Google Opens Its AI Chips to Oracle, Taking Aim at Nvidia’s Grip

Google Opens Its AI Chips to Oracle, Taking Aim at Nvidia’s Grip
Google began developing its Tensor Processing Units in 2013 for its own services, building a vertically integrated AI stack around Search, models and cloud infrastructure. That approach kept the chips largely inside Google’s walls—even as Nvidia came to dominate the wider AI-computing market.

The strategy started to shift in the second quarter, when Google opened TPUs to outside data centers. On Tuesday, the company and Oracle went further: Oracle will offer the chips to its cloud customers beginning next year, the first time a hyperscaler other than Google will provide access to TPUs. The agreement turns a once-internal technology into a direct distribution play against Nvidia’s GPUs.

For Google, the opportunity is customer choice as companies spread workloads across several chip suppliers. Anthropic already uses Google TPUs alongside Amazon Trainium, AMD and Nvidia hardware, saying it selects silicon on price-performance and efficiency for each workload. OpenAI is also planning to buy TPUs, according to reporting cited by the companies’ watchers.

The bullish case is substantial. Google CFO Anat Ashkenazi has said TPU revenue should ramp meaningfully in 2027, while cloud chief Thomas Kurian called the chips a meaningful contributor to third-quarter results. Analysts at TD Cowen and Wells Fargo project tens of billions of dollars in annual TPU sales as the business matures.

But the expansion has a price. Kurian has argued that Google’s own servers pay back in under two years, with systems using its silicon doing so in roughly half that time because the company controls the full stack. Selling through outside partners may broaden demand, yet TD Cowen estimates external TPU sales carry a roughly 20% operating margin, below Google Cloud’s approximately 35% margin. Nvidia remains vastly larger, but Google is now testing whether reach can outweigh the economics of keeping its best chips close to home.

https://foxvector.com