The two have formed a joint venture to build a US-based AI compute-as-a-service business on Google’s TPUs, with $5bn of Blackstone equity, $25bn total deal value including leverage, and 500 MW of capacity targeted for 2027.
Blackstone and Google have formed a joint venture to build a US-based AI compute-as-a-service business on Google’s tensor processing units, Blackstone announced on Monday.
Blackstone will contribute $5bn in initial equity and take majority ownership, with the total deal value reaching roughly $25bn, including leverage. The first 500 MW of data-centre capacity is targeted for 2027.
The structural piece worth reading carefully is the TPU framing. The new business is a ‘TPU cloud’ rather than a general-purpose hyperscaler unit. The competitive frame is CoreWeave, the NVIDIA-aligned neocloud that went public in 2024 and now sits at the heart of the hyperscaler-adjacent AI-compute trade.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!
The venture is, in effect, Google’s answer to CoreWeave: take the chip architecture Google has spent a decade refining, layer in Blackstone’s infrastructure-finance capacity, and sell capacity to enterprise customers who want non-NVIDIA silicon at compute-as-a-service economics.
The financial structure is a recognisable Blackstone infrastructure-fund template. The $5bn equity contribution sits inside what the wire is describing as a $25bn total project value, with the remaining $20bn implied to come from debt financing of the underlying data-centre and equipment assets.
Blackstone’s existing infrastructure portfolio includes the QTS Realty Trust data-centre platform it acquired in 2021, and the JV is the kind of asset-heavy, long-tenor cash-flow vehicle the firm has been increasingly leaning on as the AI-infrastructure category has scaled.
What this means for Google is a separate question. The company has spent the past year selling TPU capacity to external customers at an unusual pace. The $40bn-and-up Anthropic deal announced in October includes five gigawatts of TPU capacity over five years and access to up to one million seventh-generation Ironwood chips. Meta signed its own TPU deal earlier this year.
The success of those external commitments has produced the well-documented internal compute problem, in which Google’s own AI researchers, including teams at DeepMind, are now queueing for the same fabric the company is selling to its external partners. The Blackstone JV is the third leg of that same external-distribution story.
The capital-cycle context is the part that gives the deal its scale. Big Tech AI-infrastructure spending will, on current 2026 capex guidance, clear $700bn this year, with Google itself running between $175bn and $185bn of capex.
A $25bn JV with Blackstone is small relative to Google’s own balance-sheet capacity. Its structural value is that it lets Google move the infrastructure-financing burden off-balance-sheet into a Blackstone-controlled vehicle while retaining the TPU supply and architecture relationship that produces the underlying margin.
The competitive read for CoreWeave and the wider NVIDIA Neocloud category is sharper. CoreWeave’s stock has been trading on the assumption that the AI-compute-as-a-service category is structurally undersupplied and that the company’s existing NVIDIA-allocation relationships give it a multi-year head start.
A Google-Blackstone TPU cloud at 500 MW of 2027 capacity is the most credible competitor that category has yet seen, both because the silicon underneath it is differentiated and because Blackstone’s financing capacity is functionally unlimited within infrastructure mandates.
The customer mix is the part neither side has resolved. The most likely buyers fall into three buckets: foundation-model labs other than Anthropic or Meta wanting TPU capacity inside a long-tenor contract; enterprise customers buying GPU compute through CoreWeave or Lambda who want a TPU alternative for inference-heavy workloads; and sovereign-AI buyers in jurisdictions where NVIDIA export controls have created procurement gaps.
Which of those three the JV optimises for will determine whether it ends up at 500 MW or grows past it.
Operational specifics, including data-centre sites and TPU generations, are still developing; the next visible proof point will be the JV’s first named anchor customer.
Get the TNW newsletter
Get the most important tech news in your inbox each week.