Export controls now a key factor in AI chip development – adding risk for the whole industry

The physics of transistors and politics of trading licenses are colliding on the AI frontier

Analysis Few of us would have imagined that national security would play such a key role in AI hardware, even dictating its development, but here we are – in a new era of export controls.

Trade restrictions on advanced AI chips are reshaping the entire semiconductor industry in unprecedented ways. Where previously the design and performance of AI accelerators were primarily driven by physics, manufacturing yields, and customer demands, they are now controlled by US export restrictions.

The US's Department of Commerce has set strict limits on how leading American technology companies, such as Nvidia and AMD, design AI chips, with restrictions on performance, memory bandwidth, and interconnect speeds. This has forced technology companies to create specialized, downgraded versions of their flagship AI chips solely to comply with regulations.

... [I]nnovation is sometimes running in reverse as companies intentionally cap capabilities to allow their global sales to grow. The risk is that US firms spend more time crippling products than advancing them, turning engineering into compliance....

These China-only variants sacrifice the capability for the sole purpose of staying below the US export threshold. The inevitable nature of such policy changes shows how the government can directly impact chip schematics. So lets explore how US export controls on AI chips have impacted chip design and the broader market, and analyze the technical and geopolitical consequences.

Design under constraint

During the Cold War era, the United States attempted every means to prevent the Soviet Union from acquiring advanced Western technology. But when the United States opened its technology to the world, the entire global technology ecosystem benefited from this.

However, under his previous administration from 2017 to 2021, Donald Trump had already begun to repurpose Cold War tactics. While the US had export controls on semiconductor and chipmaking equipment in place for years, a new phase of escalated controls on China began in 2018, when Jeff Sessions, US Attorney General at the time, announced an initiative to fight what he described as "Chinese economic espionage." By 2019, the US was pressuring the Netherlands to halt the sale of ASML's most advanced EUV lithography tools to China.

The concern that China would advance with the most advanced technological tools intensified over the pandemic and the advent of generative AI. Suddenly, every market in the world demanded AI chips, and who better than Nvidia with its advanced GPUs to support the growing generative AI technology space.

In 2022, the US Department of Commerce’s Bureau of Industry and Security (BIS) announced sweeping new export rules explicitly on AI chips. These rules banned the export to China of top-tier GPUs, such as the Nvidia A100 and H100, as well as similar chips from AMD. The policy changes defined technical threshold criteria.

One of the key ways to limit the development was to throttle interconnects and bandwidths. High-speed chip-to-chip communication, important for scaling AI training across many GPUs, was an early target of US export controls. The October 2022 regulation prohibited the export of any GPUs with an aggregate interconnect bandwidth exceeding 600 GBps. The intent was to restrict only the most advanced AI capabilities while allowing older technology to flow.

But the company quickly developed a China-focused GPU, the A800, with reduced performance to meet the new limits. The A800 was essentially an A100 with its NVLink interconnect bandwidth dialed down to 400 GBps from 600 GBps. Similarly, when Nvidia's next-gen H100 was banned, a China-specific H800 was introduced, featuring approximately 300 GBps of interconnect bandwidth compared to 900 GBps on the original H100. These modified chips also had slightly lower throughput and memory speeds, using firmware and clock adjustments.

European stakeholders responded by planning to de-Americanize parts of their supply chain...

Beyond interconnects, newer US rules emphasize total compute power. The 2023 BIS update set a total processing performance ceiling and even a performance density limit. Complying with these meant chip designers had to either disable functional units, lower clock frequencies, or otherwise cripple the chip's peak throughput.

In effect, the product key specifications of the newer variants are determined by US export regulations rather than solely by engineering considerations. Such firmware-capped GPUs can be viewed as safety valves; they physically could run faster, but the vendor locks them down to produce export-compliant part numbers.

High-bandwidth memory stacks have even been throttled. For instance, the H800 uses slightly lower memory clock rates and delivers approximately 12 percent less memory bandwidth. This adjustment was likely made to reduce overall data throughput and model training performance.

However, after a series of regulatory updates, by late 2023, Nvidia H20 was the only high-end AI accelerator that complied with the export controls and slipped below the new limits.

Unlike normal product development, where improvements in speed and efficiency are celebrated, here innovation is sometimes running in reverse as companies intentionally cap capabilities to allow their global sales to grow. The risk is that US firms spend more time crippling products than advancing them, turning engineering into compliance.

Market and global impact

One immediate effect of US chip export limits is that customers, particularly in large markets like China, are reevaluating their supply chains. Companies now anticipate that any given high-performance chip could suddenly become unavailable due to a policy change.

Initially, Chinese AI companies and cloud providers hoarded the supply while it lasted, as the bans expanded to cover even the downgraded substitutes. The Huawei Ascend 910 series AI accelerators experienced a sudden surge in orders.

The diversification push isn't confined to China. European tech companies and research centers, also alarmed by US export decisions, are now seeking alternatives. A striking example came when the US announced in early 2025 that certain EU countries would face the same AI chip import limits as China.

European stakeholders responded by planning to de-Americanize parts of their supply chain. According to an analysis by the European Center for International Political Economy, EU companies that rely on Nvidia GPUs are now acknowledging their vulnerabilities. They are seeking alternative suppliers from South Korean, Taiwanese, or domestic European providers.

The US continued expanding its export controls. In late 2024 and January 2025, Washington proposed an "AI Diffusion Rule" to globally curb AI chip exports. The regime would have divided countries into tiers, allowing unrestricted export to a select 18 allies, but limiting AI chip sales to most other nations.

Facing pushback, the US withdrew the AI diffusion rule in mid-2025, opting instead for a more calibrated approach. By mid-2025, there were also signs of slight easing, for instance, the administration allowed Nvidia H20 and AMD MI308 to resume sales to China under special licenses.

The risk is fractured supply chains and weakened trust, even among US allies.

Long term strategic risk

Export controls on AI chips have led to the downgrading of export-friendly models. This has had a range of consequences across the semiconductor and tech industries. Some of these outcomes were intended to slow a rival's AI progress, while others are unintended side effects impacting businesses and innovation globally.

The change in approach from Chinese companies meant that millions of dollars that would have gone to Nvidia R&D are now funding Huawei's advancement instead. In effect, US export controls created an opening for China's nascent AI chipmakers to gain market share at home, as local buyers diversify their suppliers.

Perhaps the most profound long term risk of aggressive export controls is that they can accelerate the emergence of a rival technology ecosystem. By attempting to slow a competitor's progress, such controls may actually motivate them to innovate more quickly in alternative ways, ultimately reducing the controlling country's leverage.

The story of export controls on AI chips shows the powerful and sometimes perverse interplay between technology and geopolitics. When policy, rather than physics, dictates chip specifications, innovation suffers. Engineers are optimizing to legal limits, not to the boundaries of what silicon could do.

Now China is racing to remove the dependency that makes it vulnerable in the first place, investing in homegrown chips and novel architectures. Export controls have become a double-edged sword, a tool to guard today's advantage, but also a catalyst for competitors to find new paths for tomorrow. ®