On July 20, 2026, Z.AI — the Chinese lab formerly known as Zhipu — began partially operating a 1-gigawatt data center it plans to fill exclusively with Chinese AI chips, with no Nvidia silicon in the design. It is the largest concrete proof yet that China's frontier labs can train and serve cutting-edge models on domestic accelerators, and it is the facility that will host future GLM platforms.
One gigawatt is not a marketing number. It is roughly the output of a large nuclear reactor, and comparable to the electrical demand of about a million US households. Building that on domestic silicon, in a country under export controls designed specifically to prevent it, is the kind of milestone that gets studied for a decade.
This piece covers what Z.AI actually built, which chips realistically fill it, how the domestic hardware compares on the numbers that matter, and the part almost everyone gets wrong — why this is a power story at least as much as a chip story.
Key Takeaways
- Z.AI began partial operation of a 1-gigawatt AI data center on July 20, 2026, designed to run only Chinese-made chips.
- Nvidia's share of the Chinese AI chip market fell from roughly 95% to about 55%, with export controls accelerating rather than slowing domestic adoption.
- Huawei launched the Ascend 950PR on March 20, 2026, claiming roughly 2.87x the FP4 compute of Nvidia's H20 — at about 600W, some 200W more per chip.
- Huawei plans around 600,000 Ascend 910C chips in 2026 and as many as 1.6 million Ascend dies across the line.
- Five AI data centers are expected to reach 1 GW or more in 2026, each run by a different hyperscaler — Z.AI's is the first built on a fully domestic stack.
What did Z.AI actually build?
Z.AI completed construction of a data center campus designed for a 1-gigawatt power envelope and has begun operating part of it, with the stated intention of populating it entirely with Chinese-made accelerators rather than restricted Nvidia parts. The facility is meant to support development and deployment of the company's GLM model family — the same line whose GLM-5.2 scores 81.0 on Terminal-Bench 2.1, putting it within a few points of frontier Western models.
Two details are easy to miss. First, "completed" and "fully operational" are different states: Bloomberg's report describes partial operation, which is normal for a build of this size and also means capacity claims should be read as design targets, not current throughput. Second, a 1 GW design envelope does not mean 1 GW of chips are installed today; it means the grid connection, cooling and power distribution were engineered for it.
That distinction is worth holding onto, because it is exactly the distinction that gets flattened in most coverage of Western gigawatt-scale projects too.
Which Chinese AI chips actually fill a facility like this?
Realistically, Huawei's Ascend line does most of the work, supplemented by Cambricon, Alibaba's T-Head and Baidu's Kunlunxin parts. Huawei is the only domestic vendor currently shipping at the volume a gigawatt campus requires.
The flagship is the Ascend 950PR, launched March 20, 2026 as the compute core of the Atlas 350 accelerator card. Reported figures put it at roughly 2.87x the FP4 compute of Nvidia's H20 — the export-compliant part China was allowed to buy — while drawing about 600W, some 200W more per chip. Coverage of the launch also notes a CUDA-compatible software stack intended to lower migration cost, plus reported ByteDance commitments of around $5.6 billion.
| Metric | Huawei Ascend 950PR | Nvidia H20 (export part) |
|---|---|---|
| Launch | March 20, 2026 | 2024 |
| FP4 compute (reported) | ~2.87x H20 | baseline |
| Power draw | ~600W | ~400W |
| Software stack | CUDA-compatible layer | CUDA |
| Availability in China | Domestic, unrestricted | Export-controlled |
The power column is the whole argument in miniature. Domestic chips are competitive on throughput and worse on efficiency — which is survivable if, and only if, you have abundant power. China does. That is not an accident of geography; it is two decades of grid investment finally paying an unexpected dividend.
Has China closed the gap on AI chips?
Not on efficiency, and largely yes on availability. Nvidia's share of the Chinese AI chip market has fallen from roughly 95% to about 55%, and the driver was export controls — restricting supply created a guaranteed domestic market that Huawei and its peers could invest against with no fear of being undercut.
Huawei reportedly shipped around 812,000 Ascend chips, about 20% of the overall market in the relevant period, and plans roughly 600,000 910C chips in 2026, scaling the whole Ascend line toward as many as 1.6 million dies. Those are not prototype numbers. They are volume-manufacturing numbers.
The strategic read is uncomfortable for anyone who expected controls to hold the line: restricting a component in a country with capital, engineers and power does not prevent the component from existing. It relocates who builds it. We saw the mirror image of this dynamic in the custom AI chip revolution, where Western hyperscalers built in-house silicon for cost reasons rather than political ones — same outcome, different motive.
Why 1 gigawatt is the number that matters
The industry has quietly changed its unit of measurement. Two years ago, AI capacity was discussed in GPU counts. In 2026 it is discussed in gigawatts, because power is the binding constraint everywhere.
Five AI data centers are expected to reach 1 GW or more during 2026, each operated by a different hyperscaler, and US grid operators are openly struggling with the interconnection queue that implies. In the United States, the wait for a large new grid connection is now measured in years — often longer than it takes to design and fabricate the chips that would fill it.
This is where the Z.AI facility becomes genuinely strategic rather than symbolic. If your chips are 30% less efficient but you can energize a gigawatt campus in a fraction of the time your competitor needs for permits and transmission upgrades, the efficiency gap stops being decisive. Compute delivered per year, not compute per watt, is the metric that determines who trains the next model.
The angle everyone missed: efficiency only matters when power is scarce
Western analysis of Chinese accelerators almost always ends at performance-per-watt, concludes "still behind," and stops. That conclusion imports an assumption from a market where power is the scarce input.
Flip the assumption and the analysis inverts. In a system with surplus generation capacity and fast permitting, a chip that burns 50% more power for 90% of the performance is not a compromise — it is a rational trade, because the constrained resource is fab capacity and delivery time, not electricity. Z.AI's facility is what that trade looks like at scale: accept worse efficiency, buy domestic supply certainty, and win on schedule.
That is also why the Nvidia SK Hynix HBM4 agreement and this data center are two halves of the same story. One party is spending enormously to guarantee access to the most efficient memory on earth; the other is designing around the need for it. Read our breakdown of the $500B Nvidia–SK Hynix HBM4 deal for the other side of the board.
For developers, the practical consequence is already visible: Chinese models are cheap, capable, and increasingly hosted on infrastructure with no Western dependency at all. That is the supply chain behind the traffic shift we documented in Chinese AI models now powering up to 46% of enterprise API traffic, and behind the pricing pressure from releases like Kimi K3.
What this changes for teams outside China
Almost nothing technically, and quite a lot commercially. Three concrete effects:
- Price floors keep falling. A lab that owns its own gigawatt campus and its own accelerators has a cost structure that does not include a hyperscaler's margin or Nvidia's. That structure is what makes sub-$0.20 per million input tokens sustainable rather than promotional — see DeepSeek V4 Flash 0731 at $0.14/M for the current benchmark of that pressure.
- Open weights keep arriving. Domestic infrastructure reduces the strategic cost of releasing weights publicly, because distribution no longer depends on Western cloud goodwill. Expect the open-weight frontier to keep being set by Chinese labs.
- Procurement, not capability, becomes the deciding question. If your organisation cannot send data to a Chinese-hosted API, the score is irrelevant. The workaround is self-hosting open weights on your own hardware, which shifts the conversation to inference infrastructure rather than vendor selection.
If you land in that third bucket, the practical path is running the weights yourself. Our guides to running LLMs locally and vLLM vs Ollama cover the serving side; the short version is that a quantized mid-size model on your own GPUs is now genuinely competitive with a frontier API for most production tasks, and it sidesteps the residency question entirely.
Frequently Asked Questions
What is Z.AI's 1GW data center? It is an AI data center campus in China, completed and partially operating as of July 20, 2026, engineered for a 1-gigawatt power envelope and designed to run exclusively on Chinese-made AI chips. It will support development and deployment of Z.AI's GLM model family.
Which Chinese AI chips replace Nvidia GPUs? Huawei's Ascend line is the primary alternative, led by the Ascend 950PR launched in March 2026, alongside parts from Cambricon, Alibaba's T-Head and Baidu's Kunlunxin. Huawei plans roughly 600,000 Ascend 910C chips in 2026 and up to 1.6 million Ascend dies overall.
Are Chinese AI chips as good as Nvidia's? On raw throughput they are competitive — the Ascend 950PR reportedly delivers about 2.87x the FP4 compute of Nvidia's export-restricted H20. On energy efficiency they lag, drawing around 600W versus roughly 400W. Software maturity remains the larger practical gap.
How much power does a 1-gigawatt data center use? Approximately the output of a large nuclear power plant, or the electrical demand of roughly one million US households. Five AI data centers are expected to reach that scale during 2026.
Did US export controls fail? They succeeded at restricting Nvidia sales and failed at preventing domestic capability. Nvidia's China market share fell from about 95% to 55%, but the vacuum was filled by Huawei rather than left empty, accelerating a domestic industry that previously lacked a guaranteed customer base.
Can I use models trained on Chinese chips? Yes — models like GLM and DeepSeek are available through their APIs and, in several cases, as open weights. The hardware they were trained on does not affect usability. Data residency and procurement policy, not silicon, are what typically block enterprise adoption.
The verdict
Z.AI's 1-gigawatt facility is the clearest evidence to date that the AI hardware world has split into two self-sufficient stacks rather than one global one. The domestic chips are less efficient and the software is less mature, and neither of those facts prevents the machine from training frontier-class models.
Our verdict: stop scoring this race on performance-per-watt. The metric that decides 2027 is delivered compute per calendar year, and on that measure a country that can energize gigawatt campuses quickly is competitive with a country that has better silicon and a five-year interconnection queue. Watch grid connections, not benchmarks.
For the other half of this story — what happens when a Western vendor spends half a trillion dollars to lock up the efficient path instead — read our analysis of the Nvidia SK Hynix HBM4 deal.
The export controls were meant to buy time. What they actually bought was a competitor with its own fabs, its own chips, and a gigawatt of power to run them.
