Google Pushes TPU as Nvidia Rival with $3.2 Billion Lake Mariner Cluster

Google is turning its in-house tensor processing units into a direct rival to Nvidia by backing a $3.2 billion AI data center cluster near Niagara Falls, New York, according to people familiar with the matter. The facility, called Lake Mariner, will rent computing power from thousands of Google’s TPUs to Anthropic.

Alphabet’s financial guarantee for the project, developed by TeraWulf and FluidStack, mirrors the financing model Nvidia has used to lock in demand for its GPUs. Over the past year, access to compute has become the primary bottleneck for AI companies, pushing hyperscalers and financiers to underwrite large clusters.

Google’s TPU Strategy Expands Beyond Cloud

Google’s TPU program dates to 2013, when chief scientist Jeff Dean realized that deploying a speech model to 100 million users would require doubling Google’s computer fleet. The seventh-generation TPU, which Anthropic already uses for training, prompted SemiAnalysis to ask in November whether it marks “the end of Nvidia’s dominance.”

• Google recently struck a $5 billion deal with Blackstone to create a new cloud-services business targeting Nvidia-aligned providers such as CoreWeave and Nebius.
• Google now sells TPUs directly to customers, not just through Google Cloud, and has released its first TPU designed specifically for inference.
• Citadel Securities, a longtime Google Cloud client, reported running key workloads 30% cheaper and up to four times faster on TPUs, according to CTO Josh Woods.

Nvidia’s Dominance Faces Growing Pressure

Nvidia still controls more than 90% of the AI chip market, supported by its CUDA software stack. CEO Jensen Huang said in an April podcast that he doubts TPUs can demonstrate a meaningful cost advantage. Some cloud providers worry that shifting spend to alternative chips could land them in what insiders call “Jensen jail,” a term describing fear of losing access to Nvidia’s most coveted silicon.

Meanwhile, competitors are not standing still. AWS recently took Nvidia Blackwell instances to general availability and announced plans to add more than one million Blackwell and Rubin GPUs this year. Mantle DC launched a GPU-as-a-service offering with 144 Blackwell GPUs, building its foundation cluster in just five months. Google’s bet on TPUs adds another layer of competition in a market that is still defined by Nvidia’s hardware and software lock-in.

What comes next depends on whether Google can scale TPU adoption beyond internal and select cloud workloads. The company’s financing moves and direct-sales pivot suggest it is prepared to invest billions to find out.