Oracle Sells OpenAI's Tools While OpenAI Anchors Oracle's Backlog

ChatGPT moves into Oracle's back office

ChatGPT and Codex are heading into Oracle's applications, data and cloud under an expanded partnership Oracle announced in Las Vegas on October 27, 2026. It brings OpenAI's models, the ChatGPT chatbot and the Codex developer tool into Oracle's applications, data and cloud. Oracle describes it as a collaboration to "redefine how the world's businesses run."

The work arrives as four new offerings aimed at employees, IT teams, security staff and founders. The pitch is simple: move from AI that answers questions to AI that carries out work across a company's systems. Keep an eye on OpenAI. It turns up on both sides of this deal, as the supplier of these tools and as a customer tied to a large share of Oracle's backlog.

From answering questions to doing the work

Oracle says businesses still struggle to get AI into the heart of their connected systems, and that it can be the single place for that, one that centralizes access to the apps, data, cloud and controls companies depend on.

Its example starts with a supplier delay that puts a customer order at risk. Michelle Meng, an Oracle product marketing director, wrote that fixing it could mean finding a substitute product, checking inventory, confirming transport and updating the order. Oracle AI Workspace is meant to keep that goal in view while it coordinates the apps, agents, data and tools. Oracle's second example is the month-end close, when the finance team checks and locks the month's books. Here Workspace links three Oracle systems, Fusion Financials, EPM (Oracle's planning and reporting software) and NetSuite, and coordinates the approved work across them.

For builders, Oracle says AI already helps employees write apps, but deploying and running them takes extra expertise. In its example, staff build and approve a launch calendar in Codex, then publish it through Oracle App Composer as one shared app. App Composer scans the code for vulnerabilities and deploys it on managed cloud infrastructure. Oracle runs deployment and operations, while IT controls access, approved data, policies and costs, as Oracle's Blair Bozada put it.

All of this is Oracle's own description. What will the enterprise pieces cost, and when can a company switch them on? The sources don't say. And as Oracle itself notes, every new connection that lets an agent do more also gives a weakness elsewhere a longer reach.

An AI that tries the locks before burglars do

Oracle's answer is Oracle AI Security Mesh. Its example is an ordering app with a missing access check, the test that keeps each signed-in user to their own data. A customer logs in to view an order and can also pull up someone else's. Mesh maps the systems a customer has approved, tests attack paths and checks whether one customer can reach another's order. An attack path is the chain of steps an attacker might follow. A separate agent repeats the test to confirm the problem is real.

Does it then rewrite a company's code on its own? Oracle says no. Coding agents help prepare fixes, the customer's team reviews and approves them, and after the changes go in, Mesh retests that the gap is closed and normal activity still works. It also ranks problems by severity, and what the agents learn stays in the customer's private knowledge base to inform later tests.

Oracle says it refined these methods by testing its own systems at scale, and that AI finds weaknesses at a pace manual testing alone cannot match. I like how plain its test questions are: can someone reach data they shouldn't, what else could they get into, and would the fix actually stop them?

The fine print on "free"

Big companies get managed security testing. Founders get the Oracle AI SMB Bundle: conditional one-time OpenAI credits, a free tier of NetSuite QuickStart and renewable annual credits for OCI, Oracle's cloud, with more services promised over time.

The OCI credits are worth $12,000 a year, which sounds like a lot of free hosting for a young company. But they last up to three years in total, subject to eligibility review. Renewal requires an active, independent business that is building or running its product on OCI, has spent earlier credits on approved purposes, keeps its information current and keeps its account in good standing. Usage beyond the credits is billed, and the sources give no rate for it. Nor do they say what a startup pays once the three years end, or once it outgrows QuickStart's free limits. NetSuite only suggests that bigger teams may be better served by full NetSuite. Free to start, then, with the price of staying left blank.

QuickStart is a free, AI-powered finance tool for early-stage U.S. tech startups, built on Oracle Autonomous AI Database and OCI. It captures bills, sorts expenses and reconciles payments in the background, shows cash, burn rate and runway (how fast money goes out and how long it lasts), and asks for approval when a step needs it.

Free means free within limits. Each month a startup can create up to 2,500 business records and send up to 800 emails. What goes unused does not roll over. Storage stops at 20GB, fine for a stack of PDFs, not for a video archive. Only four bank and card accounts can be connected. Connected tools can have at most two requests to QuickStart running at the same time, a cap on simultaneous traffic rather than on how many tools a startup links. So a partnership pitched as redefining how businesses run lets each tiny one send 800 emails a month from QuickStart. To be fair, that is the stated target: NetSuite aims QuickStart at teams of typically five or fewer.

Whose chat window is it?

Another detail on the QuickStart page: it connects to ChatGPT and Claude through MCP, an open standard for plugging AI assistants into outside software. Oracle framed a separate announcement at the same show differently. On October 26, 2026, it unveiled 25 agentic applications, AI software that carries out multistep tasks toward a goal a person sets, running on an execution engine called Fusion Claw. They cover HR, finance, supply chain and customer management, with jobs such as payroll management and scheduling deliveries.

Oracle co-CEO Mike Sicilia said in his opening keynote that AI should let people barely think about the software at all. Fusion Companion takes what users ask for and uses business context, data and permissions to coordinate the agents, workflows and approvals behind it.

On paper, Companion's list is shorter. The Fusion release names only ChatGPT as a supported assistant for it, while QuickStart, the startup finance tool, also names Claude. Futurum says support for Microsoft Copilot, Gemini or Claude would show how open Oracle means its access layer to be. Oracle is not first in line, either. Salesforce announced Agentforce access inside ChatGPT a year ago. SAP, Workday and Microsoft want their apps reachable through assistants too. Who owns the front door is still open.

The best reference customer is Oracle

If the assistant becomes the front door, you want proof that people walk through it. Oracle co-CEO Clay Magouyrk and OpenAI CEO Sam Altman kept the keynote on adoption, and Altman said GPT-6 drew OpenAI's strongest adoption of any new model.

The hardest numbers came from an OpenAI case study, and its customer is Oracle. It counts 130,000 active ChatGPT users and more than 95,000 Codex users inside the company, and says Oracle's recruiting team cut research time by 98%. They come from OpenAI, and the sources offer no independent check. The reference customer just happens to be the seller.

The October 26 Fusion release named no customers running its agentic apps in production and disclosed no pricing, and the sources show no independent production results for the new OpenAI suite. And money never came up. As BeInCrypto reported, neither CEO said how Oracle funds its $664 billion contract backlog.

The same partner on both sides of the ledger

The software rides on hardware. CNBC reported that Oracle will offer Jalapeño, an OpenAI chip built to run trained AI models, on OCI, with Cerebras accelerators arriving in 2027. In August, OpenAI said Jalapeño runs some open-source models 1.5 times faster than competing silicon.

Oracle's backlog, revenue signed but not yet earned, reached $664 billion, up $209 billion from the year before, and BeInCrypto reports that about half of it is tied to OpenAI. In its fiscal first quarter, Oracle spent $5 billion more on capital projects than its business brought in. CNBC put that quarter's capital spending at $28.5 billion, more than three times what it spent a year earlier. In the same quarter, Oracle sold $20 billion of common stock. OCI is growing fast too: its revenue more than doubled from a year earlier, to $7.4 billion for the quarter.

I think the products look sensibly fenced. In Security Mesh, for one, a team approves each fix, and the system retests it to check that the gap is closed and normal work still runs. But the partnership is a loop: Oracle resells OpenAI's tools while OpenAI is the customer behind about half of Oracle's backlog. The software story reassures only while OpenAI keeps paying for the capacity. If that demand slows, BeInCrypto warns, the reliance becomes the headline.

So here is my question. Could selling software through ChatGPT ever earn enough to matter next to a datacenter bill this size, when the sources do not yet say what any enterprise piece will cost?