You could see it as Google’s way of making more of its own TPU chips and keeping up with the insatiable need for AI infrastructure. The demand for this kind of computing has left traditional clouds in the dust, and buyers are looking for something other than an all-Nvidia setup. This is the first time Google has made such a concerted effort to put its Tensor Processing Units in front of the open market for both training and inference.
The numbers don’t lie: Big Tech is on track to spend over $700 billion on AI build-out by 2026. Alphabet has even hiked its 2026 capex outlook to between $180 and $190 billion, from the $175-185 billion range before. It’s a sign of how much compute means to their bottom line now. And with a new inference processor and a top-tier training chip out the door last month, you can tell they are in a hurry to get models into production.
Inside the venture: capital, capacity, control
On the ground, sources say Blackstone will be the one with the equity and the control. The plan is to offer a full service-compute, software, the works, all running on Google’s hardware. They have their eye on 500 megawatts in 2027 and want to be in a position to grow from there. Some of the data centres are already in the mix, some of them still being built. Bloomberg puts the total value of the deal at $25 billion if you count in the leverage. A Google veteran, Benjamin Treynor Sloss, will be at the helm.
Here are the milestones to watch next:
– First 500 MW online in 2027
– Investment could reach $25 billion including leverage
– Benjamin Treynor Sloss named CEO
– Compute-as-a-service built on TPUs
Chips, software and services
“Together, we are accelerating AI transformation,” says Google Cloud’s Thomas Kurian. “This joint venture with Blackstone is about meeting the demand for TPUs, which are built for efficiency and performance in the AI era.”
Competitive pressure on Nvidia and neocloud rivals
It’s no secret that most of the big AI labs are using what CoreWeave has to offer with its Nvidia gear, and there are plenty of neoclouds popping up to fill the void. By putting TPUs out there in force, Google is not only upping the ante with Nvidia but also giving model makers a way to work without relying on CUDA.
People have been wondering for a while if Google would let anyone use its TPUs. They’ve made some inroads, like the side-deal with Meta and the one with Anthropic for a million or so of their chips. But with Blackstone behind it, this is a step further, with its own rules and dedicated room to run.
Blackstone’s AI buildout and governance
Blackstone has been making no bones about being the top dog in data centres. They’ve been piling on bets from power to property, from the 2021 purchase of QTS to the 2024 AirTrunk deal. They have stakes in OpenAI, Anthropic and CoreWeave. Chairman Stephen Schwarzman has put a figure of $150 billion on their data-center book, with another $160 billion in the pipeline.
They’ve even set up a unit called Blackstone N1 to put all their AI eggs in one basket, with Jas Khaira in charge. This Google project is their second foray through BXN1, after a $1.5 billion venture with Anthropic to put AI tools in front of businesses. “We see a generational opportunity to put capital to work on AI infrastructure,” says Blackstone President Jon Gray. “There is an unprecedented need for compute, and this company is in a position to answer it.”
