Table of Contents
The new Google-Blackstone-powered provider targets 500 MW of capacity by 2027
In sum – what we know:
- A $25B commitment – Blackstone is putting in $5B in equity, with leveraged debt potentially bringing the total deal value to $25B.
- TPUs as a service – Customers buy bundled TPU compute, networking, and operations rather than colocation, with ex-Google Cloud exec Benjamin Treynor Sloss running it.
- A challenge to Nvidia – The venture expands TPU reach beyond Google Cloud and positions Google more directly against Nvidia-centric clouds like CoreWeave and Nebius.
Google and private equity giant Blackstone are teaming up to build a U.S.-based AI cloud provider built around Google’s custom Tensor Processing Units. The new joint venture will sell data center capacity, networking, operations, and TPU compute as a bundled service — effectively a turnkey AI cloud aimed at organizations that want serious training and inference horsepower without standing up their own infrastructure.
Blackstone is putting in an initial $5 billion equity commitment, drawing on its existing digital infrastructure portfolio to build and run the physical data centers. With leveraged debt financing layered on top, the total investment value could reach up to $25 billion, according to reporting from Reuters and Bloomberg. Long-time Google Cloud exec Benjamin Treynor Sloss will be CEO of the new venture.
The initial target is roughly 500 megawatts of capacity coming online in 2027, with explicit plans to expand from there. That’s a very large footprint for an accelerator-dense build, and it positions the venture among the more ambitious single AI infrastructure projects currently underway.
Core of the deal
The structure of the deal is fairly clean. Blackstone owns a major stake and takes on the job of building, financing, and operating the physical infrastructure. Google supplies the TPUs, the management software, and the operational expertise — effectively contributing technology and know-how in lieu of (most of) the capex. For Google, that means scaling TPU capacity without having to fully carry the buildout on Alphabet’s balance sheet.
The product is sold as compute-as-a-service, not colocation or bare data center space. Customers buy capacity on TPUs hosted in the venture’s facilities, with data center operations and networking included. That’s a turnkey environment, not a DIY hosting model, and the revenue model leans on long-term contracted AI compute capacity, with room for value-added managed services on top.
Initial scale is 500 MW, but the venture is explicitly designed to grow beyond that first phase. No concrete next-phase targets have been disclosed.
Technology and infrastructure
Customers will run training and inference workloads directly on Google’s custom TPUs — application-specific accelerators designed in-house for AI workloads. The exact generation on offer (v5p, Trillium, something newer) hasn’t been publicly specified.
The physical infrastructure is what you’d expect for accelerator-dense AI builds. High power density, advanced cooling — almost certainly liquid cooling at this scale — and networking engineered for distributed training across large clusters. The 500 MW figure implies multiple campuses or a multi-phase build, though specific sites haven’t been announced.
On the software side, Google is contributing its AI development stack, which typically means TensorFlow, JAX, and the orchestration tooling needed to manage TPU clusters at scale. How tightly this integrates with existing Google Cloud regions — and how easy it’ll be for customers to move workloads between the two — is an open questions.
Strategic motives
For Google, the logic is straightforward. TPUs have historically lived almost entirely inside Google Cloud, and Google has lagged Nvidia badly when it comes to third-party ecosystem reach — though Google has said that it would be offering its TPUs to external customers. This venture creates a major new distribution channel for TPUs outside the usual Google Cloud quotas and regions, and positions Google more directly against Nvidia-centric AI clouds like CoreWeave and Nebius. It also offers strategic flexibility against Microsoft, which has tied itself to OpenAI and CoreWeave, and Amazon, which is leaning on its own Trainium and Inferentia chips.
Offloading capex matters too. AI-grade data centers are extraordinarily expensive — accelerators, power, cooling, networking, all of it — and partnering with Blackstone lets Google accelerate the buildout without absorbing the full balance-sheet hit.
For Blackstone, the pitch is “generational opportunity.” That’s the language Blackstone President and COO Jon Gray has used, comparing AI infrastructure to early bets on telecom towers or logistics. With more than $1.3 trillion in assets under management, Blackstone can source massive debt financing and lean on existing relationships with utilities, local governments, and construction firms to move quickly on power, land, and permits. AI accelerators also dramatically increase revenue per MW versus traditional data centers, which makes a TPU-dense joint venture an unusually high-yield form of infra exposure. And having Google as an anchor technology partner is the kind of differentiation that’s hard for rival infra investors — Brookfield, KKR, DigitalBridge — to match.
Open questions
Plenty is still unknown. Pricing models, service tiers, and network egress charges haven’t been published, and practitioners will be watching those closely once product details emerge. The integration story with Google Cloud is also unclear — will this be a fully separate provider, or effectively an extension of Google Cloud with its own SKUs? How easily workloads can move between the two will shape how attractive the venture looks to customers already running on GCP.
There’s a real software risk too. Nvidia’s CUDA stack is deeply entrenched, and many AI frameworks and models are tuned for it first. Customers may be wary of porting costs and ecosystem lock-in, especially for workloads that aren’t already TPU-friendly. If Google’s TPU roadmap slips, or competitors leapfrog on performance, the venture’s hardware advantage could erode quickly.
Then there’s the macro question. The bullish case assumes AI demand keeps climbing for years and 500 MW of TPU capacity stays fully utilized. The bearish case is that AI spending normalizes or overshoots, leaving specialized capacity stranded. Hard to call which way that breaks, but the risk is real enough to mention.
Finally, execution. Building 500 MW of AI-grade data centers by 2027 means navigating supply chain constraints on transformers, cooling equipment, and networking gear, plus delays in securing utility power. Blackstone’s infra experience mitigates some of that, but none of it is trivial — and it’s the kind of risk that tends to show up late, when there’s the least slack left to absorb it.
