Key Takeaways

  • OpenAI signed a 1.8 gigawatt multiphase power commitment in Ohio, with Nvidia backstopping a total of 4.25 gigawatts.
  • OpenAI is actively leasing Google Tensor Processing Units (TPUs) alongside discussions with neoclouds for both TPU and GPU capacity.
  • Anthropic leased 1.7 gigawatts of capacity in a single quarter, including 360 megawatts located in Queensland, Australia.
  • Hardware exclusivity is dead; frontier labs now run heterogeneous infrastructure stacks to match specific silicon to specific model workloads.

The Hardware Truce Between Bitter Rivals

Corporate rivalry stops at the substation fence. OpenAI and Google compete head-to-head across consumer products, developer APIs, and foundational model research. Yet when power and silicon constraints hit, old alliances dissolve.

Jordi Hays pointed out the oddity directly: “OpenAI plans to buy Google's chips as part of its AI buildout. Um, so now we have TPUs.”

This shift is not about corporate friendship. It is pure arithmetic. Frontier labs have consumed the near-term supply of ready data center power and top-tier Nvidia GPUs. Waiting eighteen months for a dedicated cluster means losing the race. OpenAI needs compute running immediately, even if that compute belongs to its main rival.

Leasing TPUs also reveals how software stacks have matured. Sourcing Google silicon once required rewriting training pipelines in custom frameworks. Now, translation layers and open compilers make cross-architecture training and inference practical. If an external neocloud or hyperscaler has idle TPU pods with power, OpenAI will buy them.

Gigawatts Are the New CapEx Yardstick

Compute discussions used to track server counts or capital spending. Today, the only unit that matters is power.

Citing data from Dylan Patel at SemiAnalysis, John Coogan highlighted the scale of these new contracts: “AI lab's largest lease ever. More massive deals in the pipeline. AMD taking share. OpenAI's quarter is one 8 gigawatt multiphase commitment in Ohio with Nvidia back stopping 4.25 gigawatt.”

At the same time, Anthropic locked down 1.7 gigawatts across multiple sites, pushing into unexpected regions. As Coogan noted, “Anthropic leased 1.7 gigawatts this quarter with notably 360 megawatts in Queensland, Australia. Let's give it up for compute down under.”

When labs chase power to Queensland, the constraint is no longer silicon supply. The constraint is grid interconnection. Building a gigawatt substation takes years; finding an existing grid node with excess generation takes priority over geographic convenience.

Coogan summarized where this dynamic leaves the market: “Every lab is using every chip and every data center and every hyperscaler and every neocloud and they're using it all for all different things and different chips will be bit better at different models and different part pieces of the inference stack.”

What to Do With This

Audit your production inference pipeline by Friday to see which parts rely on proprietary Nvidia kernels. Run a small test deployment on AMD Instinct or Google TPU instances through an independent neocloud provider. If your code is locked to a single chip family, you are absorbing someone else's supply chain risk for zero margin benefit.