Google’s AI chips are suddenly everywhere except inside Google’s own cloud. OpenAI is reportedly planning to buy Google’s Tensor Processing Units outright, according to The Information, and Oracle has announced it will offer Google TPUs on Oracle Cloud Infrastructure starting in 2027. Together, the two stories show the same shift: Nvidia’s biggest customers want a second source of AI compute, and Google is now willing to sell its custom silicon to rivals.
The important distinction is what is confirmed and what is only reported. The Oracle deal is confirmed: Oracle and Google announced it at Oracle AI World in Las Vegas on October 27, 2026. The OpenAI purchase is not confirmed: The Information reported it citing a person familiar with the matter, OpenAI declined to comment to Reuters, and Google did not immediately respond. No price, quantity, or delivery date has been disclosed for any OpenAI purchase.
What was reported about OpenAI buying Google TPUs
OpenAI already rents Google TPUs. Reuters reported in June 2025 that the ChatGPT maker had begun renting Google’s AI chips, and last week OpenAI said its frontier models would be available through Google Cloud. Moving from renting to buying would be a bigger step, giving OpenAI direct control over hardware it can deploy in its own data centers rather than only inside Google’s cloud.
Why would OpenAI want Google chips at all? It is already one of the largest buyers of Nvidia graphics processors, which it uses to train and run its models. But TPUs have gained traction as a cheaper alternative for many AI workloads, and Google began selling them directly to some customers earlier in 2026. For a company burning through compute as fast as OpenAI, a second supplier means more bargaining power, more available capacity, and less risk if one supplier cannot deliver.
OpenAI is also building its own chips. In June 2026 it unveiled a custom AI chip developed with Broadcom, reportedly codenamed Jalapeno, which it intends to deploy by the end of 2026. Buying TPUs would not replace that effort. It would fill the gap while its own silicon ramps up, and it would give OpenAI experience running a second hardware architecture at scale.
What Oracle and Google actually announced
Oracle will offer Google TPUs through Oracle Cloud Infrastructure in calendar year 2027, ramping up over several years. Oracle is the first hyperscale cloud provider outside Google to offer TPUs to external customers, a notable break from the usual pattern where each cloud giant keeps its custom chips inside its own data centers.
The deployment will use Google’s eighth-generation TPU portfolio: TPU 8t for training, which Google says delivers nearly three times the compute per pod of the previous generation, and TPU 8i for inference and agentic workloads, where a model answers requests and runs multi-step tasks. Oracle will supply the power, cooling, and networking needed to run the chips at scale.
Oracle’s senior vice president for cloud infrastructure, Karan Batta, described the deal as a multi-year, multi-gigawatt partnership, and said Oracle would carry future TPU generations from day one. No initial capacity, region, or price was announced, so customers cannot buy TPU capacity on Oracle Cloud yet. The practical takeaway for now is direction, not a product you can order today.
Oracle made the announcement alongside other AI infrastructure moves at the same event, including integrating OpenAI’s ChatGPT and Codex into its enterprise software and adding Cerebras wafer-scale hardware to its cloud lineup. The message is clear: Oracle wants its cloud to be the place where customers can pick whichever AI hardware fits a given job, rather than being locked to one chip family.
Why Google is selling its chips to rivals
For most of the TPU’s history, Google kept the chips as an internal advantage. Selling them changes the business model. Google now earns revenue from the chips themselves, not just from the cloud services that run on them, and wider adoption makes its software tools more familiar to developers. That matters because the real lock-in in AI is often software: once teams build workflows around a chip’s tools, switching becomes expensive.
Google’s customers are also lining up. Deals with Anthropic and others have driven demand for TPUs, and Alphabet expects to begin recognizing a small share of revenue from TPU sales agreements by the end of 2026, with most of it expected in 2027, according to its chief financial officer. Alphabet was due to report quarterly results on October 28, giving investors their first chance to test whether that chip revenue is actually arriving.
There is also a competitive logic. Amazon, Microsoft, and Alphabet together plan to spend as much as $600 billion on AI infrastructure in 2026, according to Reuters. At that scale, every large AI company wants more than one hardware supplier, and Google would rather be that second supplier than watch customers buy only from Nvidia or build their own chips entirely.
What this means for Nvidia
This is not the end of Nvidia’s dominance. Nvidia remains the default choice for training large models, and OpenAI’s previously announced plans include deploying at least 10 gigawatts of Nvidia systems alongside a framework worth up to $100 billion. TPUs are best seen as a second source, not a replacement, especially for inference work where cost per request matters more than raw training speed.
Still, the direction is uncomfortable for Nvidia. Its largest customers, including OpenAI, Google’s cloud rivals, and other frontier labs, are all investing in alternatives: Google’s TPUs, Amazon’s Trainium, Microsoft’s Maia, Meta’s custom silicon, and OpenAI’s own Broadcom chip. The more credible those alternatives become, the harder it is for any single supplier to dictate price and availability.
For smaller companies, the practical effect will arrive slowly. If you run AI workloads on a cloud platform, the main benefit over the next year or two will be more choice and, hopefully, lower prices as TPUs become available on Oracle Cloud and more providers compete. If you are choosing a cloud today, there is no need to wait for TPU availability on Oracle: the 2027 date, the lack of pricing, and the multi-year ramp mean this is a roadmap item, not a purchasing decision yet.
What to watch next
Three things will show whether this shift is real or just headlines. First, whether OpenAI or Google confirms a purchase, and at what scale. Second, what Alphabet says about TPU revenue in its quarterly results. Third, whether Oracle names regions, prices, and a launch month for TPUs on its cloud during 2027. Until at least one of those lands, treat the OpenAI purchase as a credible report, not a done deal, and treat the Oracle deal as a confirmed partnership with details still to come.
For more on the companies reshaping AI infrastructure, see our coverage of OpenAI Dots and its always-on AI agents, our explainer on Meta Muse for small business, and our look at the the NVIDIA RTX Spark launch. The hardware race behind those products is exactly what the TPU news is about.
Frequently asked questions
Is OpenAI buying Google TPUs?
Not confirmed. The Information reported that OpenAI plans to buy Google TPUs, citing a person familiar with the matter. OpenAI declined to comment and Google did not immediately respond, and no terms were disclosed.
When will Google TPUs be available on Oracle Cloud?
Oracle and Google say TPUs will arrive on Oracle Cloud Infrastructure in calendar year 2027, ramping up over several years. No month, region, or price has been announced.
What are Google TPUs?
TPUs, or Tensor Processing Units, are Google’s custom AI chips, designed for training and running machine learning models. The eighth generation splits into TPU 8t for training and TPU 8i for inference and agentic workloads.
Does this mean OpenAI is dropping Nvidia?
No. OpenAI remains a major Nvidia customer, and the reported TPU purchase would add a second supplier alongside Nvidia GPUs and OpenAI’s own Broadcom chip, not replace them.
Should a small business wait for TPUs on Oracle Cloud?
No. With no launch month or pricing announced, treat this as a 2027 roadmap item. Choose cloud AI services on today’s price, performance, and availability, and revisit the choice once Oracle publishes concrete TPU offerings.