← Back to feed

Google is pushing TPUs into the neocloud channel — the layer of compute providers like CoreWeave, Lambda, Nebius, and Crusoe that sit between Nvidia and the AI labs, startups, and enterprises that need GPU access without hyperscaler contracts. The problem, according to reporting from The Information, is that Nvidia’s relationships in that market are entrenched and the major players are not switching.

CoreWeave, Nebius, and Lambda have told Google their active clusters and customer discussions remain GPU-focused. Nscale, a fast-growing neocloud, said the same. Google’s TPU sales today go through Google Cloud on direct enterprise terms, reaching Anthropic, Meta, and Apple. That is a different motion — hyperscaler contracts with Tier 1 labs — from the neocloud channel, where smaller providers resell compute capacity to developers who need GPU access on more flexible terms.

Why the Channel Matters

Nvidia wins in neoclouds through a combination of tooling lock-in and relationship momentum. Developers train on CUDA. Engineers know the stack. Switching silicon mid-project is expensive regardless of whether the alternative hardware is faster, and Nvidia has spent years building procurement relationships with every provider of consequence in the market.

But the neocloud market is also where a disproportionate share of growth is concentrated. CoreWeave’s Q1 2026 revenue doubled to $2.08B with a $99.4B backlog. Lambda recently closed a $1B credit facility. These are fast-expanding businesses. Getting TPUs into a CoreWeave equivalent would give Google distribution to customers who have never signed a Google Cloud agreement and may never do so.

The barrier is not technical. TPUs already run Anthropic’s Claude workloads at scale. The barrier is the customer’s CUDA habituation and the fact that the major neoclouds are not incentivised to add a second compute architecture when demand for their current Nvidia-based clusters is outstripping supply.

The Blackstone Workaround

Google’s near-term move is a TPU-based neocloud built with Blackstone, structured to begin renting compute to AI labs, financial firms, and high-performance computing customers in 2027. Blackstone committed $5B to the venture, targeting 500MW of capacity.

The Blackstone vehicle does not require converting existing Nvidia neocloud relationships. It enters the market as a new neocloud that happens to run on TPUs rather than H100s or Blackwell. The strategic test is whether Blackstone can attract customers to a TPU-native environment — or whether customers will benchmark the tooling against the GPU neoclouds they already know and find the friction too high to justify switching.

Smaller neoclouds are the other opening. Google’s pitch is clearer to providers who do not already have deep Nvidia relationships: a differentiated compute product rather than a head-on substitution argument. Whether that produces meaningful market share before Nvidia extends its advantage with the next hardware generation is the timeline question.

Context

Google already sells TPU capacity externally at scale. Anthropic’s inference runs on TPUs through Google Cloud. The issue is not whether TPUs can support frontier AI workloads — they demonstrably can. The issue is the distribution architecture: Google Cloud’s direct enterprise motion versus the neocloud intermediary channel that Nvidia controls through years of embedded relationships.

The Blackstone vehicle is a bet that a purpose-built TPU neocloud can compete for the financial and HPC segments that are not already committed to Nvidia. It is a narrower ambition than displacing Nvidia in the existing neocloud market, and it is structured around a 2027 timeline that gives Google time to refine the story before the mainstream GPU contracts at CoreWeave and Lambda come up for renewal.