The long view is the only honest view. Seen from enough altitude, a supply chain is just a sequence of decisions that hardened into habit — and the habit in question here is coming apart in public, documented quarter by quarter. In its 2026 annual risk filing, NVIDIA says something a company rarely volunteers: under current rules, it cannot build a data-center product that wins approval from both the United States and China at once, and it is “in effect excluded” from China’s data-center computing market. That sentence, buried among the risk factors, is a star chart of an entire era ending.
The exclusion, in the company’s own words
Read the sequence like a light curve, because it has the shape of one. In April 2025, the H20 accelerator needed a license to ship. By August 2025, some licenses were being granted. By February 2026, small-case approvals for the H200 began trickling through. And now the 10-K states the end state plainly: no product can satisfy both governments at the same time. That is not a pause. A pause is a temporary dimming, a cycle that returns. This is a spectrum reading showing the line has moved and will not move back on its own.
Let me think about what “in effect excluded” actually does to a market, because the phrase is doing quiet work. It does not say the company is banned. It says the company’s product cannot legally exist in the shape the market wants it in. A data-center accelerator that cannot be sold without a case-by-case license is not a product; it is a petition. Customers do not design multi-year infrastructure around a petition. They design around whatever they can reliably order — which, increasingly, is not from this vendor.
I will be careful here, because it would be easy to overstate. NVIDIA’s filing describes the constraint; it does not describe the commercial counter-moves, and companies have a history of finding creative routes around written rules. But the document is the signal that matters, because it is the vendor itself conceding that the engineering problem has no current solution. When the seller publishes the verdict, the negotiation is already over.
The replacement, showing up in shipments
Meanwhile, on the other side of the data, the numbers have started to look beautiful in their own right. The National Development and Reform Commission reported on August 28 that China’s integrated-circuit exports reached 1.49 trillion yuan in the first seven months of 2026, up 91.6 percent year on year. That is a striking figure, and I want to handle it honestly rather than cheerfully. Exports are not the same as domestic deployment. A 91.6 percent rise in exports says a great deal about global demand for Chinese-made chips and about manufacturing muscle, but it does not by itself prove that domestic data centers are being re-equipped.
The domestic story has its own receipts, and they are the numbers that keep me awake. IDC counts about 4 million AI accelerators shipped in China in 2025, of which roughly 1.65 million — 41 percent — came from domestic manufacturers. Two years ago, that share would have been a rounding error and a talking point. Now it is a line in a market report. I keep coming back to that 41 percent; it is the number that separates the policy narrative from the procurement reality. A share that large does not ride on sentiment. It rides on purchase orders, and purchase orders are made by engineers who need something that actually runs.
Let me correct my own instinct here, because I made an assumption on first reading that the numbers did not support. I wanted to read the 91.6 percent export jump as proof that domestic data centers are being re-equipped — but exports measure products leaving the country, not machines being installed at home. Those are two different columns of the ledger, and conflating them would have made the story cleaner than it is. The honest version is that the two columns point in the same direction, which is enough, and I should not have needed to sharpen it beyond the facts.
There is a companion figure I do not want to leave on the table. The Commission’s August briefing also noted that major domestic wafer foundries ran capacity utilization above 90 percent in the first half of the year. That is a real number, but let me state precisely what it does and does not prove. It proves the fabs are full — that demand is meeting the installed production base, and that the export line and the domestic line are both pulling on the same capacity. It does not prove that the chips coming off those lines match the leading edge on every benchmark. I am not going to claim what the data does not say. The data says the factories are busy; that is remarkable enough on its own.
The policy layer, no longer decorative
And then there is the policy layer, which the evidence says is no longer ornamental. The Security and Reliability Review announcement for 2026, second batch, for the first time brought AI training and inference chips into the national reliability review, with nine domestic AI chips from seven companies earning the top Grade I rating. What does that mean in plain terms? It means a shelf has been built with a label on it, and only certain chips are allowed on the shelf. In a market where the state is the largest customer of its own infrastructure, the label is not advisory. It is a specification — a procurement door that quietly opens and quietly closes.
I want to be honest about the limits of this reading, too. A reliability rating certifies that a chip passed a defined review; it does not measure whether the chip out-performs a foreign rival on cost or speed. But the question was never whether the label would appear on every benchmark. The question was whether the market could be steered, and the combination of a 41 percent domestic share with a formal rating mechanism answers that question with a yes that is documented rather than asserted.
Two supply chains, diverging
Put the two sides together and the picture is hard to avoid. On one side, an American company formally writes itself out of a market in its own regulatory filing — not under editorial pressure, but as a statement of engineering reality. On the other side, a market is being supplied by domestic silicon at 41 percent, with an export line growing at 91.6 percent and a policy shelf that now has names on it. Those are not three separate facts. They are one fact viewed through three instruments, and all three point in the same direction.
This is what the long view settles. For years the conventional reading was that export controls would slow China’s advance while the gap remained bridgeable — a temporary brake on a converging road. What the documents now suggest is different. The two compute ecosystems are not just drifting apart in speed; they are being organized on different premises. One is optimizing for whatever the export license allows. The other is being built, deliberately and with policy support, to run without that permission at all. Deep time has a way of settling arguments, and it is settling this one by letting both sides proceed as if the other did not exist.
Let me walk through what that means for a reader who just wants the signal, not the cosmology. First, every company building data-center infrastructure in China now assumes a domestic supply chain as the default, with foreign chips as a possible exception rather than the other way around. Second, the foreign vendor’s own filing removes the ambiguity that used to justify waiting — there is no longer a realistic near-term scenario in which the old product line returns at scale. Third, the policy instruments have moved from encouragement to qualification: the market does not merely prefer domestic chips now; it has a formal mechanism for saying which ones are acceptable.
There is a wider angle worth noting, and it concerns everyone watching from outside the two countries. The reordering of the China market does not shrink the global demand for compute; it reroutes it. The same hyperscale build-out that used to run on a single, dominant accelerator family now runs on two separate pipelines, each with its own design rules, its own software stacks and its own suppliers. That is not a zero-sum story for the industry as a whole — it is a duplication story. Two ecosystems get built instead of one, which is precisely why the numbers on both sides can look so strong at the same time. The 91.6 percent export growth and the 4-million-unit domestic market are not competing claims; they are two halves of the same re-plumbing.
I also want to say something about time, because it is the dimension that most forecasts get wrong. The interesting question was never whether domestic chips could close the gap this quarter; it was whether the gap behaves like a ladder or like a wall. A ladder can be climbed and the control policy becomes a speed bump. A wall forces an entirely different route. The documents from this period — the 10-K concession, the grade-one ratings, the 41 percent share — all read less like speed bumps and more like the early masonry of a wall being built from both directions. How thick it gets, and whether it eventually acquires gates, is the story of the next several years, not of this quarter.
I find the phrase “measured wonder” keeps fitting this story in an unexpected way. Wonder, because the scale of the reordering is genuinely breathtaking — 4 million accelerator units in a year, a 1.49 trillion yuan export line, a supply chain being re-plumbed in under two years. Measurement, because every one of those claims sits on a document I can point to: a risk factor in a 10-K, a commission press briefing, an IDC market report. The awe is held in check by the footnotes, and that is exactly how it should be. Evidence keeps the awe honest.
Let me close with a concrete moment, the kind that stops me when I am reading these documents at night. It is a small thing: the February 2026 decision to allow small-case approvals of the H200. Not a reversal, not an opening — a trickle, administered case by case. Somewhere in that process, an engineer’s request crossed a desk, a license officer compared it against a rulebook that was itself still being written, and a stamp came down on a single shipment while the broader question stayed unanswered. That is the moment the split stopped being a policy and started being a practice. The rest, from the 10-K sentence to the 41 percent, has just been the paperwork catching up to that moment.
So here is the verdict, stated plainly. The split of the global compute supply chain is no longer a trend line you can argue about; it is an accounting fact, documented in a risk factor, a market report and a policy bulletin within a single quarter. What matters next is what both sides build with their half of the sky, and how the two halves re-connect or fail to re-connect over the years ahead. On clear nights, data like this comes back beautiful — and what is beautiful about it is that it finally lets you measure, rather than guess, where the two worlds diverge.