Blackstone and Google are launching a dedicated TPU cloud with Blackstone committing $5 billion in initial equity to bring 500 MW of compute capacity online by 2027. This is the first purpose-built third-party distribution channel for Google's accelerator silicon. Total deal value including leverage is approximately $25 billion, per the Wall Street Journal: $5B in equity from Blackstone funds and roughly $20B in debt against the underlying data center and equipment assets.
The service model is distinct from Google Cloud. The new U.S.-based company — unnamed at launch — will offer data center capacity, operations, networking, and TPU compute bundled as compute-as-a-service, giving customers a procurement path to Google's chips that bypasses the standard GCP interface. CoreWeave built the template on Nvidia's H100s; this JV follows the same model on Google's silicon. Google will supply TPUs, software, and services. Blackstone brings data center infrastructure depth — the firm is the world's largest global data center provider and owns QTS Realty Trust, acquired in 2021. The Wall Street Journal reported that site locations have already been identified, with some under construction.
Stack specifics: Google TPUs as the compute layer — chips purpose-built for AI training and inference in production for over a decade — with Google Cloud's managed software and services above them and Blackstone's physical infrastructure below. TPU v6 has been deployed since early 2026 for edge workloads. Benjamin Treynor Sloss, who spent over two decades building and operating Google's global infrastructure, steps out of Google to run the new entity as CEO. Thomas Kurian, CEO of Google Cloud, described the TPUs as "optimized specifically for efficiency and performance in the AI era."
Anthropic, Citadel Securities, and Google's own Gemini all run production workloads on TPUs. Google first shipped the chips in 2015; the design is now over a decade into production use and purpose-built for AI training and inference, with a documented efficiency advantage for agentic AI applications. That profile is narrower than Nvidia's broadly applicable GPUs, but for top AI labs, capital markets firms, and high-performance computing, the TPU track record is substantial.
No pricing, utilization targets, latency benchmarks, or cost-per-exaflop figures were disclosed at launch. The 500 MW figure is a power envelope, not a workload commitment; neither Google nor Blackstone specified chip counts, cluster topology, or interconnect fabric. The $25B total deal value represents roughly 14% of Google's $175–185B guided 2026 capex, positioning this as a meaningful off-balance-sheet financing vehicle—one that lets Google monetize its TPU supply relationship without adding to its own reported infrastructure spend.
The integration risk for any team evaluating this JV as a compute source is software compatibility. Google's TPUs run optimally on JAX and XLA; workloads built on PyTorch/CUDA require non-trivial porting effort. The new commercial channel changes procurement options, not chip architecture requirements. Blackstone made a parallel infrastructure bet with Anthropic earlier in May, signaling the firm is building a portfolio of compute positions across multiple AI supply chains rather than making a single-stack wager.
If your stack already runs JAX or you are greenfielding inference for agentic workloads and want a non-GCP procurement path for TPU capacity at scale, track this JV—but hold any migration planning until pricing, SLA terms, and network egress rates are publicly disclosed.
Written and edited by AI agents · Methodology