Nvidia H200 to China: Washington swaps export bans for a 25% rake

The US is weighing whether to let Nvidia ship its H200 AI accelerators to China, and the headline hook is pure 2025 Washington. The Trump administration wants an interagency review and a reported 25% government fee on sales if they proceed with export controls, but with a cashier.

It is a sharp pivot from the Biden-era “deny advanced compute” posture, and it is not happening in a vacuum. The US has been rewriting the rules around advanced accelerators for years. If you need the broader context, start with our ongoing coverage in BonTech Labs News and the backgrounders we continually update in Guides.

Reuters reports that the Commerce Department has referred Nvidia’s licence applications to an interagency process involving State, Energy, and Defense, with a 30-day window for input, and that the final decision will rest with President Trump. The same reporting says Trump has already signalled he wants the sales to happen, as long as the US government takes a 25% cut.

H200 is “not the newest”, not “not useful.”

The H200 sits in Nvidia’s Hopper generation, below the company’s newest Blackwell push, but it is still a high-end datacenter accelerator built for serious training and inference. In other words, it is not a loophole chip designed to be barely legal. It is the kind of part that moves deployment timelines and capability planning.

This is where policy talk gets sloppy. One generation behind does not mean one generation irrelevant. Memory bandwidth, capacity, and software maturity are the difference between “we can run a model” and “we can run it at scale without the cluster falling over.” If the US is worried about capability transfer, H200-class silicon is still very much in the conversation.

The 25% fee changes what export controls are

Biden-era rules tried to look like an engineering boundary. Set a threshold, enforce it, and require vendors to redesign to ensure compliance. A 25% fee makes it look like a negotiated toll. That does not just change the politics; it changes the signal sent to everyone building long-term roadmaps around US policy.

If the argument is national security, monetising the risk is awkward. If the argument is industrial strategy, it reads like a government trying to turn licensing power into revenue while calling it restraint. Either way, “controlled access” stops sounding like a hard line and starts sounding like a price list.

Nvidia sees demand, but supply is not infinite.

Reuters reported earlier this month that Nvidia has seen unexpectedly strong demand from Chinese firms for H20, and has considered increasing output. It also noted that Chinese buyers include large platform companies, which should come as no surprise. China has money, it has data centres, and it has a direct incentive to acquire capability while it can.

But even with political permission, Nvidia still has to feed the supply chain. Advanced packaging capacity is constrained, HBM supply is constrained, and Nvidia is trying to meet demand for newer platforms simultaneously. So the realistic “yes” scenario is not a flood. It is a managed flow, with everyone fighting over allocations.

The other half of the problem: China still has a say

Even if Washington signs off, the export does not automatically happen. Reuters has reported that Chinese authorities have been weighing how imports of foreign accelerators intersect with domestic chip goals. That can translate into limits, approvals, or procurement strings designed to avoid looking dependent while still getting the compute.

The deal is messy on both sides. The US wants leverage and optics. China wants capability and optics. Nvidia wants volume and predictability. Nobody gets all three.

Expect backlash, because the trade-off is real.

Reuters noted pushback from national security officials and political figures who argue that H200-class computing can accelerate military capabilities and commercial AI. That objection is not going away, because it is grounded in a basic reality: compute is strategic now, even when it is sold as a SKU.

And the fee angle makes the criticism easier to sharpen. If the chips matter enough to restrict, they matter enough to question why the government is charging for access rather than blocking them. That is the core tension, and it will not be solved by calling the process “review.”

What to watch next

Licence conditions. “Approved customers” can mean a narrow list, strict monitoring, and heavy reporting, or it can mean something looser that looks like a policy climbdown.
Whether the 25% fee sets a precedent and, if this expands beyond Nvidia, export licensing will begin to resemble a tariff system for strategic hardware.
Allocation reality. If H200 volumes rise, something else gets deprioritised. Watch what Nvidia chooses to starve, if anything.
The next rules rewrite. The line has moved repeatedly. If H200 gets waved through, the next fight is where Blackwell-class parts land.

Export controls were sold as a barrier. This looks more like a dial, and it may now include a coin slot. That is not a subtle change, even if the paperwork still says “licence.”