Google selling TPU access externally is the biggest structural threat to Nvidia's dominance yet

Google's decision to expand external sales and cloud access to its Ironwood-generation TPU chips beyond its own internal workloads and existing Google Cloud customers represents a more direct challenge to Nvidia's dominance of the AI accelerator market than any of the custom-silicon efforts at Amazon or Meta, precisely because Google's TPU architecture has years of proven production use training and serving Google's own frontier models, giving it a credibility that newer custom-chip entrants lack. Anthropic's disclosed large-scale TPU commitment, alongside a growing list of other AI labs and enterprises evaluating TPU access specifically to diversify away from Nvidia allocation constraints and pricing, has given Google's external chip business a credible reference-customer base within months of the expanded availability. The economics driving customer interest are straightforward: Nvidia's gross margins on its data-centre GPUs have remained extraordinarily high throughout the AI boom, reflecting its effective pricing power as the default choice for any lab that wants frontier-grade training performance, and any credible alternative that offers comparable performance at a meaningfully lower cost per unit of useful compute represents real savings at the scale frontier labs now operate. Google's ability to offer TPU access at competitive pricing is helped by the fact that TPU manufacturing costs benefit from years of amortised development investment that Google originally justified purely on internal-workload economics, before external sales became a meaningful additional revenue opportunity. Nvidia's competitive response has focused on software ecosystem lock-in - the CUDA programming stack, and the enormous body of existing tooling, libraries and engineering expertise built around it - as the moat that makes switching costs prohibitively high even for customers who might save money on raw chip costs by moving to TPUs or other alternatives. Whether that software moat holds as inference frameworks increasingly abstract away the underlying hardware is one of the central open questions in the infrastructure competitive landscape. For Indian AI infrastructure and cloud-reselling businesses, which have built partnerships primarily around Nvidia-based GPU cloud offerings, Google's TPU expansion introduces a new variable in capacity-planning and partnership decisions, particularly for the segment of Indian AI startups price-sensitive enough that a meaningfully cheaper compute alternative could shift where they choose to train and serve their own models. What to watch: how much external TPU capacity Google actually allocates beyond its own internal demand given how compute-constrained the company remains for its own model training, whether any other frontier lab beyond Anthropic makes a large public TPU commitment, and whether Nvidia responds with more aggressive pricing at the risk of compressing its own margins.

Original source: CNBC