New YorkGoogle partnered with Blackstone on $5B AI cloud.
Google supplies Tensor Processing Unit inference chips; Blackstone provides data center real estate, the companies said.
The first 500 megawatts of capacity are scheduled for 2027, Blackstone and Google said.
Sources: India Today
WHY TPUs EXIST
Google built the Tensor Processing Unit in 2015 because GPUs were too general-purpose for the matrix multiplications that dominate neural networks. A TPU strips out graphics circuitry and dedicates silicon to systolic arrays — grids of multiply-accumulate units that pump tensors through in lockstep.
TRAINING VS INFERENCE
Training a model happens once and costs billions; inference happens billions of times and costs a fraction per call. The economics of AI are increasingly inference-bound — which is why Google is selling inference-optimized TPUs, not the training behemoths that grab headlines.
THE POWER BOTTLENECK
A modern AI data center campus draws 100-500 megawatts — comparable to a small city. The constraint on AI growth in 2026 is not chips or capital but grid interconnection: utilities quote 4-7 year wait times for new substations in Virginia, Texas, and Arizona.
WHY BLACKSTONE
Hyperscalers used to build their own data centers. As capital intensity exploded, private equity stepped in: Blackstone bought QTS for $10B in 2021 and has since become one of the largest data-center landlords on earth. The pattern is the cloud equivalent of asset-light retail — Google owns the silicon and the software; Blackstone owns the concrete and the substations.
THE OLIGOPOLY BENEATH THE OLIGOPOLY
Only three companies design competitive AI accelerators at scale: Nvidia (GPUs), Google (TPUs), and AMD. All three depend on TSMC to fabricate them, and TSMC's leading-edge nodes run on ASML's EUV lithography machines — of which fewer than 200 exist globally. The AI stack narrows to a handful of choke points.
THE 2027 LAG
500 MW scheduled for 2027 reflects how slow physical infrastructure is relative to model release cycles. A frontier model trains in months; the data center that serves it took years to permit, build, and energize. The gap is why compute capacity, not algorithmic progress, increasingly sets the pace of AI deployment.