← Back to feed

Google is in active discussions to supply its Tensor Processing Units to South Korean data center operators on a priority basis, according to reporting by Seoul Economic Daily and confirmed by parallel coverage in Korean financial media. The approach marks one of the most aggressive moves yet in Google’s nascent TPU external-sales business, which launched only in April 2026.

The timing is deliberate. Nvidia’s GPU delivery backlog currently runs to approximately 30 weeks, creating a meaningful window for alternative accelerator supply. South Korea has committed to 550 trillion won — roughly $404.9 billion — in private investment to build 8.4 gigawatts of AI data center capacity by 2029. S&P Global projects total Korean data center investment through 2035 at 1,200 trillion won, which would make it the second-largest market in Asia-Pacific after China.

The revenue projection that is driving Google’s pitch

Morgan Stanley has modelled Google’s direct-sale TPU revenue at $84 billion in 2027, rising to $108 billion in 2028. Those are not incremental forecasts: they represent a business line that did not exist commercially before this year.

The shift in Google’s TPU strategy is structural. TPUs were previously deployed exclusively within Google’s own cloud infrastructure — running Search, Gemini, and internal services — or leased to cloud customers via GCP. In an April 2026 conference call, Google formally announced it would supply TPUs to external data centers outside of GCP. The decision was driven by projections of explosive demand growth for AI agents and physical AI workloads.

Anthropic as the proof-of-demand case

The external demand signal Google appears to be citing most aggressively is Anthropic. The Claude developer has already committed to scaling its TPU capacity from 1 gigawatt this year to 5 gigawatts next year — a fivefold increase over 12 months that requires supply agreements with manufacturers rather than spot-market procurement.

That trajectory is exactly what Google needed to validate the external TPU model before approaching sovereign markets.

Korea’s specific appeal

South Korea is emerging as a commercialisation hub for physical AI, with the government actively seeking alternatives to Nvidia-dependent supply chains. Korea’s situation is asymmetric: strong domestic AI demand from Samsung, SK Hynix, and their technology ecosystems, combined with a build-out timeline that runs into Nvidia’s production constraints.

Industry officials in Seoul told local media that Google has identified Korea as one of its primary TPU business bases. The full-stack offer — Google’s own chips, its Gemini model family, and its cloud infrastructure — is positioned as an alternative to Nvidia-dependent configurations that require sourcing GPUs, server platforms, and software stacks from multiple vendors.

Google joined Nvidia and AMD in the race for Korea’s AI infrastructure contracts, the only hyperscaler in the three-way contest: competing on hardware it designed for its own cloud before opening that supply externally.

What this means for the TPU market

The April decision to open external TPU sales was driven by projections, not existing demand. The Korea outreach suggests Google is moving quickly to convert those projections into contracted revenue before the Nvidia backlog clears. At $108 billion projected by 2028, even partial execution would make TPU direct sales one of Google’s largest revenue lines.

The risk for Google is execution: manufacturing TPUs at external-data-center scale requires supply chain discipline that the company has not previously demonstrated for this market. Nvidia’s backlog is a problem for buyers; it is also evidence of demand that sustains pricing power. If Google’s TPU supply is reliable, it changes the accelerator procurement calculus for any government or hyperscale operator planning capacity 12 to 18 months out.