Google and Blackstone Launch $25 Billion TPU Cloud Company to Challenge Nvidia's AI Compute Grip

Google’s in-house AI chip program has powered some of the most impressive model training runs in history, but for most of that time, its Tensor Processing Units remained a competitive secret rather than a commercial product. That changes with this week’s announcement.

Google and Blackstone, the private equity and real estate giant, have formed a US-based joint venture to build, operate, and sell TPU-based cloud compute as a managed service. Blackstone is making an initial equity commitment of $5 billion, with plans to scale total investment to as much as $25 billion through leverage as the business grows. The announcement came on May 18, one day before Google’s I/O keynote, suggesting the company is in an aggressive mode of infrastructure expansion heading into the back half of 2026.

The new company is not a rebranding of Google Cloud. It is a distinct commercial entity — with its own CEO, its own sales motion, and its own customer targets — designed to sell TPU compute directly to enterprises that want Google-quality AI infrastructure without navigating the full Google Cloud stack.

The Man at the Helm

Benjamin Treynor Sloss has been named chief executive of the new venture. Treynor Sloss is a Google veteran of more than two decades who built and operated Google’s global infrastructure and reliability engineering organization — the systems that underpin Search, Gmail, YouTube, and every other Google service at planetary scale. His appointment signals that this is not a paper joint venture: Google is deploying one of its most operationally experienced executives to run it.

The CEO choice also carries a message to enterprise customers. Treynor Sloss is not a sales or marketing figure. He is the person who has kept Google’s infrastructure running at scale longer than almost anyone else. His credibility in conversations with CIOs and infrastructure architects is exactly the kind of trust signal a new compute vendor needs.

Why This, Why Now

The timing reflects supply-side pressure and competitive reality converging simultaneously.

On the supply side, the demand for AI compute has outpaced what existing hyperscaler commercial models can absorb at the speed the market requires. Combined Big Tech AI infrastructure spending is projected to exceed $700 billion in 2026 — up from earlier forecasts of $600 billion, a revision that itself would have seemed extraordinary twelve months ago. To build data centers at the pace the demand signal requires, Google needs capital from outside its own balance sheet. Blackstone’s $100+ billion in real estate and infrastructure investing expertise provides a faster route to permitted, connected, and constructed capacity than building organically.

On the competitive side, the stand-alone AI compute business model has been validated at scale. CoreWeave — the GPU cloud provider backed by Nvidia — went public in early 2025 and established that enterprises will pay premium prices for specialized AI compute infrastructure managed by a focused operator rather than a general-purpose hyperscaler. Cerebras Systems completed the largest tech IPO of 2026 earlier this month, raising $5.55 billion at a peak valuation of $86 billion, confirming that the market assigns extraordinary value to differentiated AI silicon and the infrastructure built around it.

Google has a differentiated asset in its TPUs that it has been systematically underutilizing commercially. The joint venture is the mechanism to change that.

What the Company Will Actually Do

The venture will sell TPU capacity as compute-as-a-service — think “GPU-as-a-service” but powered by Google’s own silicon. Enterprises buy access to training and inference infrastructure without managing hardware themselves, and without being locked into the broader Google Cloud ecosystem of databases, analytics tools, and managed services.

The first 500 megawatts of data center capacity is targeted to come online in 2027. For context, a single modern AI data center campus typically runs at 100 to 200 megawatts; 500 MW in a single venture’s first tranche represents a meaningful fraction of current industry buildout pace.

Google contributes the TPU hardware, firmware, and software stack, along with ongoing model services and engineering support. Blackstone contributes equity capital and real estate development expertise — critical in an environment where data center site permitting, power procurement, and construction timelines are frequently the bottleneck that determines whether an infrastructure plan survives contact with reality.

The Anti-Nvidia Thesis

For Google, this is as much a strategic play against Nvidia as it is a revenue opportunity. Nvidia’s H100 and H200 GPUs remain the default training and inference hardware for most AI labs and enterprises, giving the company extraordinary pricing power and customer lock-in that has made it the most valuable semiconductor company in history.

Google’s TPUs offer a genuine architectural alternative. The company trained Gemini 3.5 Flash — announced at I/O 2026 the following day — primarily on TPU infrastructure, and the efficiency gains at Google’s own scale of operation have been well documented internally. The commercial question has always been whether Google could recreate those economics for external customers who lack Google’s proprietary software stack. The new joint venture is a $25 billion bet that the answer is yes, and it is building a dedicated commercial operation to find out.

Competitors are exposed in different ways. Amazon has Trainium and Inferentia, custom chips that have found some enterprise adoption but have not displaced Nvidia as the default. Microsoft has invested in custom silicon for Azure but has not taken the step of spinning out a dedicated compute-as-a-service company. Neither has built the kind of capital structure or independent commercial entity that Google and Blackstone have assembled.

The Blackstone Angle

For Blackstone, this is a high-conviction long-term infrastructure bet. The firm has been building its data center portfolio for years, and the Google partnership provides a guaranteed anchor customer for capacity it would otherwise need to lease to third parties at uncertain future rates.

Jon Gray, Blackstone’s President, has said in recent months that data center infrastructure is the firm’s single largest active investment theme. The Google partnership provides technological differentiation — TPU-powered infrastructure rather than commodity GPU clusters — that justifies premium pricing and reduces exposure to the Nvidia supply chain volatility that has created operational challenges for pure-GPU data center operators throughout 2025 and 2026.

The leverage structure — starting at $5 billion equity and scaling toward $25 billion total — mirrors the financing models Blackstone has used for other large-scale infrastructure investments, where an initial equity commitment anchors a capital structure that expands as assets are built and revenue is contracted.

What Comes Next

The joint venture is expected to begin customer onboarding in 2027, with pricing tiers and initial capacity allocation to be announced before year end. Enterprise buyers in financial services, healthcare, and government — segments that have historically been cautious about pure-cloud AI deployments — are likely priority targets, given the potential to offer dedicated, single-tenant TPU capacity with Blackstone’s physical infrastructure underpinning the deployment.

The venture needs to answer one key question: can TPUs perform well enough on workloads designed by customers who built their models and pipelines around Nvidia’s CUDA ecosystem? Google’s answer is yes. The market’s answer will come when the first enterprise contracts start executing in 2027.