← Back to feed

Google is backing a $3.2 billion data center project at Lake Mariner near Niagara Falls, New York, that will run on its custom Tensor Processing Units rather than Nvidia GPUs. The structure involves three parties: TeraWulf operates the physical facility, Fluidstack brokers cloud capacity from it, and Anthropic purchases compute on the other end to train its Claude models.

The financing arrangement mirrors a strategy Nvidia has used to lock in hardware demand at scale. Nvidia has long helped customers secure leases and funding for large GPU clusters, effectively underwriting the infrastructure that runs its own silicon. By guaranteeing that infrastructure gets built, Nvidia ensured its chips were embedded in the largest AI training deployments. Google is now applying the same logic to TPUs.

Why This Matters

Google’s TPUs were originally designed for internal services, primarily Search and YouTube. Positioning them as a commercial training alternative to Nvidia’s H100 and B200 chips requires more than claiming the hardware is competitive. It requires demonstrating that large, dedicated TPU clusters can be financed and operated for third-party customers at scale.

Lake Mariner is that demonstration. If Anthropic successfully trains frontier models on TPUs, it becomes a credible proof point that the Google silicon stack can compete with Nvidia in the use case that matters most: frontier AI training. It also deepens Google’s existing investment relationship with Anthropic, tying compute infrastructure to hardware vendor in a way that creates durable dependency.

The compute supply chain matters now in ways it didn’t two years ago. Demand for H100 and B200 chips consistently outpaces availability. Lab CEOs sign waitlists, grey-market Nvidia B300 servers sell at 82% premiums in China, and every major lab is simultaneously lobbying for more allocation and exploring alternatives. For Anthropic, a long-term TPU supply agreement with Google backing provides a hedge against Nvidia supply constraints.

The Competitive Context

Amazon, Microsoft, and Meta are all building proprietary AI chips to reduce Nvidia dependence. Amazon’s Trainium chips power much of the internal AWS workload. Microsoft is developing its Maia line. Meta’s MTIA targets inference at scale. But none has completed the move from internal use to third-party commercial training deployments at frontier scale.

Google is the furthest along. The company has invested in TPU manufacturing at significant scale across multiple generations, and its Cloud TPU product has been available externally since 2018. The Lake Mariner deal moves TPUs from a cloud rental product into a purpose-built, dedicated training facility backed by Google’s own capital.

Nvidia’s position remains dominant. Jensen Huang was signed onto a memory wafer at SK Hynix during Computex in a visual that captured the supply chain’s psychology: even the CEO of the world’s most valuable publicly traded company performs deference to memory manufacturers because demand for HBM outstrips supply. HBM is now 63% of AI chip component costs.

The Structure

Google finances and guarantees the project. TeraWulf, a data center operator with experience in energy-efficient compute infrastructure, manages the physical facility. Fluidstack, the cloud broker that aggregates distributed compute capacity, stands between operator and customer. Anthropic receives the TPU compute through Fluidstack without directly managing the facility.

This three-layer structure is common in hyperscale AI infrastructure deals but uncommon when the hardware vendor is providing the financing. Google is taking the hardware-to-infrastructure vertical integration one step further by owning the financial backstop for a TPU cluster it does not directly operate.

If the Lake Mariner model works, it gives Google a replicable blueprint for TPU deployment at the scale needed to compete seriously with Nvidia as an AI chip alternative for training workloads.