Z.ai built a gigawatt of AI compute with no Nvidia inside

Bloomberg reported on July 20 that Z.ai — the Chinese lab that used to call itself Zhipu — has switched on part of a data center drawing roughly a gigawatt of power that runs, by its account, on Chinese-made accelerators alone. No Nvidia anywhere in the racks. Several clusters, each holding more than 10,000 chips. For two years the working assumption in Washington and on Wall Street was that U.S. export controls would keep Chinese labs a training-generation behind, because no blacklisted company could ever assemble enough advanced silicon in one building. A gigawatt of domestic chips, powered up and pointed at a model, is the first real dent in that assumption. And it lands on the exact company the United States blacklisted to stop it.

What Bloomberg reported, and what it deliberately left open

Treat the specifics as reporting, not as a press release. Bloomberg says Z.ai is partially operating a site of about one gigawatt, filled with several clusters of 10,000-plus chips, all of them Chinese-made, to train its GLM family of models. A gigawatt is a useful yardstick on its own: it is roughly the draw of 750,000 homes, which puts this among the largest compute sites any Chinese lab has stood up. The same scale question is now reshaping electricity markets on the U.S. side, where regulators are fighting over who pays for AI’s power appetite.

Now the gaps, because they matter. Partial operation is not a finished frontier training run — a site can be energized and still be months from a full end-to-end pretraining pass. Bloomberg did not name the chip vendor; the domestic field is led by Huawei’s Ascend line, Cambricon, and Alibaba’s in-house accelerators, and any of them could be inside. And Z.ai itself has not published an official spec sheet for the facility. So the honest framing is a strong, well-sourced report of a gigawatt-class domestic build — not an audited disclosure. Anyone who tells you China has “matched” U.S. training capacity off this single story is reading past the caveats.

Scale is also relative. A gigawatt is enormous for a Chinese lab operating without Nvidia, but American hyperscalers are now announcing single campuses measured in multiple gigawatts, with power delivered in phases over years. So the right read is not “China has out-built the U.S.” It is that the compute floor for a sanctioned Chinese lab has moved from “can it get chips at all” to “can it run a gigawatt of its own.” That is a different, and much harder, question for export policy to answer.

An Entity-List company that had no choice but to build its own stack

Here is the part that should stick. Z.ai has been on the U.S. Commerce Department’s Entity List since January 2025, added on the way out of the previous administration for what the government described as helping advance Chinese military modernization through AI. That designation cut the company off from legal access to Nvidia silicon. It did not slow Z.ai down so much as redirect it — straight into a fully domestic hardware-and-software pipeline, earlier and more completely than peers who were still buying Nvidia where they could.

The software half of that pipeline is the part outsiders tend to skip. Running a training job of this size is not just a matter of owning chips; it means a mature compiler, collective-communication libraries, and a framework that can keep tens of thousands of accelerators in sync without Nvidia’s CUDA ecosystem underneath. Z.ai’s reported answer is Huawei’s Ascend accelerators paired with Chinese-built training frameworks — the kind of full-stack substitution that takes years and only happens when a company is forced to commit to it. Export controls supplied exactly that forcing function.

And the proof already shipped. In June, Z.ai released GLM-5.2, a 744-billion-parameter mixture-of-experts model with roughly 40 billion active parameters per token and a usable one-million-token context window. The weights went up on Hugging Face under an MIT license — free to download, self-host, and use commercially. The company says the model was trained end to end on Huawei Ascend accelerators, with no Nvidia hardware at any stage. On the independent Artificial Analysis Intelligence Index it scored 51, the highest of any open-weights model to date, matching frontier Western systems on several long-horizon coding tasks. So the data-center report is not arriving out of nowhere. The model a domestic gigawatt would train has already been demonstrated at smaller scale, on the exact stack the export controls tried to deny.

That is the strategic context builders keep underrating. The best Chinese models teams actually deploy — GLM, plus the DeepSeek line and Moonshot’s open-weight Kimi releases — increasingly ship as open weights trained on hardware Washington cannot embargo. Deny the chip, and you accelerate the substitute.

The export-control scoreboard just moved a square

Line the two tracks up side by side and the irony is hard to miss. On the sanctioned track, the U.S. spent late 2025 and early 2026 loosening the rules — allowing Nvidia’s H200 into China under a 25% revenue tax paid to the Treasury. Yet as of mid-July, a senior Commerce official described the volume actually shipped as “trivial”, with the deal stuck in limbo between U.S. licensing and Beijing’s own procurement rules. On the domestic track over the same stretch, a blacklisted lab lit up a gigawatt of homegrown compute. The controlled channel delivered almost nothing; the workaround delivered a data center.

The demonstration effect is the part that travels furthest. Z.ai is not the only Chinese lab under U.S. restrictions, and it is now the clearest proof that a domestic gigawatt is buildable rather than aspirational. Every other sanctioned or hedging lab in the country can point at it to justify the same capital and the same all-domestic roadmap. Controls work best when the workaround looks expensive and uncertain; once someone has visibly walked the path, the deterrent value drops for everyone behind them. That is a slower, compounding cost of the export regime that quarterly shipment figures do not capture.

None of this is unique to China. The U.S. hyperscalers are running their own campaign to reduce Nvidia dependence — see Amazon’s move to turn custom-silicon capacity into a supply contract and AMD and Cerebras splitting inference across non-Nvidia parts. The difference is motive and speed. American firms are diversifying to cut costs and hedge a single supplier. Z.ai is doing it under duress, on a national-policy clock, because the alternative is not shipping at all. That pressure tends to produce faster, uglier, more committed engineering than a comfortable hedge does.

What this actually changes for you

If you build products on open Chinese models, the supply story just got more durable. GLM and its peers will keep shipping regardless of whether Nvidia silicon ever clears Chinese customs, because their training no longer depends on it. That is a genuine reliability gain — your model roadmap stops being hostage to an export negotiation. But the compliance ledger gets heavier in the same motion. You are now leaning on an Entity-List vendor’s stack, which raises real questions about license durability, data residency, and whether your own organization’s procurement or export policy even permits it. Treat “open weights from China” as a supply-chain and governance decision, not just a line on a cost comparison.

If you hold AI-hardware exposure, the clean thesis that China physically cannot train frontier models without American chips now carries an asterisk. It is not falsified. Parity between domestic-silicon checkpoints and Nvidia-trained peers still has to hold up on independent benchmarks over full training runs, not demos. Domestic fabrication yield — how many good Ascend-class chips China can actually produce — is a real ceiling on how fast Z.ai scales past one gigawatt. Packaging and memory supply are others. But the direction of travel is no longer in doubt, and pricing a hard technological moat that is visibly eroding is its own kind of risk.

There is a security dimension underneath all of this that will not stay theoretical for long. A compute base that no longer depends on foreign hardware is also a compute base that is harder for outsiders to observe, meter, or throttle. Export controls were never only about slowing capability; they were a visibility mechanism, a way to keep a hand on the tap. A fully domestic gigawatt removes that hand. For enterprises, the practical version of the same problem is smaller but real: if a model you depend on is trained and served entirely inside a jurisdiction you have no leverage over, your continuity plan cannot assume you will get warning before access, licensing, or terms change.

What to watch from here is concrete: an official confirmation from Z.ai with a named chip vendor; evidence that a full frontier run, not just partial operation, completes on the domestic stack; the leaderboard placement of the next GLM trained at gigawatt scale; and the power, yield, and packaging constraints that will decide whether one gigawatt becomes five or stays a showcase. Those are the numbers that turn a striking report into a settled fact.