US chip export controls: Nvidia's $4.5bn charge and the fight over what happens next

US chip export controls began in October 2022 with a single Bureau of Industry and Security rule. Four years on, they have reshaped Nvidia's balance sheet, split Washington's AI industry into two camps and pushed China's biggest model developers toward building their own silicon.

The policy logic was simple. Advanced AI chips need specialised manufacturing, and the US and a handful of allies controlled the inputs. Restrict the exports, and China's military and intelligence services lose access to the computing power needed to train frontier models. That argument has held up in public for four years. The evidence underneath it is messier.

The numbers Nvidia is carrying

In April 2025, the US government sent Nvidia a letter. No fine, no indictment, just notice that a licence would now be required to export its H20 chip to China. According to NeuralWired's 2026 guide, Nvidia disclosed a $4.5bn charge against earnings within days and projected a further $8bn in lost H20 revenue for the following quarter.

The H20 was itself a compliance product. It had been built to sit below the performance thresholds set in October 2023, after BIS closed the A800 and H800 workaround by switching from raw compute metrics to performance density. Nvidia's share of China's AI chip market fell from almost 95% at the start of the Biden administration to 50%, Jensen Huang said at the Computex forum in Taipei in May 2025, as reported by The Guardian.

Huang put the total write-off at "billions of dollars" and called the controls a "failure" because they had spurred Chinese developers rather than slowed them. "The local companies are very, very talented and very determined, and the export control gave them the spirit, the energy and the government support to accelerate their development," he told the Computex tech show.

Anthropic and Nvidia trade accusations

The split inside the US industry went public on 1 May 2025, when Nvidia and Anthropic clashed over the shape of the controls. Anthropic, which is backed by billions from Amazon and relies heavily on Nvidia hardware to train its models, had argued in a blog post the day before for tighter enforcement.

It described Chinese smuggling tactics that included chips hidden in "prosthetic baby bumps" and "packed alongside live lobsters", and cited the 2022 arrest of a woman carrying chips into China and the 2023 seizure of "computer display cards" with a lobster shipment in Hong Kong. Anthropic proposed lowering the export threshold for Tier 2 countries, tightening rules to reduce smuggling risk and increasing funding for enforcement.

"American firms should focus on innovation and rise to the challenge, rather than tell tall tales that large, heavy, and sensitive electronics are somehow smuggled in 'baby bumps' or 'alongside live lobsters'," an Nvidia spokesperson told CNBC.

The same spokesperson rejected the use of policy to limit competitiveness. "China, with half of the world's AI researchers, has highly capable AI experts at every layer of the AI stack. America cannot manipulate regulators to capture victory in AI," he said.

Anthropic stood by its submission. "Anthropic stands by its recently filed public submission in support of strong and balanced export controls that help secure America's lead in infrastructure development and ensure that the values of freedom and democracy shape the future of AI," a company spokesperson said. The company's core argument was that compute access is the key strategic chokepoint in frontier AI, a position that puts it directly against its own chip supplier's commercial interest.

DeepSeek moves into silicon

The clearest test of whether the controls worked is what Chinese labs did next. Ars Technica reported on 7 July 2026, citing Reuters and three people familiar with the matter, that DeepSeek has been working on a move into silicon for about a year, meeting potential hardware partners and hiring engineers. The focus is data centre chips for inference, not training, and the goal is likely to reduce reliance on both Huawei and Nvidia.

DeepSeek is not alone. Huawei controls about half of China's data centre chip market, according to Ars Technica, and Alibaba and Baidu have been making similar moves. The same logic is now visible in the US: OpenAI and Broadcom jointly announced Jalapeño, OpenAI's first chip designed for inference at scale, and Anthropic has been exploring custom chip design without publicly visible milestones.

That convergence matters for how the policy gets judged. If every large AI company eventually designs its own accelerators, export controls on merchant silicon become a narrowing tool. They still bite on the training frontier, where Nvidia's top parts have no close substitute. They bite much less on inference, which is where the money is heading.

The enforcement gap

Anthropic's smuggling claims point at a number the policy debate rarely mentions. NeuralWired cites a ratio of 20 to 1 between the value of a single smuggling case, involving Super Micro, and the enforcement budget available to chase it. That is the arithmetic behind Anthropic's call for more funding, and behind Nvidia's dismissal of the anecdotes as tall tales.

The legal machinery has expanded regardless. The Foreign Direct Product Rule extends US jurisdiction to chips made abroad with US tools, which makes Washington the licensing authority for TSMC in Taiwan and Samsung in South Korea. The Entity List has grown to hundreds of Chinese semiconductor and AI companies, with 42 added in March 2025 alone, per NeuralWired. In September 2025, BIS removed the named Chinese facilities of Samsung and SK Hynix from the validated end-user programme, effective 31 December.

China has answered in kind, banning exports of gallium, germanium and antimony to the United States within days of the December 2024 package, according to Referently's timeline. Its commerce ministry called the latest US warnings "typical unilateral bullying and protectionism" after Washington cautioned other countries that using Huawei's Ascend chips could breach US rules.

Allies have been lukewarm. The chief executive of ASML, restricted from shipping its most advanced equipment to China, said the US-led restrictions had become more "economically motivated" over time and expected push-back, CNBC reported in September 2024. South Korea's trade minister said Seoul should get more incentives if Washington wants it to comply with additional curbs.

What is actually still standing

Referently's timeline shows the core China controls have survived the political churn. Chipmaking equipment limits, Entity List denials, the US persons rule and the foreign direct product rules all remain in force. What has moved is the top end. The AI Diffusion Rule was rescinded, H20 licences resumed in July and August 2025, and BIS shifted licence review for H200-class chips headed to China and Macau from a presumption of denial to case-by-case review, with conditions on US supply and independent testing.

The Affiliates Rule, which applies Entity List restrictions to companies at least 50% owned by listed parties, was suspended for a year after the Trump-Xi meeting in Busan. It is due to return on 10 November 2026 unless BIS extends it, and China's own suspended mineral controls expire around the same date. BIS has said it is drafting a replacement for the AI Diffusion Rule.

So the question now is not whether the controls exist. It is whether they are a chokepoint or a subsidy. Huang's answer, delivered in Taipei, is that they handed Chinese developers the government support and the urgency to build without Nvidia. Anthropic's answer is that the alternative is worse. Both companies are describing the same four years from opposite ends of the same supply chain, and neither has published the data that would settle it.