China's Z.AI Flips the Switch on a 1-Gigawatt Data Center Built Entirely on Chinese Chips — and the Global Cloud Industry Is Watching
Z.AI, the Chinese AI company formerly known as Zhipu, has fired up a 1-gigawatt data center powered completely by Chinese-made processors, marking the first time a facility of that scale has run without a single NVIDIA GPU inside. The center, now operating at partial capacity in an undisclosed Chinese city, represents Beijing's most concrete bet yet that it can build world-class AI infrastructure without American chips.
The facility's 1 GW power draw puts it in the same league as the largest Western data centers operated by Google, Microsoft, and Amazon. For context, that's enough electricity to power around 750,000 US homes at any given moment. And every watt runs through accelerators designed by Chinese companies — primarily Huawei's Ascend series and Cambricon's MLU processors.
"China needs to run its own race on AI infrastructure," Chen Ying, a Beijing-based tech policy analyst, told local media after the announcement. "Relying on restricted NVIDIA hardware for the country's most critical AI compute creates a dependency the government cannot accept."
The project lands at a moment when the global data center industry is facing a reckoning on multiple fronts. New York became the first US state to impose a year-long moratorium on hyperscale data center construction, citing grid strain and environmental concerns. Oracle is staring down a potential $7 billion collateral bill tied to its Wisconsin data center buildout, as reported by the Financial Times. Redfin polling found that most Americans now say "not in my backyard" to new AI data centers, as covered by Redfin News. And a growing chorus of analysts is asking whether the entire data center boom — fueled by hundreds of billions in debt — is heading for a subprime-style collapse.
The 1-Gigawatt Milestone and What It Means
Z.AI's 1 GW center is not just a technical achievement — it's a political statement. The company said the facility has started partial operations and will provide the computing capacity needed to train its most advanced GLM family of language models. Z.AI has also built or operated several clusters containing more than 10,000 chips each, suggesting the company is scaling fast.
The timing is not accidental. Moonshot, the Beijing-based startup behind the Kimi K3 model that recently challenged OpenAI and Anthropic on benchmark leaderboards, suspended new subscriptions on Sunday to prioritize computing resources for existing members. The move underscored a hard reality: Chinese AI companies have the talent and the models, but they are hitting an infrastructure ceiling.
China plans to spend around 2 trillion yuan ($295 billion) over the next five years on data center construction, the government has said. That's nearly double what the US Department of Energy estimates American hyperscalers will invest over the same period. Z.AI itself has raised billions through a Hong Kong IPO and subsequent share sales, and the company is reportedly on track to generate $1 billion in annual recurring revenue after hitting its 2026 sales target in July.
But the chip question remains the biggest variable. Huawei's Ascend 910B and the newer 910C are the closest domestic alternatives to NVIDIA's H100 and B200, but independent benchmarks suggest they still trail on raw floating-point performance and software ecosystem maturity. Cambricon's MLU370 and the newer MLU590 have carved out a niche in inference workloads, but neither has the mindshare that CUDA enjoys among AI developers worldwide.
"Domestic chips can run domestic models," said a Shanghai-based semiconductor analyst who declined to be named because they consult for multiple Chinese chip firms. "The gap with NVIDIA is closing, but it is not closed. A 1 GW data center running on Huawei Ascend is more impressive as a supply-chain statement than as a competitive-computing statement. Still, you have to start somewhere."
The Subprime Data Center Shadow
While Z.AI is pushing ahead with one of the world's biggest AI compute builds, analysts in the West are growing increasingly nervous about the financial architecture propping up the wider data center boom. Ed Zitron, writing in his Where's Your Ed At newsletter, published a detailed analysis this week comparing the current AI data center buildout to the subprime mortgage crisis that triggered the 2008 financial meltdown.
The argument is structured around Special Purpose Vehicles (SPVs) — separate legal entities that own the chips, the debt, and most of the risk for a given data center project. Companies like CoreWeave have raised tens of billions of dollars through non-recourse SPV loans, meaning that if the customer fails to pay, investors get squeezed based on their seniority in the debt structure. CoreWeave alone sits on roughly $30 billion in debt, much of it tied to contracts with OpenAI, Meta, and Microsoft.
Bloomberg estimates there is over $500 billion in outstanding AI data center debt, with at least $200 billion held by private credit funds — roughly 8 percent of all outstanding private credit loans. Nikkei Asia reported this week that Meta, Google, Amazon, Microsoft, and Oracle have accrued around $1.65 trillion in outstanding debt over the last five years, with additional "off-balance-sheet" obligations that do not appear on their published financial statements.
The off-balance-sheet mechanism works like this: BlackRock raises $12 billion to build a data center for Meta. A holding company — say, "Project Sopaipilla Holdings" — is formed. BlackRock owns 80 percent and Meta owns 20 percent. The holding company buys NVIDIA GPUs and pays construction firms. Meta's lease payments theoretically pay down the debt. But the $12 billion never appears on Meta's balance sheet because Meta is not the "primary beneficiary" of the VIE (Variable Interest Entity). Meta's long-term debt sits at around $58.7 billion on its books, but it has at least $39 billion in additional data center exposure that is not consolidated.
"That's the definition of systemic risk hiding in plain sight," Zitron writes. "The vast majority of the AI data center compute revenue is contingent on the continued ability of two unprofitable, unsustainable AI companies to raise tens or hundreds of billions of dollars a year."
Oracle's $7 Billion Problem and the Wisconsin Standoff
The risks Zitron describes are already surfacing in real projects. Oracle faces a potential $7 billion collateral call on its Wisconsin data center, the Financial Times reported this week. The facility, part of the Stargate AI infrastructure initiative announced jointly with OpenAI and SoftBank, may require Oracle to post additional capital if certain milestones around power delivery or customer commitments are not met.
The Wisconsin project is emblematic of the tension between hyperscale ambition and physical reality. Building a multi-gigawatt data center requires power substations, transmission lines, water rights, and construction labor — all in increasingly short supply. The Stargate project alone envisions five 1 GW campuses in Abilene, Texas, with a total cost estimated at $100 billion or more.
New York's decision to pause hyperscale data center construction for a year, approved by the state legislature last week and expected to be signed by Governor Kathy Hochul, adds regulatory uncertainty on top of the financial and engineering constraints. The law is the first of its kind in the US and could set a precedent for other states grappling with the grid impact of AI compute.
Communities are pushing back too. A Redfin survey published July 22 found that the majority of Americans oppose data center construction in their neighborhoods, citing noise, water usage, and aesthetic concerns. In Georgia, a private company has seized homes under eminent domain to make way for a $16 billion AI data center, prompting lawsuits and local protests that have drawn national attention.
Cloud Providers Face the Crosscurrents
For the major cloud platforms — AWS, Microsoft Azure, and Google Cloud — the situation creates an unusual set of pressures. On one hand, AI compute demand from their own internal workloads and from Anthropic and OpenAI is enormous. On the other, the cost of building the infrastructure to meet that demand is mounting so fast that even trillion-dollar companies are looking for ways to push debt off their books.
Meta recently confirmed that it has started a cloud business to sell excess AI GPU capacity, and Bloomberg reported earlier this month that the company is in talks to lease compute to Anthropic. That is a remarkable turn for a company that spent most of the last two years buying GPUs for its own AI research. Zuckerberg's team appears to have concluded that it has built more capacity than its own models need — and that renting the surplus to an AI lab desperate for compute beats leaving GPUs idle.
The neocloud ecosystem is also under strain. Nebius raised a $775 million debt facility backed by contracted cashflows and deployed GPU infrastructure — meaning that if the contracts fall through, lenders are left holding silicon that was already deployed. Runpod has $120 million in annual recurring revenue and Lambda had $114 million in revenue as of mid-2025. Neither is anywhere close to justifying the valuation multiples that investors have assigned to the sector.
For readers tracking the cloud infrastructure sector, our Semiconductors coverage provides deeper analysis of chip supply dynamics, and our Cloud and Edge Computing coverage covers how these trends affect data center operations.
What Comes Next
Z.AI's 1 GW milestone is a genuine feat of engineering and supply-chain management. It proves that China can build hyperscale AI infrastructure using domestic components — a capability the US government's chip export restrictions were designed to delay. But it also highlights how far behind the domestic ecosystem still is in software tooling, developer mindshare, and raw per-chip performance.
For Western markets, the bigger story is the mounting evidence that the data center boom is running ahead of real demand. Sightline Climate's February report calculated that there is 190 GW of data center capacity in planning globally, against roughly $100-120 billion in annual AI compute revenue. Zitron estimates that 70-90 percent of that compute demand comes from just two unprofitable companies, OpenAI and Anthropic. If either stumbles, the debt-backed SPV structure that funds most new construction could unravel quickly.
The data center industry has never seen a buildout of this scale, financed this way, with this much concentration of demand. Whether Z.AI's national-chip data center is the template for the future or a costly detour depends on whether the models it trains can generate enough value to justify the infrastructure. For now, both China and the West are betting billions that the answer is yes — and neither side is entirely sure the bet pays off.
Sources: Bloomberg, Financial Times, Yahoo Finance, Sightline Climate, Nikkei Asia, Tom's Hardware, Where's Your Ed At, Politico, Redfin.