InfrastructureCompute, chips & cloud
Oracle to offer Google TPUs on its cloud from 2027 as OpenAI reportedly plans to buy the chips
Google is moving its custom AI chips beyond its own cloud. Oracle Cloud Infrastructure will offer TPUs starting in 2027, after Google began selling TPU systems into customer data centers. OpenAI reportedly plans to buy TPUs, joining Anthropic's large commitment.
From the CNBC · Technology article of October 28, 2026 and cited sources: TPUs now reach buyers through three channels: Google Cloud, TPU systems installed in customer data centers (first revenue came in Q2 2026) and, from 2027, Oracle's cloud. [2] [3] [7] [9]
SOURCES[2] Google Cloud Press Corner: Oracle Brings Google TPUs to OCI googlecloudpresscorner.com[3] Constellation Research: Google Cloud starts shipping TPU… constellationr.com[7] FourWeekMBA: Google sells TPU systems into customer data… fourweekmba.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- TPUs now reach buyers through three channels: Google Cloud, TPU systems installed in customer data centers (first revenue came in Q2 2026) and, from 2027, Oracle's cloud. [2] [3] [7] [9]
- Original source
- CNBC · Technology article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 28, 2026 article published
- Covered
- Oct 28 edition
- Evidence
- 13 sources across 12 sites
- 1CNBC: Google alters its cloud playbook, enlisting partners…cnbc.com
- 2Google Cloud Press Corner: Oracle Brings Google TPUs to OCIgooglecloudpresscorner.com
- 3Constellation Research: Google Cloud starts shipping TPU…constellationr.com
- 4Anthropic: Expanding our use of Google Cloud TPUsanthropic.com
- 5DCD: Google and Anthropic confirm 1GW dealdatacenterdynamics.com
- 6Tom's Hardware: Google splits its TPU into two chipstomshardware.com
Reach: not measured yet
TPUs now reach buyers through three channels: Google Cloud, TPU systems installed in customer data centers (first revenue came in Q2 2026) and, from 2027, Oracle's cloud. Oracle is the first other hyperscale cloud to sell TPU access. It will supply power, cooling, networking and hardware-level isolation between tenants, with capacity ramping over several years. Pricing and named customers have not been disclosed. [2] [3] [7] [9]
Google's two biggest rivals in AI models are now TPU customers. Anthropic committed to as many as one million TPUs, with more than a gigawatt of capacity coming online in 2026. OpenAI, which has rented TPUs through Google Cloud since 2025, reportedly plans to buy them and is hiring engineers to optimize its models for the chips. [1] [4] [5] [12]
The eighth-generation TPU is split into two chips. The TPU 8t is for training and scales to up to 9,600 chips per job. The TPU 8i is for inference, with more high-bandwidth memory and on-chip SRAM to ease memory bottlenecks. Google's figures of about 2.7x better training price-performance and 80% better inference price-performance than the previous generation are vendor claims with no independent benchmark. [6] [10] [11]
The economics favor Google but are not settled. SemiAnalysis estimates that Ironwood TPUs cost Google about 44% less to own and operate than Nvidia GB200 servers, but that is Google's internal cost, not a rental price. TD Cowen expects external TPU sales to earn lower margins than Google Cloud overall, and 2027 TPU revenue forecasts range from $62.5B to about $99B. [1] [13]
TPUs are becoming available across several clouds and on-premises, and Anthropic is a major buyer. Implication: practitioners gain a credible alternative to Nvidia and more bargaining leverage, though most volume arrives in 2027.
Executive brief
Oracle Cloud Infrastructure (OCI) will offer Google's tensor processing units (TPUs) starting in 2027. This is the first time another hyperscale cloud will sell access to Google's custom AI chips (Google Cloud Press Corner). Google kept TPUs inside its own cloud for more than a decade, then began selling TPU systems for customers' own data centers in Q2 2026 (Constellation Research). On the same day as the Oracle news, The Information reported that OpenAI plans to buy TPUs, which CNBC then confirmed (CNBC). Google's two main rivals in AI models now both depend on its chips, and that gives Nvidia a credible competitor at large scale.
What changed and event timeline
Anthropic commits to up to 1M TPUs
Anthropic expanded its use of Google Cloud to as many as one million TPUs, with well over a gigawatt of capacity coming online in 2026. The deal is worth tens of billions of dollars (;).
Eighth-generation TPU splits in two
At Cloud Next, Google announced two chips for the first time: TPU 8t for training and TPU 8i for inference ().
First TPU system revenue
Google Cloud recorded its first revenue from TPU systems delivered to customer data centers, with Blackstone named as an example. Cloud revenue rose 82% to $24.8B, and most TPU revenue is expected in 2027 (;).
$63B lease vehicle for Anthropic
Bank of America, Citi, Morgan Stanley and Blackstone are financing a special-purpose vehicle that will buy Broadcom accelerators and networking gear and lease them to Anthropic. Deliveries start in 2027 ().
Oracle adds TPUs; OpenAI reportedly buying
At Oracle AI World, Oracle said OCI will offer TPUs from 2027, ramping over several years. The same day, a report said OpenAI plans to buy TPUs (;).
Capabilities and access
- Products: TPU 8t (training) and TPU 8i (inference). Google says the 8t delivers nearly 3x the compute per pod of the previous generation (Google Cloud Press Corner).
- Ways to buy: Google Cloud; TPU systems installed in customer data centers (since Q2 2026); OCI from calendar 2027.
- On OCI: Oracle supplies the power, cooling and data-center networking. Its Acceleron SmartNIC provides hardware-level isolation between tenants.
Read the full section
- Products: TPU 8t (training) and TPU 8i (inference). Google says the 8t delivers nearly 3x the compute per pod of the previous generation (Google Cloud Press Corner).
- Ways to buy: Google Cloud; TPU systems installed in customer data centers (since Q2 2026); OCI from calendar 2027.
- On OCI: Oracle supplies the power, cooling and data-center networking. Its Acceleron SmartNIC provides hardware-level isolation between tenants. Oracle is also adding Cerebras accelerators and OpenAI's "Jalapeño" inference chip (SiliconANGLE).
- Not disclosed: pricing, named OCI customers and OpenAI's deal terms.
Technical analysis for researchers and developers
- Separate chips for training and inference.
- Workloads Google targets: dense and mixture-of-experts training, post-training, and long-context reasoning inference (Google Cloud Press Corner).
- Porting work: OpenAI is reportedly hiring engineers to optimize frontier models for TPUs, which suggests GPU-to-TPU porting still takes dedicated effort (FourWeekMBA).
Read the full section
- Separate chips for training and inference. The TPU 8t reportedly has 216 GB of high-bandwidth memory (HBM), 128 MB of on-chip SRAM and up to 9,600 chips per training job. The TPU 8i reportedly has 288 GB of HBM, 384 MB of SRAM and a new "Boardfly" network topology aimed at the memory bottleneck in inference (GIGAZINE; NAND Research).
- Workloads Google targets: dense and mixture-of-experts training, post-training, and long-context reasoning inference (Google Cloud Press Corner).
- Porting work: OpenAI is reportedly hiring engineers to optimize frontier models for TPUs, which suggests GPU-to-TPU porting still takes dedicated effort (FourWeekMBA).
- Missing details: the release does not say which frameworks OCI will support.
Claims and evidence
- Price-performance: Google says TPU 8t has about 2.7x better training price-performance than Ironwood, and TPU 8i about 80% better inference price-performance.
- Cost versus Nvidia: SemiAnalysis estimates that an Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server.
- Payback: Thomas Kurian, Google's cloud chief, says servers built on Google's own chips pay for themselves in roughly half of the under-two-year payback of its AI servers overall (CNBC).
Read the full section
- Price-performance: Google says TPU 8t has about 2.7x better training price-performance than Ironwood, and TPU 8i about 80% better inference price-performance. Both figures are vendor-reported, and no independent benchmark appears in the reviewed sources (NAND Research).
- Cost versus Nvidia: SemiAnalysis estimates that an Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server. This is Google's internal cost, not what a customer pays to rent one (CNBC; SemiAnalysis glossary).
- Payback: Thomas Kurian, Google's cloud chief, says servers built on Google's own chips pay for themselves in roughly half of the under-two-year payback of its AI servers overall (CNBC).
- TPU revenue forecasts vary widely: TD Cowen expects $62.5B in 2027, and Wells Fargo about $99B (CNBC).
Context and prior work
- Google began developing in-house chips in 2013 (CNBC).
- OpenAI has rented TPUs through Google Cloud since 2025 (FourWeekMBA).
- Anthropic's IPO filing lists chips from Google, Amazon, Nvidia and AMD (CNBC).
Read the full section
- Google began developing in-house chips in 2013 (CNBC).
- OpenAI has rented TPUs through Google Cloud since 2025 (FourWeekMBA).
- Anthropic's IPO filing lists chips from Google, Amazon, Nvidia and AMD (CNBC).
- Q2 2026 cloud revenue: Google $25B, AWS $42B, Azure $29B. Google is growing fastest (CNBC).
Limitations, safety and contested findings
- "First outside Google" is worded inconsistently. SiliconANGLE describes it more narrowly, as the first supply to another hyperscaler (SiliconANGLE).
- The $63B hardware may not be TPUs. CNBC says the vehicle buys TPUs. Reuters describes it only as "Broadcom's AI accelerators" (Investing.com/Reuters).
- Lower margin: TD Cowen estimates a 20% operating margin on external TPU sales, against about 35% for Google Cloud overall (CNBC).
Read the full section
- "First outside Google" is worded inconsistently. CNBC calls the Oracle deal the first TPU access outside Google's infrastructure, yet also reports Q2 sales into outside data centers. SiliconANGLE describes it more narrowly, as the first supply to another hyperscaler (SiliconANGLE).
- The $63B hardware may not be TPUs. CNBC says the vehicle buys TPUs. Reuters describes it only as "Broadcom's AI accelerators" (Investing.com/Reuters).
- Lower margin: TD Cowen estimates a 20% operating margin on external TPU sales, against about 35% for Google Cloud overall (CNBC).
- Google's own needs come first: CEO Sundar Pichai said frontier AGI work gets TPU capacity before cloud customers (Constellation Research).
Business and practitioner implications
- Real alternative to Nvidia: TPUs can now be used on Google Cloud, on OCI or on-premises, which gives buyers negotiating leverage and room to run on more than one platform.
- Debt-financed compute: Lease structures with residual-value guarantees from the chip vendor, like the $63B Broadcom vehicle, are becoming a template for paying for AI hardware (Investing.com/Reuters).
- Google's trade-off: Mizuho describes the Anthropic relationship as Google "arming its closest model competitor." Google gains revenue but supplies capacity to rivals (CNBC).
Read the full section
- Real alternative to Nvidia: TPUs can now be used on Google Cloud, on OCI or on-premises, which gives buyers negotiating leverage and room to run on more than one platform.
- Debt-financed compute: Lease structures with residual-value guarantees from the chip vendor, like the $63B Broadcom vehicle, are becoming a template for paying for AI hardware (Investing.com/Reuters).
- Google's trade-off: Mizuho describes the Anthropic relationship as Google "arming its closest model competitor." Google gains revenue but supplies capacity to rivals (CNBC).
- Timing: For most buyers, meaningful volume does not arrive until 2027.