Google and Blackstone plan $5B TPU cloud venture to rival CoreWeave

If you wanted to train your massive AI model on Google’s proprietary TPUs, you had one choice: buy them inside the Google Cloud ecosystem.

Now, that gate is coming down.

Google and private equity giant Blackstone are joining forces to launch a massive US-based AI infrastructure company. Blackstone is leading with $5 billion in equity for a reported majority stake, and Reuters notes the venture could reach up to $25 billion with debt financing.

The goal? Bring 500 MW of dedicated AI compute online by 2027 as a standalone "compute-as-a-service" provider, powered exclusively by Google's custom Tensor Processing Units.

The TPU goes neocloud

The venture will be led by Benjamin Treynor Sloss, Google’s longtime Chief Programs Officer who oversaw data center design and networking for the tech giant.

Instead of forcing buyers into the full Google Cloud suite, the new venture will sell raw TPU data center capacity, networking, and operations directly to enterprise customers. Google Cloud CEO Thomas Kurian noted that the joint venture is designed to satisfy skyrocketing appetite for efficiency-focused AI silicon.

The hardware line will likely include Google's newly unveiled processors: the TPU 8t for training and the TPU 8i for inference. The TPU 8i packs a CAE module that Google claims slashes latency by 5x, while the 8t sports a SparseCore memory retrieval system alongside TPUDirect custom networking.

Taking on CoreWeave with 500 MW

This setup positions the venture squarely against the booming "neocloud" market dominated by CoreWeave, Crusoe, and Lambda.

While those players built multi-billion-dollar businesses reselling scarce Nvidia GPUs, Google and Blackstone are taking the opposite approach: dedicated capacity built entirely on proprietary TPUs, backed by deep industrial infrastructure.

Blackstone already manages $1.3 trillion in assets, including over $150 billion in data centers following its acquisitions of Airtrunk ($16.1B) and QTS ($10B).

As HyperFrame analyst Stephen Sopko pointed out, a 500 MW deployment isn't typical IT growth—it's an industrial asset class. The race is shifting from who has the chips to who has the capital, power agreements, and long-duration capacity.

Why it matters

For enterprise builders and AI labs, this marks the first time Google’s elite silicon will be available at massive scale outside the traditional hyperscaler walled garden. If you are struggling with GPU shortages, price gouging, or simply want an alternative to Nvidia's CUDA tax, a standalone TPU cloud offers a viable, dedicated escape hatch.

For the broader AI ecosystem, this is a direct challenge to Nvidia's grip on the neocloud boom. By partnering with Blackstone's balance sheet, Google doesn't have to carry the entire capital burden of new multi-gigawatt facilities on its own books while expanding TPU distribution.

We think this structure is brilliant: Google gets a high-margin chip customer, Blackstone gets premium tenants for its data center and power assets, and the market gets desperately needed silicon diversification. The catch will be developer friction. Google's TPU software stack is world-class for internal teams, but convincing mainstream engineers to ditch CUDA pipelines for JAX and TPUDirect remains the real battle.