Source: Z.ai completed construction of a 1GW data center that houses only Chinese chips; Z.ai built or operates several computing clusters, each with 10K+ chips
Z.AI has completed construction of a giant data center that houses only Chinese-made chips, a big step in Beijing's efforts …
Bloomberg
Context & Ripple Effects
The reported facility extends a visible buildout of domestic-chip AI infrastructure: Alibaba and China Telecom previously launched a southern China site powered by 10,000 Alibaba Zhenwu chips, while Alibaba was reported to have shipped more than 100,000 Zhenwu 810E units.
It also gives physical scale to prior policy direction. Guidance reported in 2025 called for state-funded new data-center projects to use domestic AI chips, making Z.ai's all-Chinese-chip deployment a consequential implementation case rather than an isolated cluster.
First-order effects
Z.ai obtains a large, dedicated footprint for workloads on Chinese-made chips, alongside its reported clusters of more than 10,000 chips.
Chinese chip suppliers gain a high-profile deployment environment in which their hardware must support data-center-scale AI training and inference operations.
Second-order effects
Other Chinese cloud and data-center operators face stronger pressure to turn domestic-chip procurement into deployable capacity, not simply smaller pilot installations; the Alibaba-China Telecom deployment provides another reference point.
Scaling all-domestic facilities shifts attention to the surrounding stack—servers, networking, cooling, power delivery, and software compatibility—where bottlenecks can determine how much usable compute the installed chips deliver.
Third-order effects
If such deployments recur, China’s AI-infrastructure market could become more vertically localized, with chip selection increasingly tied to policy eligibility and domestic supply-chain availability.
The pattern makes AI compute a long-duration infrastructure commitment: the competitive question shifts from access to individual accelerators toward who can finance, power, and operate large domestic-capacity fleets efficiently.
The trend: This is one data point in the shift from domestic AI-chip substitution at the component level to sovereign, utility-scale computing infrastructure.
Zhipu, creators of GLM, just brought a massive data center online They officially started operating a 1GW data center containing ONLY Chinese-made chips to train next-gen GLM models without the need for Nvidia chips ig they will join the very big OSS model club soon [image]
Holy: China now has a 1-gigawatt AI data center built without a single Nvidia chip! Via Bloomberg: zAI has reportedly completed construction of the facility and begun partial operations. It runs exclusively on Chinese-made chips and will support the development of the company's […
No, China does not have a 1-gigawatt AI data center. Some sanity on this... 1. The actual claim is “several clusters of 10,000+ chips”... that's ~75-100 MW of actual load. The facility is billed at 1 GW. It probably has 1GW of energy, but it certainly isn't powering 1GW of
Xi should be grateful that the Council on Foreign Relations has such China hands as Chris So many errors. more like 680K, 950DT is better than 910C in everything except fp16 FLOPS (also irrelevant), we know of 60K and 100K clusters with domestic chips and ofc datacenter!=chips [i…
actual physical independence would be the franz ferdinand moment for taiwan invasion china would emerge as intelligence leader w its own chip supply chain, and US risks losing the next century
Always take a large grain of salt 🧂 whenever journalists were unable to personally visit the site and confirm it through “this person said.” 🧂🧂🧂 We've been through too many intentionally misleading Chinese AI stories in the past 2.5 years that this should be part of
@ChrisRMcGuire >A 1GW data center would be the biggest data center in the world. The largest functioning data center in the world by power draw is xAI's Colossus II One note since this can be confusing, this is by IT power. If we account for PUE, Colossus II and the Amazon-Anthro…
Those export bans sure helped to keep AI acceleration technology out of China's hands, didn't they? Do people really think banning open-weight models will be any easier? Hint: It won't. They're 1s and 0s. Even harder to restrict access.
Amazingly, when I crunched what it takes to serve Kimi K3, I found not only Huawei's next-gen Ascend 950DT on track to beat the Nvidia B300 on the all-important tokens-per-watt metric, but also the current CloudMatrix 384 generation already at parity at the 50 tok/s speeds API [i…
Some back of the envelope math as to why I think reports that https://z.ai/ is running a 1GW data center using only Chinese chips are impossible: 1. China doesn't make enough chips. A 1 GW facility running Huawei's CloudMatrix systems needs ~500K Ascend 910Cs.
@kimmonismus And remember that unlike most other countries, China is not energy constrained and actively building out nuclear power plants that can fuel all these datacenters. After being the factory of the world, I'm convinced this will allow them to transform into the token fac…
This reminds me of something. A few days ago, I read the transcript of an expert call hosted by DB. It said that shipments of homegrown Chinese chips had already reached 2.5 million units this year. I scoffed at the figure, thinking there was no way it could be true—but now I'm […
@suchenzang Yeah, this was my big concern with that. Ultimately it was an attempt to lock down the field and bypass the need to compete internationally. That was never going to work. It's hard to lock out competition. It's especially hard to lock out competition in an emergent fi…
@ChrisRMcGuire Dude, when have you ever been right about anything related to Chinese semiconductors? You gambled with America's future and found enough politicians dumb enough to buy into your export control policies and it backfired. And this is the first of many more cases
@suchenzang 1) Jensen was just advocating for NVIDIA's and his self-interest. They make more money if they have more customers for their chips. 2) China was building their own stack either way. Self-reliance has long been in their five year plans.