality is increasingly decoupled from raw chip access.[4] Training efficiency gains of 2-4x per year erode hardware advantages faster than export rules can be updated.
3
The policy has pivoted from security to revenue extraction. The progression from export ban to 15% tax on H20 to 25% tax on H200[6] reveals a government that has abandoned containment in favor of monetization. This undercuts the national security rationale and signals to allies that chip controls are negotiable instruments, not strategic commitments.
4
The biggest loser is NVIDIA, not China. NVIDIA went from 95% China market share to 0%,[1] took a $4.5B inventory write-down,[13] and now pays 15-25% of any resumed China revenue to the US government. Meanwhile, China's "Four Little Dragons" raised $8B+ in IPOs to build the replacement ecosystem. The controls redistributed market share from a US champion to Chinese competitors.
Key Terminology
BIS — Bureau of Industry and Security; Commerce Dept. agency administering export controls
EAR — Export Administration Regulations; legal framework governing controlled exports
wspm — Wafer starts per month; standard measure of fab production capacity
DUV — Deep ultraviolet lithography; older process SMIC uses to approximate EUV results
HBM3 — High Bandwidth Memory 3rd gen; stacked DRAM used in AI accelerators
PFLOPS — PetaFLOPS; 10^15 floating-point operations per second
N+2 — SMIC's 2nd-gen 7nm-class process node; used for Ascend 910C production
AI Diffusion Rule — Biden-era 3-tier export framework; issued Jan 2025, rescinded May 2025
MoE — Mixture of Experts; architecture enabling DeepSeek V3's compute efficiency
Four Little Dragons — Biren, Enflame, MetaX, Moore Threads; China's leading GPU startups
Sources: Bureau of Industry and Security, US Department of Justice (Operation Gatekeeper), Congressional Research Service (R48642), NVIDIA SEC filings, TrendForce, SemiAnalysis, CNAS, Tom's Hardware, CNBC, CSIS, DeepSeek Technical Report