Nvidia Faces Google TPU Push as Oracle Opens Its Cloud

What happened

Nvidia's dominance in AI accelerators is facing its most organized challenge yet. Google is taking its tensor processing units (TPUs) beyond its own data centers. Oracle has agreed to offer the chips on Oracle Cloud Infrastructure (OCI) starting in 2027 48. Oracle announced the deal on October 27 at its AI World conference in Las Vegas 10. It is the first time a hyperscale cloud other than Google's has sold access to TPUs 910.

The timing was deliberate. Alphabet made the news a day before its third-quarter earnings report 12. The Information separately reported that OpenAI, long one of Nvidia's marquee customers, plans to buy Google's TPUs 6.

Google has been developing its own silicon since 2013 1. However, it only began selling TPUs to outside customers in the second quarter of this year, including to so-called neo-cloud providers 14.

The strategy shift

For years, Google sold its chips, models and cloud services as one integrated package. Amazon and Microsoft tried to copy that full-stack pitch 15. Routing TPUs through a rival cloud breaks from that model.

The effort is led by Google Cloud chief Thomas Kurian, who spent more than two decades at Oracle before joining Google in 2018 17. Some outlets describe a broader channel push involving system integrators and distributors, not just Oracle 3. Most reporting, though, centers on the Oracle agreement as the concrete milestone 149.

Amin Vahdat, Google's CTO for AI infrastructure, said the Oracle arrangement would let more organizations use TPUs. He added that it would give them more flexibility in building complex AI applications 4. Under the deal, Oracle supplies the power, cooling and data-center networking, while the TPU systems keep Google's chip-to-chip interconnect 8.

Oracle will add Google's newer TPU 8t and 8i chips to its platform 9. Pricing, regions and initial capacity have not been disclosed. The companies say capacity will ramp over several years 8.

Why it matters for Nvidia

The numbers behind Google's push are getting large. Analysts expect Google Cloud revenue to grow roughly 90% year over year, after 82% growth in the prior quarter 2. Projections cited in several reports put TPU revenue at $9.4 billion in the second half of 2026 and $62.5 billion in 2027 57. Anthropic plans to use up to one million TPUs, supported by a $63 billion financing vehicle that Broadcom described and that Blackstone and three major banks are backing 7.

"It's very possible that they could eventually take market share from Nvidia," said Chris Ballard of Check Capital Management 2.

Price is the second pressure point. In September, Google published on-demand pricing for its seventh-generation Ironwood TPU:

• $12.00 per chip-hour in Iowa
• $13.20 per chip-hour in London

That makes the comparison against Nvidia hardware a simple spreadsheet exercise 11. Meanwhile, Nvidia's large customers are reportedly being told to expect higher costs on Grace Blackwell and Vera Rubin systems as memory prices rise 11.

Independent benchmarks support Google's cost pitch, with caveats. SemiAnalysis found Ironwood 18.5% cheaper per token than Nvidia's B200 and 34.4% cheaper than the B300. But under heavy load of 256 concurrent users, Ironwood took 2.25 times as long as the B300 to return the first token 12. Latency-sensitive applications may still favor Nvidia.

Software may matter more than either result. TorchTPU, developed with Meta, lets PyTorch run natively on TPUs 12. That chips away at the developer lock-in that has long protected Nvidia's position.

The reading: a multi-vendor future, not an Nvidia collapse

Nvidia is not being displaced, and Oracle's own announcements show why. At the same event, Oracle added Nvidia's Vera Rubin alongside chips from AMD, Intel, SambaNova and Positron. It also brought OpenAI's ChatGPT and Codex into its business applications 10. TechRadar quoted Oracle's view that the next generation of AI will require "a diverse range of compute" 9.

As I read it, the shift is about Nvidia's leverage more than its leadership. When customers can choose TPUs inside the same cloud that already holds their data 8, Nvidia's pricing power faces a real alternative for the first time. Cheaper inference and PyTorch support add to that pressure.

Nvidia remains the default for many workloads, especially where response speed is critical. But the projected TPU revenue figures, if they hold, suggest the AI chip market is moving from a near-monopoly toward a genuine competition. One outlet even framed the move as a challenge extending into crypto-related computing 5, though that angle is thinly supported.

The clearest signals to watch are Oracle's eventual TPU pricing and capacity, and how much of OpenAI's spending actually moves away from Nvidia.