To build a serious rival to Nvidia, Google has reached for Nvidia’s own playbook.
A Wall Street Journal investigation lays out how Google is using financial guarantees and what the industry calls “circular financing”, the very tactics that helped make demand for Nvidia’s chips unstoppable, to win data-centre customers for its own silicon.
The clearest example sits on the southern shore of Lake Ontario, a short drive from Niagara Falls.
The $3.2bn guarantee by the falls
At an AI data-centre cluster in western New York known as Lake Mariner, Google has provided a $3.2bn financial guarantee, according to the WSJ. The site’s developers, TeraWulf and the Google-backed cloud provider FluidStack, will rent computing power from thousands of Google’s tensor processing units, or TPUs, to Anthropic.
The guarantee helps the data centre raise cheaper debt, which is exactly how Nvidia has repeatedly stoked demand for its own chips. “They want to be in the game, they don’t want to be left behind,” TeraWulf co-founder Nazar Khan told the paper.
Circular financing, the Nvidia way
The other borrowed tactic is circular financing, in which some of the money a chipmaker invests flows back to it as chip purchases. Google is backstopping a string of Anthropic projects on the same logic: a $7bn data centre called River Bend near Baton Rouge, and another $1.4bn in guarantees for a computing lease in Colorado City, Texas.
It all sits on top of an already-vast Google and Broadcom compute agreement, and feeds the same machine as a roughly $35bn private-credit deal, arranged by Apollo and Blackstone, that buys Google TPUs and leases them to Anthropic.
Going direct, and going for the kill
Google is no longer keeping its chips to itself.
In May it said it would start selling TPUs directly to customers and unveiled its first chip built specifically for inference, the fast-growing job of serving AI queries rather than training models. It also struck a $5bn deal with Blackstone to launch a cloud company aimed squarely at Nvidia-backed providers like CoreWeave and Nebius, and said this month it would raise $85bn in equity, largely for AI infrastructure.
The pitch is starting to land: Citadel Securities, an early adopter, says it runs some workloads at 30 per cent lower cost and up to four times faster on TPUs.
Nvidia is not impressed
Jensen Huang has played the threat down. Nvidia’s “market reach is far greater than any TPU or ASIC can possibly have”, he said in April, arguing Anthropic is Google’s only significant outside TPU customer and challenging Google to prove its chips are cheaper.
Nvidia still holds north of 90 per cent of the AI-chip market, defended by its CUDA software and plug-and-play hardware, and some smaller clouds say they fear losing their Nvidia allocation, “Jensen jail”, if they stray. Huang’s public line is more relaxed: “Nothing gives me more joy than when you buy everything from Nvidia. But it gives me tremendous joy if you just buy something from Nvidia.”
Why it matters
Google is not the only challenger. AMD, Broadcom and Cerebras are circling, and Amazon is running the same play with its $50bn Trainium business and its huge Anthropic clusters. But none of them can match Google’s balance sheet, and that is the point.
Under Amin Vahdat, promoted in December to run its AI infrastructure, Google has stopped being polite about competing, even as he insists “it’s not zero-sum” because “there’s so much demand out there”. When the richest challenger starts copying the exact financing flywheel that built Nvidia’s empire, Nvidia’s 90 per cent moat finally faces a real test, and the AI build-out takes on yet more debt-fuelled, circular risk.