Where Are They Now
Covering Chinese policy and rhetoric on external events and actors, military and security issues, economy and technology, and bilateral relations with India.
From Chips to LLMs: Have the Export Controls Prevented China from Catching Up?
Amit Kumar and Satya S. Sahu
As geopolitical contestation and the desire to dominate new and emerging technologies intensify, export controls have acquired a greater salience. The latest in the series is the US directive to suspend access to foreign entities and individuals around the world from accessing the most “advanced and powerful” frontier LLM model overnight. The tussle is not being fought more fiercely anywhere than in the domain of AI.
Earlier on May 21, 2026, Taiwan’s Keelung District Prosecutors’ Office executed search warrants across 12 locations on the island, detaining three individuals for allegedly using forged documents to smuggle Super Micro servers – equipped with restricted NVIDIA chips – to China, Hong Kong, and Macau. The raids followed a US Department of Justice indictment in March that charged Super Micro’s co-founder, Wally Liaw, and two associates with conspiring to divert approximately US$ 2.5 billion worth of NVIDIA-powered servers to China through a front company.
The Super Micro case is not an isolated incident. Over the past 12 months, the US Bureau of Industry and Security has announced nearly US$ 420 million in combined penalties related to the illegal smuggling of semiconductor technology to China.
Furthermore, one must look at these developments against the DeepSeek’s V4 preview release, a 1.6-trillion-parameter AI model. Last year, DeepSeek’s R1 release sent the US AI stocks tumbling, for achieving comparable performance at a fraction of the cost and compute power which was previously thought necessary. However, the R1 model had primarily been trained on NVIDIA H800 GPUs stockpiled before export restrictions tightened.
Contrary to R1, however, V4 has been built to run inference workloads on Chinese homegrown chips, such as the Huawei Ascend 950 and Cambricon silicon. It marked the first instance of a frontier Chinese model being co-optimised with domestic inference hardware from the outset, with DeepSeek and Huawei’s teams working together prior to launch. The training stack remains a more contested picture. Multiple sources suggest a mix of Nvidia H800S and Huawei Ascend 910C chips was used. Nevertheless, that the full inference pipeline has shifted to domestic silicon, with eight Chinese chipmakers completing same-day adaptation, is in itself, a significant inflection.
Together, these two developments raise an uncomfortable question: how effective were the US export controls in constraining China’s AI progress? The fact that the US approved the sale of NVIDIA’s H200 chips in December last year only adds fuel to the fire.
There’s no simple answer. In fact, depending upon how one defines the intended objectives and metrics of success, the answer could vary.
The answer can vary
On the frontier hardware front, the US position is stronger. NVIDIA’s Blackwell architecture delivers roughly three times the processing power of Huawei’s best chip, the Ascend 950, which recently surpassed NVIDIA’s H20 chips in inference performance.
Furthermore, China’s dependence on American hardware is eroding faster than most anticipated three years ago. Its AI chip self-sufficiency rose from below 10 per cent in 2020 to 41 per cent by 2025, while NVIDIA’s China market share has fallen from over 95 per cent to under 60 per cent. Even more striking is China’s resolve to further reduce this dependency. The US Commerce Secretary Howard Lutnick confirmed on April 22 that NVIDIA had not sold a single H200 to China, because “The Chinese central government has not let them buy the chips,” referring to Beijing’s January 2026 directive that instructed agents to block NVIDIA H200 imports, and government officials told domestic companies not to purchase the chips unless necessary. Thus, the outcomes point to mixed results for the US.
On the frontier model level, however, the export controls haven’t succeeded in creating a significant gap, prima facie. While DeepSeek acknowledged that it still trails closed-source leaders on broad world-knowledge tasks, on competitive coding benchmarks, V4 outperformed GPT-5.4 while trailing other leading models by a thin margin on software engineering tasks. The Stanford report estimated that the performance gap between the best US and Chinese AI models had closed to 2.7 per cent by March 2026, down from over 30 percentage points in May 2023. Moreover, the fact that the V4 model represents a documented transition from American hardware to a fully domestic Chinese stack in just over a year is, in itself, a significant achievement.
Denied easy access to NVIDIA GPUs, the Chinese chose to optimise for efficiency rather than wait for hardware parity. One of the ways included leveraging a mixture-of-experts architecture – wherein only certain parameters are activated depending upon the type of query to compensate for weaker chips. Thus, while the gap at the chip level grew, the gap at the model level shrank – suggesting that hardware advantage isn’t translating into superior model capability.
Diffusion continues to be a challenge
However, DeepSeek’s nationwide adoption, diffusion, and deployment remain constrained by three interconnected factors: mass availability of chips, cluster assembling, and interconnect bandwidth.
To begin with, China is still struggling to mass-produce the homegrown chips. SMIC has demonstrated 7nm production using older DUV lithography through multi-patterning, but its 5nm process – needed for next-generation AI accelerators – operates at approximately 20 per cent yield without access to EUV lithography tools, far below commercial viability. In practice, this means that even when China can design chips capable of running frontier models, it cannot yet produce them at scale for an economy-wide rollout.
The challenge of ‘cluster efficiency’ represents the most significant physical barrier for the deployment of frontier models like V4. In the modern AI landscape, intelligence is a product of scale. The more parameters, the higher the intelligence. However, as models grow to encompass 1.6 trillion parameters – which DeepSeek’s V4 does – inference workloads for them must be fragmented across thousands of chips, thereby creating a massive distributed system, called a cluster. Under such conditions, the brute speed of the individual chip isn’t the primary constraint. It’s the delay incurred – referred to as communication tax – when thousands of chips attempt to coordinate their calculations in real-time. In a large model, chips spend up to 50% of their time on communication, thereby slowing down the model’s overall performance.
This is where the third structural challenge emerges. To overcome this barrier and improve cluster efficiency, models like DeepSeek’s V4 use a mixture-of-experts architecture. But it demands constant high-speed data exchange. American frontier models rely on NVIDIA’s NVLink – a high-throughput interconnect – to manage this and reduce latency. Chinese clusters, constrained by export controls, operate on significantly narrower bandwidth, and so adding more domestic hardware does not necessarily translate into proportional gains in usable performance.
In this light, the relaxation granted for H200 chips offers little respite to China. After the easing of restrictions last year, each of the 10 approved Chinese companies can purchase up to 75,000 H200 chips. Statistically, it remains a marginal concession when measured against the sheer scale of the US’s hyperscale investment.
To put it in perspective, Meta housed over 1.3 million GPUs by the end of 2025, and plans to deploy millions more. xAI’s Colossus 2 facility alone houses approximately 555,000 Blackwell GPUs, and has plans to further expand to 1 million chips. The asymmetry reinforces the reality that while China may maintain ‘frontier parity’ through algorithmic ingenuity, the US will likely retain a commanding lead in total aggregate compute capacity.
However, the smuggling challenge operates on a different plane from the sanctioned H200 sales. The chips being diverted through networks like Super Micro’s are often frontier-grade hardware. While their volumes may not match the scale of US hyperscale deployments, even modest quantities of smuggled frontier chips can meaningfully aid China’s cutting-edge research.
Amit Kumar is a Staff Research Analyst with the Takshashila Institution, Bengaluru. Satya S. Sahu is an Associate with The Asia Group’s South Asia practice based in New Delhi.
Eye on China is a weekly newsletter curated by the Indo-Pacific Studies Programme at The Takshashila Institution, a public policy think-tank based out of Bengaluru, India.
 | A guest post by
|
A follow-on thought after reading Tyler Cowen's recent post on the restrictions placed on Anthropic's latest models.
This piece focuses primarily on the hardware side of the equation. But what if the next bottleneck is not hardware at all, but deployment?
One of the most interesting conclusions here is that export controls succeeded in widening the chip gap, yet failed to create a comparable gap in model performance. Chinese firms responded by optimising for efficiency, and the result was that the gap in chips widened while the gap in models narrowed.
Cowen's argument suggests a possible next step in that story. Suppose the US succeeds in maintaining a meaningful lead in frontier models. If national security concerns then limit who can access those models, where they can be deployed, and how widely they can be used, America could end up facing a different version of the same problem you identify for China.
Possessing a capability and successfully diffusing that capability are not the same thing.
That observation reminded me of last week's Military-Civil Fusion piece and its emphasis on conversion efficiency. The question may increasingly become not who possesses the greatest capability, but who can most effectively convert capability into economic value, widespread adoption, and geopolitical influence.
China's challenge appears to be converting domestic hardware into deployment at scale. The US may eventually face the opposite challenge: maintaining a capability lead while restricting access to the very capabilities that create strategic advantage in the first place.
Viewed through that lens, the competition starts to look less like a race for better chips and more like a race to convert technological leadership into real-world influence.
This is one of the more accessible and tightly argued Eye on China pieces I've read recently.
What I particularly appreciated is that it asks a concrete question, "did export controls work?" and then arrives at a nuanced answer rather than forcing a binary conclusion.
My summary of the article would be:
1. Export controls succeeded in widening the hardware gap between the US and China.
2. Chinese firms responded to those constraints by aggressively optimising for efficiency rather than waiting for hardware parity.
3. As a result, the gap in model performance has narrowed even as the gap in chips has widened.
4. China still faces serious challenges in scaling deployment because of chip manufacturing, cluster assembly, and interconnect bottlenecks.
5. Therefore export controls have slowed China but have not prevented it from remaining near the frontier.
Or, in a single line: the gap in chips widened, while the gap in models narrowed.
My only criticism is that the article occasionally buries its strongest observations under more technical exposition than necessary. The discussion of DeepSeek, Huawei chips, cluster efficiency, NVLink, yields, smuggling networks, and hyperscaler comparisons all support the same central conclusion, but the reader has to work a bit harder than necessary to extract it.
In fact, I think the most interesting observation in the piece is that the US has managed to increase the hardware gap while China has managed to reduce the model-performance gap. That tension explains much of what has happened in AI over the last few years.
A piece built around that central idea would be even stronger and perhaps 25–30% shorter without losing any substantive content.