Google is escalating its campaign to break Nvidia's grip on artificial intelligence chips, deploying creative dealmaking to push its Tensor Processing Units (TPUs) into the wider market.
The strategy, detailed in exclusive reporting by The Information's Amir Efrati, targets the "neo-clouds": specialised GPU cloud providers, many of which began life as cryptocurrency miners.
Hundreds of these firms exist, but only about half a dozen matter, and Google has been in talks with them about adding TPUs to their offerings.
The pitch is diversification, freeing these providers from total dependence on Nvidia, alongside a technical argument that TPU designs have remained stable while Nvidia's architecture changes radically with each generation.
Those frequent shifts create genuine installation headaches for the data centre operators tasked with deploying them.
Financial firepower
The competition is increasingly being fought with balance sheets rather than benchmarks.
Nvidia has long used its financial muscle to support its largest customers, and Google is now considering matching that approach by offering backstop deals to lenders.
Under these arrangements, Google would guarantee payments if businesses that borrow to buy TPUs cannot find renters or buyers for the chips.
Google holds a structural advantage here: unused TPUs can simply be absorbed into its own cloud operations, whereas Nvidia lacks a cloud business of comparable scale to soak up stranded hardware.
A joint venture with Blackstone to build a TPU-based cloud provider extends the same logic.
The friction is already visible, with reports suggesting Nvidia became aware of Google's discussions with neo-cloud provider Nscale and may have offered additional incentives to discourage TPU adoption, though Nscale has said on the record that this is not its position.
Jensen Huang is said to monitor Google's chip programme closely and regards the company as a significant competitive threat.
An awkward embrace
The rivalry is complicated by mutual dependence.
Google remains one of Nvidia's largest customers, buying GPUs at scale for a cloud business that serves external clients, and it currently needs Nvidia's supply as much as Nvidia needs its custom.
Meanwhile the external TPU business is gaining real traction, with Anthropic, Apple and Meta among the clients, and Meta emerging as a significant customer.
The next constraint is manufacturing, since TSMC is the bottleneck through which all chip ambitions must pass, and Google secures its capacity via Broadcom as intermediary.
Allocations for 2027 production are being determined now, and the capacity Google wins will indicate how seriously TSMC takes the TPU business against competing demands from Nvidia and others.
The stakes extend beyond chips: with gigawatt data centres costing $50 billion to $60 billion and rising, the companies able to guarantee that spending will shape the infrastructure of the entire AI economy.
