Google spent a decade building chips few outsiders could touch. That wall is coming down—selectively.
On May 19, 2026, Google and Blackstone unveiled a joint venture to build a standalone, U.S.-based TPU cloud: a compute-as-a-service platform built entirely on Google’s custom silicon. Blackstone is committing $5 billion in initial equity, taking a controlling stake in a structure expected to scale to roughly $25 billion with leverage.
The ambition is physical as much as financial. The venture is targeting 500 megawatts of capacity by 2027, with plans to expand well beyond that. Running the operation is Benjamin Treynor Sloss, a 20-year Google infrastructure veteran—an early signal this is not a passive capital vehicle.
This isn’t just a partnership. It’s a structural shift in how Google distributes its most valuable hardware.
This Is Google’s Answer to CoreWeave
For the past two years, the AI infrastructure boom has had a clear winner: GPU clouds built on NVIDIA silicon. The most visible example is CoreWeave, which turned access to GPUs into a multi-billion-dollar business.
Google’s Tensor Processing Units have always been the technical counterweight—often more efficient for specific workloads, especially at scale. But access has been tightly controlled through Google Cloud.
This deal breaks that model.
Instead of bundling TPUs inside its cloud stack, Google is effectively spinning up a parallel access layer—a dedicated TPU supply channel that competes directly with GPU-first providers.
As Thomas Kurian has framed it publicly, the goal is to give organizations “more options to access accelerated compute capability.” Translation: TPU supply, without the constraints of shared cloud environments.
The Balance-Sheet Play Few Are Saying Out Loud
There’s a second layer to this deal—and it’s financial.
AI infrastructure is becoming brutally expensive. Industry-wide capex is projected to exceed $700 billion in 2026, with Google alone pushing well past $175 billion. Building data centers at that scale stresses even Big Tech balance sheets.
This joint venture is a workaround.
By shifting the burden of land, power, and construction into a Blackstone-controlled entity—specifically its new AI infrastructure arm, BXN1, led by Jas Khaira—Google offloads capital intensity while retaining control over the highest-margin layer: chips and software.
Blackstone, meanwhile, gets exposure to one of the fastest-growing asset classes in the world.
As Blackstone President Jon Gray described it, the moment represents a “generational opportunity to invest capital at scale.”
The Real Bottleneck Is Power, Not Models
The 500 MW target isn’t arbitrary—it reflects the real constraint shaping AI’s next phase.
The industry doesn’t lack models. It lacks power, cooling, and deployable infrastructure.
That’s where Blackstone’s track record matters. The firm has spent years acquiring and operating physical infrastructure assets, including QTS Realty Trust and AirTrunk. Those are not side bets—they’re the blueprint.
This venture is less about software innovation and more about industrial execution: securing land, negotiating power contracts, and building facilities fast enough to meet AI demand curves.
What This Means for Enterprise Buyers
This platform won’t replace Google Cloud. That’s not the point.
What it creates is a third access model for AI compute:
- Foundation model labs looking for long-term, dedicated TPU capacity
- Enterprise AI teams seeking predictable inference pricing outside shared cloud constraints
- Sovereign AI buyers needing controlled, large-scale infrastructure
The technical nuance matters here. Early signals suggest the platform will lean heavily toward inference workloads, likely powered by newer TPU generations such as “Ironwood,” rather than purely competing in massive training clusters.
The Bigger Test Comes in 2027
The real question isn’t whether this infrastructure gets built.
It’s whether customers follow.
For years, the AI ecosystem has been anchored to NVIDIA by default. This deal creates the first credible, large-scale alternative distribution channel for non-NVIDIA silicon.
If enterprises adopt TPUs at scale outside Google’s walls, the competitive landscape shifts.
If they don’t, this becomes just another well-funded infrastructure bet.
Either way, the experiment is now live—and the results start arriving in 2027.
Related: The $803B AI Illusion: How Big Tech Is Paying Itself to Fake Demand
