Blackstone secures the majority stake in Google's new TPU cloud.

      Blackstone and Google have established a joint venture to create a US-based AI compute-as-a-service company utilizing Google’s tensor processing units (TPUs), as announced by Blackstone on Monday. Blackstone is providing $5 billion in initial equity while taking majority ownership, with the overall deal value estimated at around $25 billion, including leverage. The initial target is to achieve 500 MW of data center capacity by 2027.

      It is important to note that the new business is framed as a ‘TPU cloud’ rather than a typical hyperscaler unit. The competitive landscape it addresses is CoreWeave, the NVIDIA-aligned neocloud that made its public debut in 2024 and now plays a significant role in the AI-compute sector adjacent to hyperscalers. Essentially, this venture represents Google’s competitive response to CoreWeave: leveraging a chip architecture Google has meticulously developed over the past decade, integrating Blackstone’s infrastructure financing capability, and offering capacity to enterprise clients seeking alternatives to NVIDIA silicon under compute-as-a-service terms.

      The financial structure follows a familiar Blackstone infrastructure-fund model. The $5 billion equity stake is part of what is being described as a $25 billion total project worth, with the remaining $20 billion expected to be sourced from debt financing linked to the data center and equipment assets. Blackstone’s current infrastructure portfolio includes the QTS Realty Trust data center platform, which it acquired in 2021, demonstrating the firm’s increasing reliance on asset-heavy, long-term cash-flow initiatives as the AI infrastructure sector expands.

      The implications for Google pose a separate set of considerations. Over the past year, Google has been selling TPU capacity to external clients at an unprecedented rate. The recent $40 billion-plus agreement with Anthropic includes five gigawatts of TPU capacity over five years and access to up to one million seventh-generation Ironwood chips. Earlier this year, Meta also entered into its own TPU agreement.

      The success of these external sales has led to a well-known internal compute challenge, where Google’s own AI researchers, including teams at DeepMind, are now waiting for access to the same resources the company is offering to external partners. The Blackstone joint venture represents the third element of this external distribution strategy.

      The capital-cycle context contributes to the deal’s scale. Big Tech spending on AI infrastructure is projected to exceed $700 billion this year, with Google alone expected to spend between $175 billion and $185 billion in capital expenditures. When compared to Google’s overarching financial capacity, the $25 billion joint venture with Blackstone appears relatively modest. However, its structural significance lies in enabling Google to transfer the infrastructure financing responsibility off-balance-sheet while maintaining the TPU supply and architectural relationship that generates the underlying profit margin.

      The competitive perspective for CoreWeave and the broader NVIDIA Neocloud sector becomes increasingly relevant. CoreWeave's stock has been trading based on the belief that the AI-compute-as-a-service market is structurally underserved and that its existing relationships with NVIDIA give it an advantage for several years ahead. A Google-Blackstone TPU cloud with 500 MW of capacity by 2027 emerges as its most substantial competitor to date, due to the differentiated silicon technology and Blackstone’s virtually limitless financing capacity within infrastructure constraints.

      The choice of customers remains a point of uncertainty for both parties. Potential buyers can be categorized into three groups: foundation-model labs apart from Anthropic or Meta seeking TPU capacity through long-term contracts; enterprise clients purchasing GPU compute via CoreWeave or Lambda who are looking for TPU alternatives for inference-heavy applications; and sovereign AI buyers in regions where NVIDIA export restrictions have led to procurement challenges. The focus of the joint venture on one of these segments will influence whether it remains at 500 MW or expands further.

      Operational details, such as data center locations and TPU generations, are still in progress, with the next significant indicator being the announcement of the JV’s first named anchor customer.