Google TPU Sales Expand to Oracle Cloud in Challenge to Nvidia

Google opens the TPU to a rival cloud

For most of a decade, Google's tensor processing units stayed in one place. Anyone who wanted them had to rent them inside Google Cloud, and that exclusivity was a big part of the pitch for Google's integrated stack. That changed on October 27. Oracle said it will offer Google's TPUs through Oracle Cloud Infrastructure, which makes it the first hyperscale cloud other than Google to sell access to the chips.19 The announcement came at Oracle AI World in Las Vegas, one day before Alphabet's third-quarter earnings.101

The chips will not show up right away. Oracle plans to bring TPUs online in calendar 2027 and add capacity over several years. Neither company has said which regions will get them, what they will cost, or how much capacity will be available at launch.810 The launch hardware will be Google's eighth generation, which comes in two versions: the TPU 8t for training and the TPU 8i for inference.9 Oracle will supply the power, cooling and data-center networking, while the TPU systems keep Google's own chip-to-chip interconnect.8 An Oracle executive described the arrangement as a multi-year, multi-gigawatt partnership and said Oracle would get each new TPU generation on the day Google launches it.10

OpenAI is the other part of the story. The Information reported that OpenAI, a longtime Nvidia customer, plans to buy Google's TPUs, and CNBC later confirmed it.62 In one week, Google's chip business picked up a new distribution channel and, reportedly, the most prominent AI lab that wasn't already a customer.

A playbook rewritten in stages

The Oracle deal looks less like a sudden pivot and more like the next step in a gradual opening. Google started designing its own silicon in 2013 but did not sell TPUs to outside customers until the second quarter of 2026.1 In April it said it would sell TPUs to selected third-party data center operators, its formal entry into the merchant accelerator market that Nvidia dominates.21 In July, CFO Anat Ashkenazi confirmed that Google had booked revenue from TPU systems delivered directly to customer data centers for the first time, though she said the financial impact so far was small.26

Other pieces were already moving. Google reached a deal for London-based Fluidstack to host TPUs in a New York data center, and it reportedly approached other neoclouds, including CoreWeave.27 Google is also building a TPU cloud called Crux AI with Blackstone.10 Because of these arrangements, the coverage is not unanimous about how much of a first the Oracle deal is. Most outlets call it the first time TPUs will be available outside Google's infrastructure.17 The Next Web points out that Oracle may not be first in practice, since TPUs are already deployed or being deployed through other partners.10 The fairest reading is that Oracle is the first rival hyperscaler to resell TPUs, not the first outside home for them. That still matters, because Oracle has been one of Nvidia's most aggressive buyers.28

CNBC frames the shift partly around Google Cloud chief Thomas Kurian, who spent more than 20 years at Oracle before joining Google in 2018.1 One Alphabet shareholder, Harding Loevner's Uday Cheruvu, put the change in culture bluntly: Google used to excel at its own projects but struggled to serve customers, and he said that mindset has shifted.1

The cost argument behind the chips

The main reason customers want TPUs is cost, especially for inference. SemiAnalysis estimates that a full Ironwood TPU deployment costs about 44% less to own and operate per chip than an Nvidia GB200 server.1 The firm explains the gap by pointing to margins: Google pays Broadcom a substantial markup, but that is still much less than what Nvidia earns on complete systems, including CPUs, switches, networking and cabling. Kurian has argued along the same lines. In September he said Google's AI servers pay for themselves in under two years, and servers built on its own silicon do so in roughly half that time.2

Independent inference benchmarks help the case, with caveats. In SemiAnalysis's InferenceX testing at 100 tokens per second per user, Ironwood cost about $0.181 per million tokens, compared with $0.222 for Nvidia's B200 and $0.276 for the B300.18 At a 20-second median response time the gap shrinks to roughly 8% against the B200, and the B200 comes out ahead around a 30-second response time.20 Ironwood also lacks native FP4 support, and its advantage disappears when Nvidia GPUs run at that precision.14 Latency is another weak spot. At 256 concurrent users, Ironwood took 2.25 times as long as the B300 to return the first token, the delay a chatbot user actually notices.12

List prices complicate things further. Google's published on-demand rate for Ironwood is $12 per chip-hour in Iowa, which is higher than at least one marketplace's on-demand B200 price, though Google's one- and three-year commitment rates come in lower.11 Large buyers pay far less than list. One SemiAnalysis estimate puts Anthropic's negotiated rate at about $1.60 per TPU-hour.11 In other words, TPU economics look strongest for big, steady, high-volume inference and weakest for small buyers or latency-sensitive chat. That is why the Oracle channel and the OpenAI talks matter more than any single benchmark.

Nvidia still sets the reference point

Despite the attention on TPUs, none of the parties involved is presenting this as a break from Nvidia. At the same event, Oracle said Nvidia's Vera Rubin NVL72 systems are now generally available on OCI and that it plans to grow its Rubin clusters to several gigawatts.10 Asked whether the TPU deal signaled a move away from Nvidia, OCI President Mahesh Thiagarajan said no and called Nvidia's chips the best.24 His explanation was scarcity: chips remain hard to get, so customers benefit from having more options.10 Oracle also added Intel, AMD, SambaNova, Positron and Cerebras hardware to its lineup.10

Nvidia is also responding on inference. Jensen Huang has said its co-designed rack systems can deliver up to 35 times higher throughput per megawatt on trillion-parameter models.21 Still, there are cracks. A reported one-quarter delay in the Rubin ramp led TrendForce to cut its estimate of Rubin's share of Nvidia's 2026 high-end GPU shipments from 29% to 22%.21 Nvidia customers have also reportedly been warned that Grace Blackwell and Vera Rubin server builds could cost more than 15% extra from early 2027 because memory prices are rising.11 Google is exposed to the same memory market, but buyers who lock in multi-year TPU commitments get some protection from price swings.11

Margins and the competitor problem

The expansion has costs for Google. TD Cowen estimates a 20% operating margin on TPU sales, compared with Google Cloud's overall margin of about 35% last quarter, which means faster TPU growth could dilute profitability.1 The revenue forecasts vary widely. TD Cowen models $9.4 billion in TPU sales for the second half of 2026 and $62.5 billion in 2027, while Wells Fargo expects about $9 billion in the third quarter alone and $99 billion next year.1 Ashkenazi has said the vast majority of TPU revenue from existing deals will be recognized in 2027.1 The size of the gap between analysts shows how uncertain they are about how quickly committed capacity will turn into recognized sales.

The customer list also creates strategic tension. Google's biggest TPU customer is Anthropic, whose Claude models compete directly with Gemini. Mizuho has described the relationship as Google arming its closest model competitor.2 Financing is becoming part of the business too. Broadcom detailed a $63 billion vehicle, backed by Blackstone and three banks, that will buy TPUs and networking gear and lease them to Anthropic for five years. William Blair analysts estimate it covers only about half of Anthropic's 2027 needs.2 Adding OpenAI would put both of Gemini's main rivals on Google silicon.

Skeptics remain. One Register columnist characterized much of the week's dealmaking as circular spending among the largest AI players.10 A more practical concern is resale: Nvidia GPUs have a deep secondary market supported by CUDA, and TPUs do not yet have one. Software is a related weakness. The external TPU developer community is far smaller than CUDA's, though Google has been working to close that gap with native PyTorch support.29

The reading

This is a real turning point, but the sources point to a slower one than the more excited "Nvidia killer" framing suggests. The cost advantages hold up for large, predictable inference workloads, and the biggest labs are buying. But Oracle's capacity does not arrive until 2027, Nvidia remains the default even at the clouds adding TPUs, and each external TPU sold earns Google a thinner margin than its cloud business. In the near term, Google probably won't displace Nvidia. The more likely result is that multi-vendor buying becomes standard practice, which lets the largest customers negotiate inference prices down. That shift alone could change who captures the margins in AI hardware.