Google and Blackstone Launch Joint AI Venture with $5 Billion Investment
The partners are launching a cloud AI company based on Google TPU chips. The project, valued at up to $25 billion, aims to bring 500 MW of capacity online by 2027 to compete with Nvidia.
Google and Blackstone Strike at Nvidia from an Unexpected Angle
While the market debates Cerebras' fair valuation and awaits a verdict on Musk's lawsuit, Google and Blackstone quietly closed a deal that reshapes the entire AI infrastructure economy. $5 billion in direct investment, a former Google top executive as CEO, a separate company to sell TPU as a service. At first glance, it's just another partnership. In reality, it's a tectonic shift in who will own the foundational layer of the AI stack and how.
The Core: What's Really Happening
For a decade, Google kept TPUs exclusively for itself. Anthropic had access through cloud contracts, Meta had a separate agreement. But Jensen Huang's core argument was simple: TPUs are not an open market commodity, CUDA remains the only universal platform, and therefore Nvidia's 90% share is unassailable.
The deal with Blackstone shatters this thesis. Google hasn't just allowed external clients to rent TPUs via Google Cloud—it has spun off TPUs into an independent business where Blackstone is the majority shareholder. A controlling stake held by a private investment firm means this is not an internal Google Cloud experiment, but a full-fledged market player with a mandate to aggressively capture market share. The CEO is Benjamin Sloss, a man who spent 20 years building Google's global infrastructure. This is not a figurehead; it's a signal: the project has secured the best possible talent.
Timeline and Context
The sequence of events leading to the deal is executed with surgical precision:
- 2015 — Google releases the first TPU. The chip is designed exclusively for internal tasks.
- February 2026 — Google ramps up TPU v6 Trillium to full power: over 1.6 million chips, 4.7x performance improvement over v5e.
- February–March 2026 — Ironwood (TPU v7) reaches general availability: 192 GB HBM3e, 7.4 TB/s bandwidth, native FP8 support.
- April 2026 — Blackstone creates BXN1, an AI infrastructure unit led by Jas Hair, who previously oversaw investments in CoreWeave.
- Late April 2026 — BXN1 closes its first deal: a $1.5 billion joint venture with Anthropic, Goldman Sachs, and Hellman & Friedman.
- May 18, 2026 — The deal with Google is announced. $5 billion in equity, a target of 500 MW by 2027, and a total budget including debt financing of up to $25 billion.
The time gap between BXN1's creation and its second major deal is three weeks. Such speed implies that negotiations with Google were conducted in parallel with the unit's formation, or possibly even preceded it.
Who Wins and Who Loses
Google wins. The company solves three problems with one move. First: capital expenditure. Alphabet allocated $185 billion in capex for 2026. Spinning off TPU infrastructure into a separate company with external financing relieves the parent corporation's balance sheet. Second: Google Cloud's margin. Selling TPU as a service through a separate legal entity isolates infrastructure costs from cloud reporting. Third: TPU market penetration. When a major client chooses between Nvidia GPUs and Google TPUs, they are now offered a "compute-as-a-service" model through a neutral provider, rather than through Google Cloud with its complex ecosystem.
Blackstone wins. The company is methodically acquiring entry points into the AI stack: data centers QTS, AirTrunk, stakes in CoreWeave, Anthropic, OpenAI, and SpaceX. Now the portfolio gains exclusive access to TPUs. This turns Blackstone into a disproportionately influential player in the AI infrastructure market. No other private investor simultaneously controls data centers, chips, and stakes in key AI companies.
Nvidia loses—but slowly. The custom ASIC market is growing at 44.6% CAGR, while GPUs grow at 16.1%. Nvidia currently holds over 90% of the market, but analysts at New Street Research predict its inference share will drop to 20–30% by 2028. The Google–Blackstone deal accelerates this process. Clients who previously bought GPUs simply because TPUs were not publicly available now have a full-fledged alternative.
CoreWeave loses. The company built its business on renting Nvidia GPUs as a service. Now Blackstone, one of its own investors, is creating a direct competitor based on TPUs. This is a classic conflict of interest, and CoreWeave's management will have to ask some uncomfortable questions.
What the Media Isn't Saying
The insight missed by headlines: Google's real motivation is not confrontation with Nvidia, but preparation for OpenAI's IPO. Here's the chain: OpenAI plans an IPO in Q4 2026 with a valuation of up to $1 trillion. OpenAI is one of the largest consumers of Nvidia GPUs. If after the IPO the company decides to diversify its supply (and it has already signed a contract with AMD for 6 GW of capacity), TPUs from Blackstone become a ready-made solution. Google simultaneously competes with OpenAI in AI models and wants to profit from its infrastructure spending. Spinning off TPUs into a separate company under Blackstone's management allows serving OpenAI without the conflict of interest that would inevitably arise from direct work through Google Cloud.
The second hidden point: 67% of all AI compute is inference. Nvidia GPUs are designed as universal accelerators; TPUs are inherently optimized for specific model architectures. At a batch size of 1, the B200 idles over 99% of its tensor cores. TPUs are radically more efficient in this scenario. Google knows this and is building its business around inference, leaving training to Nvidia. The question is not whose chip is faster in synthetic benchmarks, but what the cost per million tokens for inference will be in two years. And the answer for Nvidia is grim.
Forecast: 30 Days and 90 Days
30 days (by mid-June 2026):
The first clients of the new company will be announced—likely Anthropic, Citadel Securities, and one major telecom operator with whom Google has already demonstrated TPU solutions at MWC 2026. Nvidia's stock price may correct by 3–5%, not due to fundamental factors, but because of the noise around the deal. Blackstone, on the other hand, will receive a positive revaluation from analysts: Wall Street loves infrastructure bets with an anchor technology partner. Morgan Stanley already forecasts production of 7 million TPUs by 2028, with each batch of 500,000 chips generating about $13 billion in revenue.
90 days (by August 2026):
Google will announce a second expansion round for the joint venture—targeting 1 GW instead of the initial 500 MW. Blackstone will raise additional debt financing, increasing the total budget from $25 billion to $35–40 billion. The first major client outside the US will be revealed—likely a Japanese or South Korean conglomerate.
Key market signal: if within 90 days AWS or Microsoft announce spinning off their chips (Trainium, Maia) into similar external structures, the "compute-as-a-service" model based on custom silicon will become the new industry standard. If not, Google will gain a temporary 12–18 month window to capture 15–20% of the inference market. Nvidia will retain dominance in model training but will begin losing ground in inference faster than current forecasts assume.
The Google and Blackstone deal is not just the news of the week. It is a blueprint for what AI infrastructure will look like in 2028: not a monopoly of a single GPU manufacturer, but a market of several vertically integrated stacks owned by private investors and leased as a utility. Those who understood this first have already invested $5 billion.
— Editorial Team
No comments yet.