The Agentic Post
Breaking
How to Build an App With AI and No Code: A Step-by-Step Beginner Guide  ·  Perplexity vs ChatGPT Search vs Google AI Mode: Which AI Search Is Best in 2026?  ·  What Are Vision-Language-Action (VLA) Models? How Robots Learn to Follow Instructions  ·  How to Translate Documents With AI Accurately (and When Not To)  ·  MCP vs A2A: The Two AI Agent Protocols Explained in Plain English  ·  Is Grok Accurate for News? What Studies Show About AI Search Citations  ·  
Home/Hardware & Robotics/Chips & GPUs
AI Chip Export Controls Explained: What the US Lets Nvidia Sell to China

AI Chip Export Controls Explained: What the US Lets Nvidia Sell to China

Chips & GPUs

US AI chip export controls in plain English: what they ban, how H20 and H200 rules changed, why China is refusing Nvidia chips, and why it matters in 2026.

AI chip export controls are US rules that stop companies like Nvidia from selling their most advanced chips to China without a government licence. They began in 2022 and have changed repeatedly since. By 2026, the US has approved some sales of Nvidia’s H200 chip, but China has told its own companies not to buy, leaving US chipmakers with almost no share of China’s AI chip market.

Why does the US restrict AI chips?

Advanced AI chips power both commercial AI and military uses like surveillance and weapons design. The US wants to keep a lead, so it limits the most powerful chips and the machines used to make them. For background on why these chips are so valuable, see our AI chip supply chain explainer.

How have the rules changed over time?

Here is the short timeline, based on a Brookings analysis:

  • September 2022: broad controls cut off high-end AI chips and chipmaking tools for China, tightened in 2023
  • April 2025: Nvidia’s China-specific H20 chip was ruled to need a licence; Nvidia took a 5.5 billion dollar write-off
  • Summer 2025: the US said it would grant H20 licences, with the government taking a 15% cut of certain sales; China then told its firms not to buy H20s
  • December 2025: the US said the more powerful H200 could be exported to China
  • May 2026: H200 sales were approved for 10 Chinese companies, but China has not let them buy; Commerce also required licences for chips going to China-headquartered firms even outside China

Why is China refusing Nvidia chips?

Brookings lists several reasons: Beijing wants its own chip industry to grow, it no longer trusts US rules to stay stable, and its models are getting by without US chips. DeepSeek’s V4 model, for example, was optimised to run on Huawei’s Ascend chips. Read our report on the DeepSeek ban for the other side of the rivalry.

What does this mean for Nvidia?

China was once a large market: Nvidia earned an estimated 12 to 15 billion dollars from H20 sales in 2024, Brookings notes. Losing it puts more pressure on sales elsewhere, and the issue is now politically hotter after Nvidia’s board opened a probe into CEO Jensen Huang over an alleged Chinese intelligence link, prompting calls in Congress for tougher chip legislation.

Do export controls affect ordinary users?

Indirectly. They shape which companies can build the biggest models, where data centers get built, and chip prices worldwide. They also explain why Chinese AI labs focus so hard on efficiency, using tricks like mixture of experts to do more with fewer chips.

What should I watch next?

Whether Congress passes new chip export laws, whether China ever allows H200 purchases, and how fast Huawei and other Chinese chipmakers close the gap. Rules in this area change often, so always check the date on any export control news.

Up Next
CoreWeave Asks Customers to Pay Up Front: What Its Q3 Results Say About AI Data Center Funding

CoreWeave Asks Customers to Pay Up Front: What Its Q3 Results Say About AI Data Center Funding

Data Centers

CoreWeave grew revenue 168% to $3.66 billion but lost $709 million, and now funds growth with about $9 billion in customer prepayments. What it means for the AI build-out.

CoreWeave, one of the biggest renters of AI computing power, reported third-quarter revenue of 3.66 billion dollars, up 168% from a year earlier, but a net loss of 709 million dollars. To pay for its huge build-out without more expensive borrowing, it is asking customers to pay in advance and is signing shorter contracts. It expects about 9 billion dollars in prepayments and deposits over the next three months.

What is CoreWeave?

CoreWeave is a “neocloud”: a company that builds data centers packed with Nvidia GPUs and rents that computing power to AI labs and businesses. Customers have included major AI developers. It is a key part of the AI supply chain, and its results are watched as a signal of whether AI spending is sustainable; see AI capex explained.

What were CoreWeave’s Q3 2026 numbers?

From Constellation Research and CNBC via AI Weekly:

  • Revenue: 3.66 billion dollars, ahead of analyst expectations
  • Net loss: 709 million dollars
  • Backlog: 132.5 billion dollars, up from 104 billion in June
  • Debt: nearly 47 billion dollars; interest paid in Q3 was 846 million
  • 2026 revenue guidance: raised to 13.0 to 13.3 billion dollars
  • 2026 capital spending: 37 to 39 billion dollars

Why is CoreWeave asking customers to pay up front?

Because borrowing has become expensive. CEO Michael Intrator said CoreWeave “will not use the unsecured debt market to fund our business at current yields, nor do we need to.” Instead it collected prepayments on 70% of deals this year, with some customers funding up to 100% of contract value up front, and it ties loans to specific customer contracts. Shorter contracts also let it re-price capacity as demand rises.

Is demand for AI computing still strong?

CoreWeave says yes. Intrator said it has “10 times as much qualified demand as available capacity through the end of 2027,” and prices for older Nvidia Ampere and Hopper GPUs rose 30% to 50% in the quarter. Contracted power rose to more than 4.6 gigawatts. That supports the view that computing is still scarce, even as power limits and local opposition slow new data centers.

What are the risks?

  • Debt: heavy borrowing makes CoreWeave sensitive to interest rates and any drop in demand
  • Contract roll-off: more than 130,000 GPUs come off contract in 2027 and about 170,000 in 2028, and must be re-rented
  • Customer concentration: a few big AI labs account for much of demand
  • Market mood: shares were down about 32% this year before results, and regulators are watching AI debt; a Fed governor this week warned of a possible large correction

What does this mean for the AI boom?

Customers paying in advance is a sign they badly want capacity, but it also shifts risk onto AI labs’ balance sheets. Combined with Oracle’s debt and Broadcom’s financing deals, it shows the AI build-out increasingly runs on creative financing. This is not investment advice.