Z.AI vs DeepSeek: The 1GW Data Center Showdown

Two Chinese AI labs. Two gigawatt-scale data center bets. One of them just finished building — and switched the lights on.

Z.AI, the Beijing company formerly known as Zhipu and best known for the GLM model family, has completed construction of a 1-gigawatt AI data center powered entirely by Chinese-made chips, with partial operations already underway. The news, first reported by Bloomberg, turns a planned build-out into a working facility — and it lands squarely in the middle of China's most interesting infrastructure rivalry: Z.AI versus DeepSeek.

Round 1: The Build — Who Actually Finished?

This is where the two projects stop being comparable on paper. DeepSeek's 1-gigawatt facility in Ulanqab, Inner Mongolia, remains a plan — a Bloomberg-reported ambition backed by a mix of new construction and leased capacity from other firms, hedged to spread cost and risk. Z.AI, by contrast, has a completed campus that is already partially operational.

The scale is hard to wrap your head around. One gigawatt is roughly the electricity draw of 750,000 homes at any given moment. Bloomberg's report describes a site built around several computing clusters, each containing more than 10,000 chips. That makes it one of the largest server hubs ever assembled by a Chinese AI lab.

Here is the scorecard so far:

Round 2: The Chips — Domestic Silicon vs the Nvidia Question

The most important detail is not the megawatts. It is the hardware — or rather, the absence of Nvidia hardware. Z.AI's facility runs exclusively on Chinese accelerators, a milestone in Beijing's campaign to replace restricted U.S. silicon in frontier AI development. The exact suppliers have not been named, but China's accelerator market has a short list of candidates: Huawei, Cambricon, Moore Threads, and Kunlunxin. Reuters previously reported that Zhipu's GLM systems already used domestic chips for inference, including Huawei Ascend parts and products from the other three vendors.

That history matters. Z.AI has been quietly validating Chinese silicon in production for inference; this facility is the leap from testing to training at scale. It is the clearest signal yet that Chinese labs are no longer treating domestic chips as a fallback — they are building frontier infrastructure around them.

DeepSeek's position is messier. The lab has been accused by U.S. officials of bypassing export controls to use Nvidia Blackwell hardware — accusations it denies, but which shadow every expansion move it makes. Its cost advantage is real: Chinese AI services now price at a tenth or less of their U.S. equivalents. But that advantage depends on access to hardware that export rules are actively trying to cut off.

Round 3: The Business — China's Anthropic vs the Cost War

The two labs are also betting on different business models. Z.AI is reportedly on track to reach $1 billion in annual recurring revenue — which would make it the first Chinese AI startup to hit that mark. It sells cloud access to its models and builds custom AI systems for enterprise customers, including state-owned firms. Analysts increasingly frame it as China's closest equivalent to an enterprise AI supplier like Anthropic: fewer consumer fireworks, more steady contracts.

DeepSeek is playing the cost war. Its models rattled global markets on price alone, and its reported IPO filing could land within the year — turning the frozen fields of Ulanqab into a pitch deck prop. Where Z.AI courts enterprise trust, DeepSeek courts scale and disruption.

The pressure from both directions is real. Beijing is preparing to spend roughly 2 trillion yuan — about $295 billion — on data centers over five years, with Alibaba and China Telecom among the biggest builders. Startups like Z.AI are racing to control their own compute rather than rent it, and the crunch is visible across the sector: rival Moonshot AI recently suspended new subscriptions to prioritize compute for existing members after its Kimi K3 launch.

The Verdict. If this were a boxing card, Z.AI wins on execution: it is the only contender standing in a completed, partially powered-up arena, with no export-control exposure in its chip supply chain. DeepSeek wins on raw ambition and price aggression — but its plan still lives partly on paper, and its hardware questions have not gone away.

The honest caveat: domestic chips still lag Nvidia in raw performance, software maturity, and cluster stability at the frontier. A gigawatt of Chinese silicon is not automatically a gigawatt of Nvidia-grade training throughput. The real test is what happens next — Z.AI's next GLM model, and whether a domestic-chip cluster can train it as efficiently as a Blackwell farm would.

For now, the scoreboard reads: Z.AI, the first mover with hardware on the floor. DeepSeek, the challenger with the cheaper punch. China's answer to Nvidia just moved from slide decks to working, gigawatt-scale infrastructure — and the race is officially on.