- Google is providing multi-billion-dollar financial guarantees to data centers that commit to its TPU chips — including a $3.2B guarantee for the Lake Mariner cluster in western New York (whose compute Anthropic will rent), a $7B project near Baton Rouge called River Bend, and $1.4B in Colorado City, Texas — replicating the financing playbook Nvidia used to dominate AI compute.
- Google struck a $5B deal with Blackstone to launch a new cloud-services company competing directly with Nvidia-backed CoreWeave and Nebius, and in May announced plans to sell TPUs directly to customers and unveiled its first inference-specialized chip — going head-to-head with Nvidia’s Groq 3 LPU.
- Citadel Securities reports running key workloads 30% cheaper and 4x faster on TPUs, and Google has raised $85B in equity to fund its AI infrastructure buildout — giving it more financial firepower than any other Nvidia challenger.
- The fear of “Jensen jail” — losing Nvidia chip allocations by buying from rivals — has been the primary brake on TPU adoption, but insiders say the Blackstone deal signals that dynamic is beginning to crack, with Jensen Huang publicly dismissing TPUs as serving only Anthropic.
What Happened?
Google is making its most aggressive push yet to commercialize its Tensor Processing Units (TPUs), deploying billions in financial guarantees to data-center developers who commit to its chips — mirroring the strategy Nvidia has used to entrench its GPUs across the industry. A $3.2 billion guarantee backs the Lake Mariner cluster in western New York, where Anthropic will rent compute capacity. A $7 billion project called River Bend is underway near Baton Rouge, Louisiana. Google also provided $1.4 billion in guarantees for a Texas facility. In May, Google announced it would sell TPUs directly to customers for the first time — previously available only through Google Cloud — and unveiled an inference-specialized chip to compete with Nvidia’s Groq 3. The commercial push accelerated under Amin Vahdat, promoted to chief technologist for AI infrastructure in December, who now reports directly to CEO Sundar Pichai.
Why It Matters?
Nvidia holds an estimated 90%+ share of the AI chip market, defended by its CUDA software ecosystem and plug-and-play hardware. The fear of “Jensen jail” — losing Nvidia GPU allocations by defecting to a competitor — has kept most cloud providers loyal. Google’s Blackstone deal is a signal that lock-in is weakening: Blackstone has close ties to both Nvidia and CoreWeave, and even its willingness to partner with Google on a competing cloud platform marks a meaningful shift. Jeff Dean’s 2013 “thought experiment” — realizing that scaling speech AI to 100 million users would require doubling Google’s entire computing fleet — set the company on a decade-long custom silicon path. That bet is now being tested commercially at scale. Citadel Securities’ reported 30% cost savings and 4x speed improvement on TPUs suggests the performance gap with Nvidia has closed materially.
What’s Next?
Google’s $85 billion equity raise gives it the balance sheet to sustain aggressive customer financing for years. The newly announced inference TPU will face its first real market test as AI labs shift compute from training toward inference — a workload where Nvidia’s dominance is less entrenched. Jensen Huang has publicly dismissed TPUs as a niche product, and Nvidia’s next-generation chips will arrive with their own inference optimizations. The deeper strategic question is whether Google’s circular financing model creates durable customer loyalty or merely buys temporary share. If Anthropic’s TPU-trained models continue to close the quality gap with GPU-trained alternatives, it would be the most powerful proof point yet that Google’s decade-long bet on custom silicon is paying off.
Source: The Wall Street Journal
