After Washington banned one of Nvidia’s flagship GPUs, grandfathered another, pushed for a “China-safe” version — and then yanked that — the Trump administration has found a new line in the sand: the H200. The White House now reportedly plans to let Nvidia export the powerful chips to China under a controlled regime, betting that limited access to last-generation hardware will protect U.S. leverage in the AI race instead of giving it away.

The move isn’t formal yet, but the contours are clear. Semafor reports that the White House will direct the Commerce Department to license Nvidia’s H200s — roughly 18 months behind Nvidia’s cutting-edge chips — for export to China as a compromise between blanket bans and a full reopening. Spokespeople for Nvidia, the Department of Commerce, and the White House didn’t immediately respond to a request for comment.

For Nvidia, the news lands like the first crack in a brick wall it has spent a year describing as permanent — and the stock jumped over 2% on the headline. The company’s chips are used to train most AI models, run Wall Street’s data centers, and keep Silicon Valley’s market caps inflated. But in China, once a $7 billion revenue stream, Nvidia’s racks have been sitting empty. U.S. export controls shut the door. Beijing’s recommendations then dead-bolted it.

“[China has] made it very clear that they don’t want Nvidia to be there right now,” Huang said at a media briefing at the company’s October GTC event in Washington. “I hope that will change in the future.” Before the bans, the ban reversals, and China’s closed door, Nvidia controlled roughly 95% of China’s advanced-AI chip market. Huang said the company was “100% out of China.” Analysts estimated those lost sales could cost Nvidia as much as $15 billion in annual revenue — and around $3 billion in U.S. tax receipts.

Nvidia hasn’t exactly been suffering in exile, though. The company walked into earnings carrying the weight of a monthlong AI panic and walked out with a $57 billion quarter that made any AI selloff look like stage fright. The data-center business alone grew into a $51 billion engine, the kind of gravitational force that reorganizes the rest of the sector around it. Wall Street had already braced for something good; Nvidia cleared the bar anyway and then raised guidance to $65 billion for the next quarter, all while assuming exactly zero from China.

“In all of our forecasts, we assume zero for China,” Huang told investors. “If anything happens there — which I hope it will — that will be a bonus.” On paper, the company has retrained itself to live without one of its most important markets. In practice, it’s been lobbying hard to make sure that “bonus” eventually exists.

Old chip, new leverage

Enter: the H200, an upgraded Hopper-generation workhorse that’s vastly more capable than the hobbled H20 chip Washington previously let ship to China, with enough memory and bandwidth to train and run giant transformer models at scale. In Nvidia’s world, it sits between yesterday’s H100 and tomorrow’s Blackwell chips — close enough to the frontier that Chinese labs could build serious supercomputers with it, just not quite at the bleeding edge President Donald Trump wants to keep locked up at home.

Trump has been unusually clear on that point. “We don’t give (the Blackwell) chip to other people,” he said on Air Force One. “The most advanced, we will not let anybody have them other than the United States.” Commerce Secretary Howard Lutnick has described the decision on H-class exports as sitting directly on Trump’s desk, with Huang on one side of the argument and national-security hawks on the other.

The president, Lutnick said recently on CNBC, is caught between “Jensen from Nvidia, who really wants to sell those chips” and “an enormous number of other people who think that that’s something that should be deeply considered.” The decision, in Lutnick’s words, is “in front of the president.” And the president has given Huang the green light.

Nvidia already agreed to one compromise this summer: handing the U.S. government roughly 15% of any China-related chip revenue in exchange for export licenses on disallowed parts, an unprecedented arrangement that blends national-security controls with a quasi-tax on individual companies. That deal covered chips such as the H20, the “China-safe” model Washington first blessed and then helped kill. Folding H200 into that same framework would turn the chip into exactly what Huang has been trying to sell in private — a way for the U.S. to tax and monitor access without walking away from the world’s second-biggest AI market.

There’s a geopolitical logic here that sounds suspiciously like Nvidia’s lobbying deck. Huang has warned repeatedly that cutting American chips out of China doesn’t freeze Chinese AI but rather forces Chinese labs and coders to optimize for someone else’s hardware. Export controls, he said earlier this year, were “a failure” because they cost U.S. firms “billions of dollars in lost sales” while pushing customers toward domestic alternatives. For a company whose moat is as much about software and standards as raw silicon, that’s the nightmare scenario.

A maybe-return to a very different China

But Beijing hasn’t exactly rolled out a welcome mat for a return. After the H20 workarounds, Chinese authorities nudged state-linked firms away from Nvidia gear and toward domestic alternatives on security grounds. “It’s up to China when they would like Nvidia products to go back to serve the Chinese market,” Huang said recently in Taiwan.

After meeting Trump at the White House last week, the Nvidia CEO admitted he has “no clue” whether Chinese buyers would even come back for H200-class chips if the U.S. opens the gate, because Beijing is now loudly prioritizing homegrown AI and domestic GPUs.

China’s Huawei moved quickly to fill the AI vacuum with its Ascend line, stitching together domestic supercomputers designed to mimic Nvidia-style GPU clusters. State-backed systems now run entirely on local silicon. The state has been busy minting "China’s Nvidias." Domestic GPU startups tripled in valuation. Moore Threads — founded by a former Nvidia executive — just debuted in Shanghai with a stock-market pop big enough to turn it into a tens-of-billions-of-dollars mascot for Beijing’s self-reliance drive. The chips still lag behind Nvidia’s best, but the money and political capital flowing into local AI silicon is the point.

But the U.S. policy ground under all of this is still shifting. Last week, a bipartisan Senate bill proposed a 30-month ban on exporting advanced AI chips such as the H200 and Blackwell to China — the SAFE Chips Act — a direct attempt by Trump-skeptical hawks to lock in a harder line. “The United States leads the AI race with Communist China due in large part to our dominance of global compute power,” GOP Sen. Pete Ricketts, a co-author of the bill, said in his announcement. “Denying Beijing access to these AI chips is essential to our national security.”

A partial reopening of China would give Nvidia a shot at reclaiming a slice of a market it has already written down to zero, while still leaving its Blackwell line — the real profit engine — off-limits for now. AMD and other AI chipmakers would expect similar treatment.

For now, it seems that the Trump administration is trying to move from whack-a-mole export controls to something more calibrated, where last year’s GPU is the pressure valve and this year’s stays locked away. But even if Commerce signs the licenses, Congress has to decide whether to pass SAFE Chips and slam the door on China; Beijing has to decide whether H200s are politically acceptable; and Chinese tech companies have to decide whether they still want to bet billions on Nvidia when their own government is leaning hard on homegrown silicon.

If the H200 deal sticks, it could restore some of Nvidia’s lost China revenue and keep U.S. chips embedded in the world’s second-largest AI market. If it doesn’t, the U.S. risks replaying the last three years: surrendering market share, hardening Beijing’s resolve, and discovering again that the real leverage was just in the hardware.

📬 Sign up for the Daily Brief