US-China AI Chip War 2026: Where Nvidia, Tariffs, and the H200 Stand Now

In 2024, Nvidia held roughly 95% of China’s advanced AI chip market and booked about $17 billion in China revenue. By mid-2026, Nvidia’s own filings describe the company as effectively foreclosed from China’s data-center compute market. The demand did not vanish. US policy, Chinese policy, or both, made the sales impossible to complete.

The US-China AI chip fight is the most consequential technology trade conflict in a generation, and it has been unusually volatile. This is the timeline, where things stand as of late July 2026, and the open questions, with contested claims attributed to their sources rather than stated as settled fact.

Can Nvidia sell H200 chips to China right now?

On paper yes, in practice barely. Nvidia holds US export licenses to sell H200 chips to about 10 approved Chinese companies, but almost none have shipped, and Nvidia reports it has generated no China data-center compute revenue.

The clearest recent marker came on July 14, 2026, when Commerce Department export-controls official Jeffrey Kessler told Congress that H200 shipments to China had been “very few” despite roughly $10 billion in approved licenses. On Nvidia’s most recent earnings call, CFO Colette Kress said the company had yet to generate any revenue from H200 sales to China and was uncertain whether any imports would be allowed, and Nvidia again excluded China data-center revenue from its forward guidance. After a visit to China alongside Nvidia CEO Jensen Huang, President Trump said Beijing had not approved H200 purchases even though the US had cleared them. The result is a stalemate: licenses on one side, no chips moving on the other.

How did the chip war reach this point?

The conflict escalated through four years of restriction, workaround, and reversal. The short version is that each US control prompted Nvidia to design a compliant chip, and each compliant chip prompted a new control.

The sequence runs as follows. In October 2022, the Biden administration imposed sweeping export controls that effectively banned Nvidia’s A100 and H100 for China, and Nvidia responded with downgraded parts (the A800, H800, and later the H20). Through 2024, that cat-and-mouse pattern repeated, and Nvidia’s China revenue fell from around 20% of its total toward roughly 13%. In April 2025, the US restricted even the compliant H20, and Nvidia took a multibillion-dollar writedown while AMD faced similar limits on its MI308. In December 2025, Trump reversed course and approved H200 sales to China under a revenue-share structure, and in early 2026 Commerce cleared about 10 Chinese firms, including Alibaba, Tencent, ByteDance, and JD.com, plus distributors, to buy up to 75,000 units each. Then the deliveries stalled.

What is the 25% tariff on Nvidia chips to China?

The 25% figure is a revenue-share arrangement, not a conventional import tariff. Under the deal Trump signed in January 2026, advanced chips bound for China route through the US for mandatory testing and certification, and the US government receives 25% of the sale revenue.

The structure was pitched as a middle path between a total ban and open sales: the US keeps oversight and takes a cut, while Nvidia keeps a foothold in a large market. In practice it has generated little, because the chips are not moving. It has also drawn criticism from multiple directions, which is covered below. Beijing, for its part, has reportedly raised security concerns about the revenue-share model itself, seeing US-routed certification as a channel for surveillance.

Why won’t the approved chips actually ship?

The chips won’t ship because requirements on both sides block nearly every transaction, and Beijing is actively steering buyers elsewhere. It is a two-sided blockage, not a single bottleneck.

On the US side, approved Chinese buyers must demonstrate security procedures and certify non-military use, and reporting indicates many buyers failed tightened checks. On the Chinese side, Beijing has told companies not to buy US chips as a retaliatory measure and has pushed “buy local” mandates favoring Huawei. US Commerce Secretary comments and Kessler’s testimony both point to Chinese government blocking of imports as a central reason nothing moves. Nvidia’s own filings frame the outcome bluntly, saying its foreclosure from China helped competitors build larger ecosystems to challenge it worldwide. Whether the fault lies more with US licensing friction or Chinese blocking is genuinely disputed, and both are operating at once.

How is this affecting Nvidia’s business?

Nvidia is losing a large China revenue stream and growing rapidly anyway, because demand elsewhere has more than replaced it. The company posted record results with China revenue essentially zeroed out of its numbers.

Nvidia reported quarterly revenue of $81.6 billion, up 85% year over year, with data-center revenue of $75.2 billion, and guided to roughly $91 billion the following quarter, all while assuming no China data-center compute revenue. That tells two stories at once: AI demand outside China is strong enough that Nvidia is growing without it, and the company is still leaving an estimated $15-20 billion in annual China revenue unrealized, revenue that domestic Chinese rivals are pursuing. Its record run is documented in our coverage of Nvidia’s record AI-driven earnings and its ecosystem investments in Nvidia’s AI empire. Major customers preparing IPOs, covered in our look at the SpaceX, OpenAI, and Anthropic listings, remain heavy buyers.

This article is reporting, not investment advice. Nothing here is a recommendation to buy or sell any security.

Is China building its own AI chips instead?

Yes, and analysts widely argue the controls accelerated exactly the domestic capability they were meant to slow. The clearest evidence is Huawei’s rise in China’s own market.

Bernstein estimates, cited via the Economist and SCMP, put Nvidia’s share of China’s AI-chip market at roughly 8%, with Huawei around 50%, a reversal from Nvidia’s former dominance. Analysts are careful to note this does not mean Huawei’s Ascend chips match Nvidia’s top parts on performance; they don’t, on current benchmarks. It means a protected market with mandatory domestic procurement is a powerful industrial-policy tool. Three developments stand out:

• Huawei Ascend is the leading Chinese alternative, and Chinese cloud providers have integrated it despite a US assessment that it was developed in violation of export controls. Huawei has also built MindSpore as an open-source rival to Nvidia’s CUDA software.
• Chinese AI labs are designing around it. DeepSeek’s V4 model, released in April 2026, was reported as explicitly optimized for Ascend processors, a signal that frontier Chinese labs are not waiting for H200 deliveries. We cover the models in our pieces on Kimi K3 and whether DeepSeek is safe.
• Memory remains the chokepoint. China’s CXMT is working toward HBM3-class high-bandwidth memory but is estimated to trail Samsung and SK Hynix by about two generations, which caps Ascend performance on the most demanding training workloads. The broader memory squeeze is in our memory shortage coverage.

What does Congress want to do about chip exports?

Congress is split from the Trump administration and internally, with several competing bills, so the legislative picture is unsettled rather than decided. Reporting here should be read as a snapshot of live debate.

The main measures are worth naming precisely. The GAIN AI Act, attached to the 2026 defense authorization bill, would give US customers a right of first refusal on advanced chips before a maker could export them; supporters say it stops US buyers waiting behind Chinese firms, while critics including analysts at CSIS argue it would damage the global competitiveness of US chipmakers and hand share to Huawei. Reporting in late 2025 indicated lawmakers moved to exclude the GAIN AI Act after Nvidia lobbying, which industry read as a win. Separately, a House foreign-affairs committee advanced bipartisan legislation for arms-sale-style congressional oversight of AI chip sales, which would also codify a multi-year ban on exporting Nvidia’s more advanced Blackwell chips to China. The administration, meanwhile, has generally favored the licensed-sales approach. The policy dimension connects to the broader export-control story we covered in the US AI export controls piece.

Who is winning the chip war?

There is no clean answer, and anyone offering one is picking a side of a genuine dispute. The honest framing is that the scorecard depends entirely on the metric.

By one reading, US controls are working: China is cut off from the most advanced training hardware, and its domestic chips still trail on performance and memory. By another, the controls backfired, handing Huawei a guaranteed home market, accelerating Chinese self-sufficiency, and costing US firms tens of billions while Chinese labs design around the gap. Critics of the sales, such as the Council on Foreign Relations, argue that permitting H200 exports narrows the US lead; critics of the controls, such as the administration’s own tech advisers, argue that bureaucratic delay is a gift to Huawei. The supply chain is visibly splitting into two ecosystems, US-aligned and China-aligned, and that bifurcation is the one trend nearly everyone agrees on. For the AI models trained on these chips, the outcome shapes what gets built, at what cost, and by whom, a thread we follow in our comparison of the leading models and our look at the 2026 AI spending reckoning.

FAQ

Can Nvidia sell AI chips to China right now?

Nvidia holds US licenses to sell H200 chips to about 10 approved Chinese companies, but almost none have shipped. A Commerce Department official told Congress on July 14, 2026 that shipments were “very few,” and Nvidia reports no China data-center compute revenue. Beijing’s “buy local” directives and security requirements on both sides have blocked nearly every transaction.

What is the 25% tariff on AI chips to China?

It is a revenue-share arrangement rather than a standard tariff. Under the January 2026 deal, advanced chips bound for China route through the US for testing, and the US government receives 25% of the sale revenue. The structure was meant to allow limited, monitored sales, but little revenue has materialized because the chips are not being delivered.

How much China revenue is Nvidia losing?

Estimates put the unrealized China data-center revenue at roughly $15-20 billion annually at current levels, down from China being around 20% of Nvidia’s data-center revenue before restrictions. Despite that gap, Nvidia posted record overall growth because demand outside China more than compensated. Figures are estimates, and Nvidia excludes China revenue from its guidance.

Is China making competitive AI chips?

China is building capable AI chips, led by Huawei’s Ascend line, and Bernstein estimates Huawei now holds around half of China’s AI-chip market. Analysts caution that Ascend chips still trail Nvidia’s most advanced parts on performance and depend on domestic high-bandwidth memory that lags Samsung and SK Hynix. They are increasingly used for inference and, per reports, for training runs at labs like DeepSeek.

Will Congress block Nvidia’s China chip sales?

It is unresolved. Several bills are in play, including the GAIN AI Act, which would give US buyers first refusal on advanced chips, and separate bipartisan legislation for congressional oversight of chip exports plus a Blackwell ban. Reporting suggests the GAIN AI Act was set aside after industry lobbying, but the debate is ongoing and the administration and Congress remain divided.