China is moving toward allowing its biggest tech companies to buy Nvidia’s powerful H200 artificial-intelligence chips as early as this quarter, a shift that could reopen a crucial channel for advanced U.S. hardware and again test the limits of Washington’s export controls.
The emerging framework, first reported by Bloomberg and highlighted Thursday by the CSIS Economics Program and Scholl Chair in a post on X, would permit “select commercial use” by private firms while keeping military and other sensitive state-linked entities on the blacklist.
“If this holds, it’s the most significant reopening of China’s access to U.S. AI silicon since the crackdown began,” said one Beijing-based semiconductor adviser who works with major cloud providers.
“This is not a full thaw. It’s a controlled leak in a very tight valve.”
A narrow door opens—carefully
Approvals could begin between January and March 2026, according to people familiar with the talks. Regulators have signaled that large consumer internet platforms, cloud-computing providers and major commercial language-model developers will be at the front of the line.
Even so, officials have been telling companies not to get ahead of themselves. Regulators have privately warned big tech firms to stop informal pre-orders and aggressive forward-buying while the rules are still being drafted, in an effort to avoid a stockpiling frenzy that could distort prices and make enforcement harder.
The new access will be tightly circumscribed. Under the draft plan, military organizations, intelligence-linked institutions, critical infrastructure operators and state-owned enterprises are expected to be excluded outright. The resulting two-track system is designed to convince Washington that high-end U.S. chips will not be flowing into China’s most sensitive sectors, even as Beijing tries to keep its commercial AI champions in the global race.
A Shanghai-based policy analyst described the approach as “a political compromise wrapped in a technical license.”
“Beijing wants its internet giants to keep pace with Silicon Valley,” she said. “But it also wants to show Washington it can live within some red lines—at least for now.”
H200: A step behind Blackwell, but far ahead of China’s last option
The H200 is not Nvidia’s newest architecture, but it represents a dramatic jump from what Chinese companies were allowed to buy just months ago.
People briefed on performance estimates say the H200 delivers roughly six times the capability of Nvidia’s export-compliant H20—a chip custom-built for China to slip under earlier U.S. thresholds. For Chinese AI developers, that gap is the difference between merely staying in business and contending at the frontier.
“The H20 felt like running a marathon in ankle weights,” said an engineer at a large Shenzhen-based AI startup. “The H200 is not Blackwell, but it finally lets us train frontier models on a realistic schedule.”
ByteDance and Alibaba are each preparing to order more than 200,000 H200 units once they get the green light, according to people familiar with internal planning. Across China’s cloud providers, model labs and internet platforms, potential demand has been signaled for more than 2 million chips, at an estimated price of about $27,000 apiece.
On paper, that points to more than $54 billion in potential orders. In reality, supply is far tighter: Nvidia’s current H200 inventory is thought to be closer to 700,000 units, and meaningful new output is not expected until the second quarter, when Taiwan Semiconductor Manufacturing Co. is due to ramp production.
Beijing’s strings: buy American, also buy Chinese
Beijing is not simply throwing open the door to Nvidia and stepping back. Regulators are weighing a “paired-purchase” rule that would require buyers to take a set quantity of domestic accelerators—such as Huawei’s Ascend 910C or chips from Biren Technology—whenever they import batches of H200s.
On the whiteboard, that fits snugly with China’s long-term ambition of semiconductor self-reliance, nudging local champions into the same racks as Nvidia. On the balance sheet, it is more complicated.
“It’s like telling a Formula One team they can have a top European engine, but only if they also buy a slower local one and keep it in the garage,” said a venture investor focused on Chinese AI infrastructure. “Nobody wants to pay twice for the same compute.”
Still, for many cloud operators and model developers, the trade-off may be tolerable. The chance to cut weeks off training cycles for large language models, recommendation systems or generative engines often matters more than the purity of any one vendor’s hardware.
Nvidia’s big bet—with the risk pushed back to China
For Nvidia, which saw its China market share plunge from roughly 95% to near zero at the height of U.S. restrictions, the possible reopening is a financial windfall wrapped in political risk.
Chief Executive Jensen Huang has told investors that the company has “fired up” its supply chain and is negotiating additional H200 capacity with TSMC. Yet the commercial terms now being floated show Nvidia has little appetite to absorb another geopolitical shock.
According to people familiar with the negotiations, Chinese customers are being asked to pay in full, up front, with no rights to cancel, reconfigure orders or claim refunds if export rules tighten again. In limited cases, companies may be allowed to post commercial insurance or physical collateral instead of cash, but the core idea is the same: Nvidia ships only after the risk has been shifted.
One Hong Kong-based trade lawyer called it “regulatory risk offloading in its purest form.”
“Nvidia remembers the billions in write-downs after the last China bans,” he said. “This time, it’s saying: if the rules change overnight, you—not we—will hold the bag.”
Nvidia has kept public messaging deliberately guarded. Huang has suggested he does not expect Beijing to issue splashy statements about any deal. The real tell, he has said, will be whether Chinese clients can quietly place and receive orders.
That ambiguity leaves maneuvering room for both governments—and keeps companies on edge.
Washington, Brussels and Tokyo watch for creep
The reported arrangement fits into a broader, tentative de-escalation between Washington and Beijing, which have paused new tariffs and dialed down some of the sharpest rhetoric around tech decoupling. But even against that backdrop, U.S. export-control hawks are zeroing in on a familiar question: how to police what counts as “commercial.”
“The real issue is end-use drift,” said a former U.S. Commerce Department official now at a Washington think tank. “Chips that start in a cloud for social media can end up supporting dual-use AI, including systems relevant to the battlefield.”
He warned that any indication H200s are touching military projects, mass-surveillance platforms or critical infrastructure could trigger another round of fast, restrictive measures from Washington.
European policymakers, who have generally tracked U.S. tech-security moves rather than writing a separate China playbook, are expected to take their cues from Washington rather than draw their own red lines around AI chips. Indo-Pacific allies such as Japan and South Korea face their own balancing act: their chipmakers want access to Chinese demand and to Nvidia’s ecosystem, even as security services worry about strengthening a potential rival.
Hovering over all of it is Taiwan, home to TSMC and therefore a central link in the H200 supply chain—and the region’s most volatile geopolitical fault line.
A fragile equilibrium in the AI arms race
Inside China’s AI community, the mood this week has been a mix of relief and realism. If implemented as described, H200 access would turbocharge training for internet giants, autonomous-driving players, robotics startups and industrial-automation firms that have struggled to match the pace of their U.S. counterparts under current hardware constraints.
But no one inside the ecosystem is confusing this with a level playing field. Nvidia’s newest Blackwell chips remain off-limits, and the underlying U.S. architecture of controls is still designed to keep China’s cutting edge from converging with America’s. Future rules could once again close the door with little warning.
“Think of this as a pressure valve, not a peace treaty,” said the Beijing policy analyst. “Both sides are still building separate tech stacks. They’re just agreeing to trade in last year’s chips for now.”
For Nvidia, China’s internet giants and policymakers in both capitals, the next few months will show whether a tightly managed form of interdependence in strategic technologies can actually hold—or whether advanced AI hardware remains too sensitive an asset to live comfortably across rival borders.
If one chip can move markets, tilt supply chains and unsettle diplomats, the question hanging over this uneasy compromise is how long it can last before the next shock arrives.
