Infrastructure

A Gigawatt Of AI Compute Now Runs Without Nvidia

Z.ai has finished a gigawatt-class AI data centre running entirely on Chinese-made chips, and bought the software layer that makes one model run across them. The second part is the news.

InfrastructureBy Servola Tech Desk2026-07-215 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

A Gigawatt Of AI Compute Now Runs Without Nvidia

Key takeaways

Z.ai, the Chinese lab behind the open-weight GLM models, has completed a gigawatt-class data centre built entirely on domestically made chips and has begun partial operation, the first reported case of a large Chinese firm doing so without any Nvidia silicon.
On the same day Z.ai closed its acquisition of Zhongke Jiahe, whose software runs one model efficiently across different Chinese accelerators including Huawei and Cambricon parts, which is the harder half of the problem and the part it chose to buy rather than build.
A gigawatt is roughly the draw of 750,000 homes, and because domestic accelerators deliver less compute per watt than the restricted Nvidia parts, this is a decision to substitute energy for efficiency, an option China has and Europe does not.
If an open-weight GLM model already sits in your stack, its supply no longer depends on US export licences, but it now sits inside a second export-control perimeter that Beijing is actively consulting industry about.
Mirror the weights of any Chinese open-weight model you run in production to storage you control this quarter, and record in your architecture notes which export-control regime governs each model you depend on.

Ten thousand chips, none of them American

Z.ai, the Chinese laboratory previously known as Zhipu and the developer of the GLM model family, has finished building a data centre in the gigawatt class and switched part of it on. The facility runs on more than ten thousand domestically manufactured semiconductors and, according to the reporting, not one Nvidia part. It is described as the largest server facility built by a Chinese AI company, and Z.ai now operates several clusters each exceeding ten thousand chips.

The claim that matters is the exclusivity, not the size. Chinese firms have been assembling large clusters for years, generally with a mixture of stockpiled Nvidia hardware, export-compliant variants and domestic accelerators. A build with no American silicon in it at all is a different statement. It says the domestic supply chain, from accelerator to interconnect to the software that drives them, is complete enough to stand up a frontier-scale training site without a licence from Washington.

Context sits behind it. China has committed roughly 2 trillion yuan, about 295 billion dollars or 273 billion euros, to national data-centre construction over five years. This facility is one instance of that programme reaching operational status rather than announcement status, and operational status is the only kind that changes anyone's options.

The acquisition that matters more than the building

On the same day, Z.ai completed the purchase of Zhongke Jiahe, an AI infrastructure software company founded by researchers from the Chinese Academy of Sciences. Its technology does something narrow and load-bearing: it runs the same model efficiently across different Chinese-made accelerators, including Huawei and Cambricon designs.

That is the actual bottleneck, and it is not a hardware problem. The reason a decade of Nvidia alternatives failed to displace anything was rarely raw silicon performance. It was that every workload had to be ported, tuned and maintained per architecture, and the porting cost exceeded the hardware saving. A domestic build using two or three different accelerator families multiplies that problem rather than escaping it.

Z.ai did not attempt to solve it in-house. It bought a team from the Academy of Sciences that had already solved it, and closed on the same day it announced the site. Read the sequencing as intent. A company confident it could paper over heterogeneous hardware with internal engineering does not acquire the capability at the moment it needs it. The purchase is an admission of how hard the layer is, and simultaneously the reason the gigawatt is usable at all.

A gigawatt buys less than it sounds like, and that is the point

One gigawatt is approximately the instantaneous draw of 750,000 homes. Stated that way it sounds decisive, and for a European reader it is worth translating into the currency that actually constrains us. It is not chips. It is grid connection.

Domestic Chinese accelerators generally deliver less compute per watt than the Nvidia parts they replace. So a gigawatt of Huawei and Cambricon silicon does not buy a gigawatt of Nvidia-equivalent training throughput. It buys meaningfully less. The strategy only works if you can put more power behind the same result, and Beijing can, because grid capacity and generation build-out are the inputs China has in surplus.

Europe is on the other side of that trade. Connection queues in the Netherlands, Germany, Ireland and the United Kingdom are measured in years, not months, and the binding constraint on European AI capacity has not been chips for some time. This is why the announcement should be read as an energy story wearing a semiconductor jacket. A competitor that can pay for inefficiency with electricity is playing a game a European operator cannot copy, whatever hardware becomes available.

What this does to a model already in your stack

GLM models are open-weight, and a considerable number of European firms are running them in production, often as the cost-controlled tier underneath a frontier model. For those firms the practical question has always been supply durability: if export controls tighten, does the thing I built on stop being maintained.

The answer just changed shape rather than improving. A domestic compute base makes continued development of these models far less dependent on American licensing decisions, which removes one failure mode. But the same self-sufficiency is precisely what makes the models worth controlling from the other end. Chinese regulators have been consulting domestic AI and chipmaking firms on tighter export rules covering advanced models and the technology behind them, a direction reported this week and not yet enacted.

So the risk did not disappear. It moved, and it doubled. A model that used to face one export-control regime now sits between two, and the two are not coordinated. For an architecture decision with a five-year horizon, that is a materially different picture from the one most European AI risk registers currently contain.

Mirror the weights, then write down which regime applies

Two actions, both cheap, both this quarter. First, for every open-weight Chinese model you run in production, mirror the weights to storage you control, in your own jurisdiction, with a recorded hash. Open weights you have downloaded are durable regardless of what any government decides later. Open weights you are pulling from a hosted repository on demand are a dependency wearing the costume of an asset.

Second, add one field to your model inventory: which export-control regime governs this. Most inventories record licence, parameter count and hosting location. Almost none record jurisdiction of control, because until recently there was only one plausible answer. There are now two, they can tighten independently, and a model can become unavailable through either without the other doing anything.

Nothing here calls for switching vendors or abandoning Chinese open weights. The cost advantage that put them in your stack is real and it is not going away, if anything a domestic compute base makes it more durable. The point is narrower. You are now depending on an artefact whose continued availability is a matter of policy in two capitals, and the mitigation costs a storage bucket and a spreadsheet column.

Frequently asked questions

Does this mean Chinese chips have caught up with Nvidia?

No, and the facility does not claim it. Domestic accelerators generally deliver less compute per watt than the Nvidia parts they substitute for, which is why the significant number here is a gigawatt of power rather than a performance figure. What has been demonstrated is sufficiency at scale without American silicon, not parity per chip.

Why is the Zhongke Jiahe acquisition described as more important than the data centre?

Because running one model efficiently across different accelerator architectures is the cost that has historically defeated Nvidia alternatives, not the raw hardware. A build combining Huawei and Cambricon parts multiplies that porting problem. Z.ai bought a team from the Chinese Academy of Sciences that had solved it, and closed on the same day it announced the site, which is what makes the gigawatt usable rather than merely large.

We run GLM in production. What should we change?

Mirror the weights to storage you control in your own jurisdiction with a recorded hash, so availability does not depend on a hosted repository. Then add a field to your model inventory recording which export-control regime governs each model. The supply risk has not disappeared with domestic compute, it has moved to Beijing and now exists alongside the American one rather than instead of it.

The temptation is to file this as a milestone in a rivalry that Europe watches rather than participates in. That misreads what was proved. The constraint that decided this build was not access to the best chips, it was the ability to put a gigawatt behind second-best ones. Europe has the opposite endowment and keeps planning as though the shortage were silicon. The firms that adjust first will be the ones that started treating grid connection, not procurement, as the thing standing between them and their own compute.

Z.aiGLMHuaweiCambriconExport controlsData centres

1.65 Trillion In AI Debt Never Reached A Balance Sheet

Five US tech giants carry 1.65 trillion dollars of AI-related contractual obligations that never appear as debt. The disclosure sits in a footnote, and it tells you what your next cloud renewal will cost.

6 min read

AI Policy

Three Months In, the US AI Testing Chief Resigned

Chris Fall is leaving CAISI after about three months. NIST director Arvind Raman takes the role in an acting capacity, and no reason has been given.

4 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles