On October 27, 2026, two things happened on the same day: Oracle announced that it would offer Google TPU chips on its cloud platform starting in early 2027—the first time Google TPUs have landed on a hyperscale cloud platform outside its own cloud system; a few hours later, The Information and Bloomberg reported that OpenAI plans to directly purchase Google's tensor processing units (TPUs). The specific procurement scale, price, and timetable have not been determined, and both companies declined to comment.
Google is turning TPUs from an internal compute asset into a chip business serving the entire industry, and its top potential buyer is its most direct competitor in AI products.
A Compute Deal Between Competitors
OpenAI and Google compete across the board in large-model assistants, search engines, and enterprise AI tools. According to Reuters, OpenAI began renting Google's AI chip compute resources as early as June 2025; on October 21 this year, OpenAI announced that it would offer its frontier models to users through Google Cloud. The potential direct TPU purchase is another extension of this collaboration line—from renting cloud compute to directly acquiring chip assets.
Such relationships are not uncommon in the tech industry, but the structural tension between OpenAI and Google is particularly pronounced: Google DeepMind and OpenAI are both competing for the same top AI researchers, GPT and Gemini are directly competing on capabilities, and Google's TPU business needs a major buyer like OpenAI to prove its commercial value.
Why Now: The Pressure of Inference Costs
Training a large model is a one-time massive expense that can be procured centrally and negotiated on price; serving a model to hundreds of millions of users, by contrast, is an ongoing cost that burns compute every hour. According to Gadgets Now, “training is a large bill paid in installments, while inference is a bill that arrives every hour”—for a service provider of OpenAI's scale, the latter cost pressure is more persistent and harder to compress.
Nvidia GPUs have an indisputable advantage in training tasks, but in inference scenarios, TPUs—with their specialized architecture and lower cost per unit of compute—are becoming an alternative for some companies. Google has used TPUs internally for years to run Gemini and search inference tasks, accumulating extensive optimization experience. If OpenAI can migrate part of its inference workload to TPUs, it could in theory lower the cost of each API call and thus gain more room in pricing competition.
TPUs differ significantly from Nvidia GPUs in software ecosystem, model adaptation, and engineering operations. OpenAI is recruiting engineers to optimize frontier models specifically for TPUs, which itself shows that migration costs are not low.
Nvidia's Position: Harmed but Not Shaken
The impact of this deal on Nvidia is real, but it needs to be viewed quantitatively. According to multiple media reports, Nvidia invested about $30 billion in OpenAI in early 2026, participating in OpenAI's roughly $110 billion financing round, and the two sides also signed a large-scale agreement to deploy Nvidia's latest Vera Rubin systems. OpenAI has so far lined up a roster of six chip suppliers: Nvidia, AMD, Broadcom, Cerebras, Amazon, and Google if the deal goes through.
This shows that Nvidia is not being replaced but diluted. OpenAI's strategy is to build multiple positions and spread risk, avoiding constraints from a single supplier in capacity, price, or technology roadmap. For Nvidia, the real risk is not losing OpenAI as a customer, but losing the pricing power of the equation that “AI infrastructure equals Nvidia GPUs.” Once TPUs are validated as a viable alternative, competitors such as Google, AMD, and Intel will have a real reference point with which to lever negotiations.
Google's Chip Commercialization Transition
Google has developed its own TPUs for more than a decade, but for a long time they were limited to internal use. This year, Google began formally selling TPUs directly to some external customers, a strategic shift in its business model. Alphabet CFO Anat Ashkenazi previously said the company expects to begin recognizing revenue from some TPU sales agreements before the end of 2026, with most related revenue expected to be recognized in 2027.
Research firm TD Cowen forecasts that Google's TPU sales will reach about $9.4 billion in the second half of 2026 and could jump to $62.5 billion in 2027. This forecast has not yet been verified, but the magnitude itself shows that market expectations for TPU commercialization are expanding rapidly. Oracle's announcement that it will offer Google TPUs on its own cloud platform is a new step by Google to expand TPU reach through partner channels, and a direct response to user concerns that TPUs can only be used on Google Cloud.
OpenAI's In-House Chip Path Advances in Parallel
OpenAI is not only working on external procurement. According to a report cited by TradingKey, OpenAI is collaborating with Broadcom to develop a custom AI processor codenamed “Jalapeno,” planned for deployment later this year, as part of OpenAI's multi-generation in-house chip roadmap. OpenAI's chip strategy runs on three parallel tracks: continuing to buy Nvidia as the core compute foundation, bringing in Google TPUs and other suppliers to share inference costs, and developing its own chips to reduce long-term external dependence.
The endpoint of this path is for OpenAI to gradually gain pricing power and supply-chain resilience at the chip level. The short-term cost is enormous engineering complexity—hardware with different architectures requires different optimization paths, and the marginal cost of model migration is not low.
One Assessment
This TPU purchase is essentially OpenAI using concrete procurement action to endorse the usability of Google TPUs. The signal that a leading model developer is willing to pay for a competitor's chips will itself accelerate the entire industry's reassessment of compute supply strategies. But a procurement intention is still a considerable distance from actual deployment: OpenAI needs to complete model optimization for TPUs, and Google needs to prove that its TPUs can meet the requirements of large-scale production deployment in engineering support and supply reliability.
Nvidia's dominant position in the AI chip market will not be shaken by this, but what it is losing is the absolute bargaining advantage of “no alternative.” As major model developers such as OpenAI and Anthropic (which already has deep cooperation with Google) vote for TPUs, the market is moving from a unipolar structure to an oligopoly of coexisting players. For the entire AI compute cost structure, this may be a more profound change than any single technological breakthrough.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接