Blackstone is putting $5 billion in equity into a new US-based joint venture with Google that sells TPU compute as a service outside Google Cloud's existing platform. 500 megawatts come online in 2027 with expansion planned after. Benjamin Treynor Sloss — the longtime Google infrastructure executive who codified SRE as a discipline — is CEO. Google supplies the silicon (the announcement references the 8th-gen lineup: TPU 8t for training, TPU 8i for inference), software, and supporting services. The JV builds and runs the data centers. This is the first credible decoupling of TPU access from the GCP stack since TPUs launched.
Sloss as CEO is the operational signal that matters. The man who built Google SRE and the methodology behind production-reliability calculations everyone else now uses is running a multi-billion-dollar AI infrastructure operator. That's a reliability-first ops CEO appointment, not a financial-engineering placeholder. 500 MW in 2027 is substantial — for context, CoreWeave's GPU footprint was roughly 800 MW in 2025, so this venture comes in at TPU-cloud-operator scale rather than boutique. Structure: Google brings the chips and the stack; Blackstone brings the capital and the operator construct; the JV brings the data-center buildout, networking, and operations.
Ecosystem effect. For three years the AI infrastructure cloud playbook was "buy NVIDIA GPUs, build a cloud, sell GPU-hours" — CoreWeave, Lambda, Nebius, Crusoe, Yotta. Google now has the structural equivalent for TPU. Until today, TPU access meant committing to GCP's full stack (BigQuery-adjacent IAM, GKE, Cloud Storage, billing). Now TPU compute is available as a standalone product. Three downstream effects: Anthropic, one of the largest TPU buyers, gets a credible non-GCP option for additional capacity; non-GCP enterprises (hedge funds, quant shops, sovereign deployments) that wanted TPU without GCP lock-in get a buyer; NVIDIA's "default AI compute" position takes a real capitalized hit. AMD MI300/MI400 and AWS Trainium remain captive in their respective clouds — Google's structural move puts pressure on AWS to do something similar with Trainium, and on AMD to find a JV partner if they want to compete at the operator layer instead of just the silicon layer.
Monday: if you want TPU for production workloads, track this venture's GA timeline carefully — 2027 lands the same window as NVIDIA Rubin shipping in volume. Workloads that are genuinely TPU-favorable (large-batch training of dense transformers, inference on Gemini-style architectures) may see a cost-curve inflection, since third-party operators historically price more aggressively than captive-cloud same-rack. Workloads that run equally well on GPU or TPU face a strategic choice: bet on the Google ecosystem maturing into a real alternative compute substrate, or stay on the broader NVIDIA stack. The Sloss appointment raises the probability this venture ships reliable capacity rather than announces it. Names to track over the next 6 months: Sloss's hires. The composition of his ops org will tell you whether this is a real reliability-first operator or a balance-sheet vehicle dressed up.