China Quietly Lets Nvidia's H200 Back In — ByteDance and Tencent Each Get 10,000 Chips

For most of 2026, the world’s most sought-after AI chip has been sitting in a strange limbo: legally cleared for export by Washington, yet blocked at the dock by Beijing. That standoff now shows its first real cracks. The Financial Times reported this week that small batches of Nvidia’s H200 — one of the company’s most powerful AI processors — have been allowed to enter mainland China, with ByteDance and Tencent each receiving roughly 10,000 chips at their mainland facilities in recent weeks, and a few other Chinese technology firms potentially following soon.

It is a quietly enormous development in the US–China AI race. Not because of the volume — 20,000 chips is a fraction of the more than 2 million units Chinese firms reportedly ordered when Washington first approved the H200 for export — but because of what it signals about where both governments’ red lines actually sit.

A chip caught between two capitals

The H200’s odyssey through export politics has been tortuous. In December 2025, the Trump administration reversed Biden-era restrictions and greenlit China-bound sales of the H200 — Nvidia’s second most powerful chip at the time — in exchange for a 25% cut of the revenue. Chinese companies placed orders worth tens of billions of dollars almost immediately; Reuters reported commitments exceeding 2 million chips at around $27,000 each.

Then came the whiplash. In January 2026, Chinese customs authorities told agents the H200 was “not permitted” to enter the country, freezing an estimated $54 billion in pending orders. Suppliers halted China-bound H200 production. Beijing’s calculus was transparent: it wanted Chinese AI giants to keep buying — but to park the hardware in places like Hong Kong or overseas data centers rather than the mainland, so as not to undercut its push for domestic chips from Huawei and others. The Cyberspace Administration of China had already directed ByteDance, Alibaba and Tencent to limit purchases of certain Nvidia chips, and reports suggested the H200 would be treated the same way.

By late January, Beijing gave a narrow green light for importing a first batch. The FT’s new reporting shows that the trickle has now actually become delivery: chips are physically arriving at mainland facilities, not sitting in bonded warehouses or Hong Kong server rooms.

Why Beijing changed its mind

The FT frames the easing as Beijing “trying to aid its leading AI companies in the race” — and the race has gotten urgent. Chinese labs are training frontier models at a pace that has surprised Western observers, and the compute gap is the single biggest constraint they face. The H200 is at least two generations behind Nvidia’s most powerful Blackwell-era chips, which remain off-limits to Chinese buyers under US export rules. But “two generations behind” is still a massive upgrade over what most mainland teams can otherwise access at scale, and the CFR has noted the H200 offers over six times the processing power of any chip previously approved for China export.

Allowing limited imports also gives Beijing something it values: leverage and optionality. China keeps its domestic substitution drive intact — the quotas stay small, approvals stay selective, and regulators can turn the tap off at any moment — while its champions get enough top-tier compute to stay competitive in the near term. It is industrial policy as a rheostat, not a switch.

For Washington’s part, the US has cleared around 10 Chinese firms — including Alibaba, Tencent, ByteDance and JD.com — to buy the H200, with conditions attached. The January 2026 BIS rule layers a 25% tariff-style surcharge and a 50% volume cap on top, meaning only a slice of those 2 million ordered chips can ever legally ship. Both governments, in other words, have converged on the same awkward answer: let some chips through, tightly metered, while pretending the broader decoupling continues.

What it means for Nvidia

For Nvidia, even a partial reopening is welcome news in what has been a brutal stretch for its China business. Jensen Huang has said publicly that US export controls erased Nvidia’s share of the China AI chip market — once 95% — “to zero,” and the company stopped expecting significant H200 revenue from China earlier this year after months of approvals, reversals, and blocked shipments. Nvidia’s China-specific chips have cycled through the H20 and toward Blackwell-derived variants like the B30A, none of which have fully escaped the political crossfire.

Ten thousand chips per customer is real but modest money — on the order of $270 million each at list price, a rounding error against Nvidia’s data center revenue. The significance is directional: if the trickle becomes a flow, the multi-billion-dollar backlog of cleared orders could start converting to revenue. If it stalls again, it confirms that Nvidia’s China business is now entirely hostage to two governments’ moods.

The bigger picture

The subtext of the FT report is that both countries are finding the full decoupling of AI hardware harder to sustain than their rhetoric suggests. China wants sovereign chips, but its top labs need Nvidia’s interconnect and software ecosystem today. The US wants to deny China compute, but it also wants the 25% cut and the leverage that comes from being the supplier. Small shipments of 10,000 chips are what a compromise looks like when neither side wants to admit it compromised.

Watch the next signals closely: whether Alibaba and other cleared firms begin taking mainland deliveries, whether Beijing formalizes a quota regime, and whether Washington’s response to mainland installations — rather than Hong Kong warehousing — triggers another round of restrictions. The H200 has spent a year being a diplomatic instrument first and a processor second. That hasn’t changed; the music has just gotten slightly quieter.