Oracle brings ChatGPT and Codex into its applications and adds Google TPUs and Cerebras to OCI

Oracle used the opening of Oracle AI World in Las Vegas on October 27, 2026 to announce a broader partnership with OpenAI that places ChatGPT, Codex and OpenAI's cyber models inside Oracle's business software and cloud, and, in separate announcements on the same day, to add two new families of AI accelerators to Oracle Cloud Infrastructure (OCI): Google's Tensor Processing Units and Cerebras inference systems. Taken together, the announcements describe a company trying to sit at both ends of enterprise AI, as the place where employees interact with frontier models and as the infrastructure on which those models are trained and served.

The OpenAI side of the news consists of four offerings. Oracle AI Workspace connects ChatGPT to Oracle Fusion Applications, Oracle NetSuite, Oracle AI Database and OCI through plugins, so that employees can query business data and run approved workflows from a conversational interface. Oracle AI Security Mesh is a managed service that uses what Oracle calls "OpenAI frontier cyber models" to find and fix vulnerabilities across OCI, other clouds and on-premises systems. Oracle App Composer lets employees build enterprise applications with Codex while Oracle handles deployment on OCI. Finally, the Oracle AI SMB Bundle packages ChatGPT Business, Codex, the new Workspace and NetSuite QuickStart with $12,000 in Oracle cloud credits per year for small businesses.

On the infrastructure side, Oracle and Google said Google TPUs will be offered on OCI starting in calendar year 2027, which both companies describe as the first time TPUs will be available from a hyperscale cloud other than Google's own. Cerebras and Oracle announced their intention to bring Cerebras inference to OCI on the same 2027 timeline. None of the announcements includes a firm general-availability date for the OpenAI products, and Oracle attaches its standard disclaimer that the timing, features and pricing of future products may change.

··········

WHAT ORACLE AND OPENAI ANNOUNCED.

The central product is Oracle AI Workspace, which Oracle presents as a single user experience combining OpenAI models with Oracle applications and data. According to Oracle, employees describe a business goal in ChatGPT and the plugins assemble the relevant Oracle tools and data to retrieve information, complete approved workflows and take actions in Oracle systems. The systems named are Fusion Applications (Oracle's ERP, HCM, supply chain and customer experience suite), NetSuite, Oracle AI Database and OCI, together with what the release calls "other enterprise tools, and data sources". Oracle says the Workspace keeps "the governance, approvals, and controls required for enterprise operations", but it has not yet published a permissions model, a list of supported workflows or a list of eligible ChatGPT plans beyond the Business tier included in the SMB bundle.

Oracle AI Security Mesh is the most unusual of the four offerings. It is described as a managed service that combines OpenAI cyber models with Oracle's own testing tools, independent validation and testing infrastructure. Customers define where testing is authorized and control its scope, while Oracle says it manages the safeguards intended to reduce the risk of outages or disruption during testing. The release does not name the specific OpenAI models involved, and it does not say whether findings are remediated automatically or handed to customer teams for approval.

Oracle App Composer extends the partnership into software creation. Users build applications with Codex, Oracle deploys and operates them on OCI, and the applications connect to approved Oracle data and workflows and to OCI services including OCI Kubernetes Engine and Oracle Autonomous AI Database. Oracle says enterprise identity, permissions, networking, audit and application lifecycle controls apply to everything built this way. App Composer also supports APEXlang, which Oracle describes as a new technology for fast generative development that avoids producing application code that is hard to maintain, an implicit acknowledgement of the concern that AI-generated internal tools become unmaintainable once their original authors move on.

The quotes in the release frame the deal at the executive level. Clay Magouyrk, Oracle's CEO, said that "OpenAI is transforming how enterprises solve business problems", while Greg Brockman, OpenAI's president and co-founder, described the expanded partnership as a way of putting frontier AI into everyday operations.

··········

WHAT CHANGES FOR CHATGPT, CODEX AND ORACLE CUSTOMERS.

For organizations that already run Oracle Fusion Applications or NetSuite, the practical change is that ChatGPT becomes a supported front end to those systems rather than a separate tool into which employees paste exported data. If Oracle AI Workspace works as described, a finance or procurement user could ask ChatGPT for the status of an approval or start a documented workflow, with Oracle's existing approval chains still deciding whether the action goes through. How far that goes in practice will depend on which workflows Oracle exposes through the plugins and on how ChatGPT workspace administrators and Oracle administrators share control, details Oracle has not yet documented.

For small businesses, the SMB Bundle is the most concrete element. It is presented as free to start, and it includes ChatGPT Business for day-to-day work, Oracle AI Workspace, Codex with App Composer, OCI for deployment and NetSuite QuickStart for core operations such as invoicing and cash flow, plus $12,000 in Oracle cloud credits each year. Oracle notes that terms, conditions and restrictions apply but has not published them, so the number of ChatGPT Business seats included, the eligibility criteria and the length of the offer are still unknown.

........

Offering	What it includes	OpenAI component	Disclosed availability
Oracle AI Workspace	ChatGPT plugins for Fusion, NetSuite, AI Database and OCI	ChatGPT, OpenAI models	No date given
Oracle AI Security Mesh	Managed vulnerability finding and fixing across OCI, other clouds and on-premises	OpenAI cyber models (not named)	No date given
Oracle App Composer	App building with Oracle-managed deployment on OCI, APEXlang support	Codex	No date given
Oracle AI SMB Bundle	ChatGPT Business, Workspace, Codex, App Composer, OCI, NetSuite QuickStart, $12,000 yearly credits	ChatGPT Business, Codex	Terms not published

........

The announcement builds on an earlier arrangement. In June 2026 OpenAI said that eligible Oracle Universal Credits could be applied to OpenAI models and Codex through OCI, letting Oracle customers draw on an existing cloud commitment rather than set up a separate purchasing path with OpenAI. The October announcements move from billing to product integration, which matters for procurement teams because spending on ChatGPT and Codex can increasingly be routed through Oracle contracts. Data Studios has previously covered a comparable move on another cloud, when OpenAI brought GPT-6 models to Google Cloud Model Garden.

··········

GOOGLE TPUS ARRIVE ON ORACLE CLOUD INFRASTRUCTURE.

The infrastructure announcement with the largest strategic weight is the arrival of Google TPUs on OCI. Oracle and Google said the TPUs will be available through OCI in calendar year 2027 and will ramp over several years, with OCI supplying the power, cooling and data center networking needed to deploy them at scale. Google's blog post on the partnership, written by Oracle's Karan Batta and Gurmeet Goindi, says OCI becomes "the first hyperscaler outside Google to offer TPUs" and that the systems are designed for direct liquid cooling at rack scale, the profile OCI's next-generation data centers are built for.

Oracle's release positions TPUs for the whole model lifecycle, from training and post-training of dense and mixture-of-experts models to production inference for reasoning, agentic and long-context applications. It describes Google's eighth TPU generation as two architectures: TPU 8t for training, which Google says delivers nearly 3x the compute performance per pod of the previous generation, and TPU 8i for inference, aimed at workloads in which many specialized agents iterate. Neither company has said which generation will be installed on OCI, in which regions, or at what price.

........

Claim	Figure	Source of the claim
Eighth-generation TPU compute	Up to 12.6 PFLOPS native FP4 per chip	Google Cloud blog
Efficiency vs previous generation	Up to 2x performance per watt	Google Cloud blog
TPU 8t training performance	Nearly 3x compute per pod vs previous generation	Oracle release
Training goodput	Over 97% across thousands of chips	Google Cloud blog
Low-latency serving	Up to 80% better performance per dollar	Google Cloud blog
OCI availability	Calendar year 2027, multi-year ramp	Oracle release

........

All of these figures are vendor claims rather than independent measurements, and the goodput and performance-per-dollar numbers describe Google's systems generally, not a measured OCI deployment. The goodput figure refers to the share of time a training cluster spends on productive computation, which Google attributes to optical circuit switching and interconnect routing that move traffic around faults without restarting jobs. The serving claim is tied to more on-chip SRAM, more HBM and a new interconnect topology that Google says cuts network hops by more than 50%.

Inside OCI, the TPUs will connect to the rest of the cloud through Oracle Acceleron, the networking layer that combines SmartNICs with hardware tenant isolation, a high-performance fabric and Zero Trust Packet Routing. The authors argue that this lets data move between Oracle databases, storage and TPU clusters without consuming host compute, and that fine-tuning and inference can therefore run next to the Oracle databases that often hold a company's system of record. On the software side, Google points to TorchTPU, its native PyTorch stack for TPUs, which it says is now open source, and to support in vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face and LiteRT. The stated aim is that teams can move a workload between GPUs and TPUs, and between Google Cloud and Oracle infrastructure, without rewriting model code.

··········

CEREBRAS INFERENCE JOINS THE OCI LINEUP.

Cerebras and Oracle announced on the same day that they intend to bring Cerebras inference to OCI, also in calendar year 2027. The offering targets AI labs, enterprises and developers building interactive applications, and it adds to OCI's existing inference options, including Dedicated AI Clusters in OCI Enterprise AI. Cerebras attributes its speed to the Wafer-Scale Engine, which places compute and memory on a single processor and reduces the data movement and inter-chip communication that slow token generation in long reasoning chains and multi-step agent workflows.

The Cerebras release is thinner than Google's. It names no supported models, publishes no throughput or latency figures and gives no regions. Andrew Feldman, Cerebras's co-founder and CEO, said that "the faster the inference, the more work AI can do without making users wait", while Mahesh Thiagarajan, president of OCI, described the addition as another ultrafast and deterministic option in Oracle's compute portfolio. Thiagarajan made the same argument in the TPU release, saying that bringing Google TPUs to OCI "gives customers access to multiple leading AI compute architectures in one cloud".

··········

WHY ORACLE IS DIVERSIFYING ITS AI COMPUTE.

Read together, the chip announcements show Oracle trying to turn OCI from a cloud built largely around GPUs into one that offers several accelerator architectures under a single network and security model. The logic Oracle and Google set out is cost per token: each stage of a workload, from training to latency-sensitive serving, could run on whichever accelerator is cheapest or fastest for it, with the enterprise data staying in Oracle databases. That pitch only works if the portability promised by TorchTPU and the open serving stacks holds up for real customer models, something no customer has yet reported on OCI because the capacity does not exist before 2027.

The timing also matters. AI compute remains scarce and expensive, as recent financing deals such as SpaceX's $40 billion GPU lease commitment from Apollo have shown, and offering TPUs and Cerebras systems gives Oracle capacity that does not depend entirely on a single GPU supply chain. For Google, placing TPUs in another company's data centers extends the reach of its accelerators to customers who have standardized on Oracle, without those customers having to move their databases to Google Cloud.

··········

WHAT IS ESTABLISHED AND WHAT REMAINS UNCERTAIN.

The established facts are clear. On October 27, 2026 Oracle announced Oracle AI Workspace, Oracle AI Security Mesh, Oracle App Composer and the Oracle AI SMB Bundle as products built with OpenAI's ChatGPT, Codex and cyber models, with the SMB bundle carrying $12,000 in yearly Oracle cloud credits. Oracle and Google confirmed that Google TPUs will be offered on OCI starting in calendar year 2027, and Cerebras and Oracle confirmed a plan to offer Cerebras inference on OCI on the same timeline.

Much remains undisclosed. Oracle has given no availability dates for the four OpenAI-based products, no list of eligible ChatGPT plans for the Workspace, no names for the OpenAI cyber models used in Security Mesh and no full terms for the SMB bundle. On the infrastructure side, the TPU generation, regions, capacity and pricing for OCI are unannounced, and Cerebras has not said which models it will serve or how fast. The performance numbers in all three announcements are vendor claims about the underlying hardware, not results measured on OCI, and Oracle's own disclaimer warns that the described features should not be relied on for purchasing decisions until they ship.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org