China didn't reopen its market to Nvidia. It built the same permission machine Washington did.

The most popular reading of this week's news out of Beijing is that China has cracked its door back open to Nvidia. After months in which the country's own regulators steered its national champions toward domestic silicon, officials have reportedly told Alibaba, ByteDance and DeepSeek that they may, after all, buy the H200 — the second-fastest AI chip Nvidia sells. Access restored, the story goes. A thaw. It is an appealing reading, and it is a category error. What happened is not that a market reopened. It is that a second government finished building the same machine the first one built, and the two of them quietly agreed to run the world's most important supply chain by permission slip. I say this as someone who used to sign the permission slips.

Let me start with the popular position stated fairly, because it deserves to be. For most of a year, the conventional wisdom held that export control was a one-directional instrument — a lever in Washington's hand, pulled to slow China's access to frontier compute. Under that frame, anything that puts more Nvidia silicon into Chinese data centers is a loosening: good for Nvidia's revenue, good for the Chinese labs that want to train larger models, and a small de-escalation in a rivalry that badly needs some. If you believe the frame, this week is unambiguously a thaw. Access beats a ban. A rationed H200 trains a better model than no H200 at all.

Now the facts, because the frame doesn't survive them. According to reporting from The Information, subsequently relayed more widely, Beijing is not opening a market. It is issuing a quota. The total number of H200s it will wave through is expected to come in under 200,000 units — by the same reporting, less than half of what the three companies asked for. Each firm has to specify how many chips it needs and detail what it intends to do with them. And there is a condition on the workload itself: the H200s are to be used for training models on public data. Inference — the part where a trained model actually answers a query, the recurring, higher-volume, more lucrative half of the business — and anything touching sensitive customer data are to run on Huawei and other domestic chips. This is not a country letting a supplier back in. It is a country deciding, chip by chip and workload by workload, exactly how much of a foreign product its own firms may use, and reserving the profitable half of the market for its national champion by decree.

Two machines, mirror images

To see why this is the same machine, look at what Washington built first. In December, after weeks of deliberation, the U.S. government approved H200 sales to "approved customers in China" — and attached to that approval a set of conditions that tell you what the instrument had become. The government would take a reported 25 percent cut of each sale. The chips would pass through a third-party testing lab. Chinese buyers could not receive more than half of what American customers got, had to demonstrate "sufficient security procedures," and could not use the chips for military purposes. Read that list again and ask what kind of policy it describes. A security regime does not take a revenue share. A tariff does not test for military end-use. What Washington built was a hybrid — a permission list that is simultaneously a security control, a trade tax, and a leverage instrument, administered case by case at the government's discretion.

Beijing has now built the mirror image. Its instrument is an import-permission regime rather than an export one, but the architecture is identical: a discretionary list, a per-applicant justification, a numerical cap set wherever officials please, and — this is the part that matters most — a restriction not merely on whether the chip may be bought but on what it is allowed to compute. The two governments are not fighting over the H200. They are doing the same thing to it from opposite ends. Both have concluded that the interesting power is not in the binary of yes-or-no, but in the dial: how many, for whom, for which workloads, at what cut. The chip now passes through two capitals' discretion before it multiplies a single matrix.

The two governments are not fighting over the H200. They are doing the same thing to it from opposite ends.

Here is the strongest version of the counterargument, because I want to be fair to it before I take it apart. Perhaps none of this framing matters and the outcome is what counts: Chinese labs get some frontier training compute, Nvidia books some revenue, Huawei gets a protected runway to build a real inference business, and the whole arrangement lowers the temperature. Everybody, on this reading, gets enough of what they want. Managed access, the argument runs, is more stable than a hard embargo that invites smuggling and retaliation. And there is something to it — a use-case partition really is a more surgical tool than a blanket ban, and a government that reserves inference for its own industry while importing training compute is at least being coherent about its industrial goals.

The procedural question the frame skips

Now ask the question the outcome-focused reading skips, the one I learned to ask from the inside: who actually enforces this, and how? Because the answer determines whether you are looking at a policy or a mood. A published rule — this chip may be sold, that one may not, here is the line and here is why — is a thing companies can plan against, courts can test, and allies can either match or reject. What both governments have built instead is a permission granted applicant by applicant, workload by workload, with no bright line, no published criteria, and no appeal. A cap of "under 200,000, less than half what they asked" is not a number that fell out of a principle. It is a setting on a dial, and a dial is a thing you turn. The same discretion that grants the 200,000th chip can deny the 200,001st next quarter, for reasons that need never be written down.

This is the category error, named plainly: we are still calling this export control, a phrase that implies a security judgment about a technology, when what both countries are actually running is industrial allocation policy conducted through a permission list, with national security as the vocabulary and market share as the goal. The tell on the American side is the 25 percent cut — a security regime does not invoice. The tell on the Chinese side is the workload partition — a genuine security control would care whether the chip is in the country, not whether it is doing inference or training, because the physical capability is the same either way. Reserving inference for Huawei is not about keeping a dangerous tool out of the wrong hands. It is about keeping a profitable market in the right ones. That is a perfectly legitimate thing for a government to want. It is just not what the words on the label say.

Why does the mislabeling matter, if the industrial goals are legitimate? Because discretion dressed as security is the version of this that cannot be audited, cannot be appealed, and cannot be relied upon — and a global compute supply chain is now hostage to exactly that. Every hyperscaler build-out, every training run scheduled a year out, every lab's capacity plan now rests on two separate permission lists that either government can revise with a phone call and no paper trail. Firms will not fight this; they will lobby it, chip by chip, because that is what a discretionary regime trains its subjects to do. You do not petition a bright-line rule. You petition a person. Both governments have just made themselves the person, and both will discover — regulators always do — that the requests never stop, because a favor granted once becomes a favor expected forever.

Discretion dressed as security is the version that cannot be audited, cannot be appealed, and cannot be relied upon — and the world's compute supply chain is now hostage to exactly that.

The better frame, and what it costs

So here is what I would do instead, with the cost of my own position admitted up front. Both governments should say what they are actually doing. Washington should decide whether it is running a security regime or a revenue regime, because it cannot credibly run both — the 25 percent cut has already answered the question, and pretending otherwise corrodes the security case every time it is invoked. Beijing should state plainly that reserving inference for Huawei is industrial policy, not safety, and defend it on those terms, which are defensible. And to the extent either government is going to gate this technology, it should do so with published criteria and bright lines that firms and allies can actually plan against — a rule that holds still long enough to be relied on, rather than a dial that moves with the diplomatic weather.

The cost of that frame is real, and I will not pretend otherwise. Bright lines are less flexible than discretion. A published rule forgoes precisely the fine-grained leverage both capitals now enjoy — the ability to reward this firm, punish that one, extract a cut here, protect a champion there, all without ever having to explain themselves. Governments reach for the dial because the dial works; it let Washington turn a ban into a revenue stream and let Beijing turn an import into an industrial subsidy for Huawei, and each did it in a single stroke. I understand the appeal from the inside better than most. That is exactly why I distrust it. The instrument that lets you calibrate everything is the instrument that can be captured by anyone patient enough to lobby it, and it leaves the people downstream — the companies, the researchers, the allied governments trying to decide whether to match your policy — with nothing stable to build on.

The comforting story this week is that the H200 news is a thaw. The more accurate story is that it is a convergence. Two rival governments, starting from opposite premises, have arrived at the same tool: a discretionary permission list that decides not just whether the world's most important chip may be sold, but who profits from every workload it runs. That is not the market reopening. That is the market being quietly nationalized from both ends at once, in a language borrowed from security and paid for by everyone who has to plan around it. A dial is not a policy. It only looks like one until someone turns it.