Fusion Apps & OCI Snapshots:
Agents That Finish the Work,
Capacity That Pays for It

What did Oracle show at AI World 2026? Oracle is running at least two businesses at very different speeds, and AI World has been largely about connecting them.

 

Oracle Cloud Infrastructure (OCI) revenue grew 121% to $7.4B in the June–August 2026 quarter. Cloud applications grew 10% and Fusion 14%. The product sessions made the case that Fusion is where Oracle's AI capacity turns into governed business work.

​

The advantage will go to companies that write down how they run, because agents can only follow the policies they're given. Gartner predicts that 33% of enterprise software applications will include agentic AI by 2028, and that at least 15% of day-to-day work decisions will be made autonomously. Finance is already behind: 63% of finance organizations say AI implementation has been slower than expected.

What is new in Fusion?

 

Agents that do the heavy computation themselves. Chris Leone, EVP of Applications Development, described Fusion Claw: a governed runtime that runs code next to the ledger, the supply chain or the workforce data. It reconciles, optimizes and simulates across very large data sets. Oracle also published Fusion setup and configuration as AI skills in a public Git repository. That makes the mechanical part of implementation something any partner or customer can automate.

Our prognosis:

• Fusion. Fusion should gain ground on SAP and Workday in data-heavy finance, supply chain and HR work. In those areas, "close to the data" is a technical requirement as much as a selling point.

• OCI. OCI's outlook rests on converting a $664B contracted backlog into revenue. Capital spending outran operating cash flow last quarter, leaving free cash flow at about –$5.4B, so buyers and investors should watch execution, not ambition.

Fusion Applications: agents that do the work

Oracle's message on Fusion has shifted from agents that help to applications that finish the work. Leone's Q&A session explained how, and it covered some ground beyond what was in the keynote.

What is Fusion Claw, and why does it matter?

Fusion Claw is a governed runtime that lets Fusion's agentic applications reason, compute and act on very large data sets. Oracle launched 25 Claw-powered agentic applications across ERP, HCM, supply chain and customer experience. Leone explained why it exists. One home-sharing customer loads about a billion records a month into its general ledger. A chatbot can't reconcile that, and copying the data to an outside AI tool isn't viable.

"Claw needs to be close to the data. You can't move 2 million records out of the system and then put it back in after you're done." Chris Leone, EVP, Applications Development, Oracle

There is a second advantage. One Claw application, with a data set, a set of permitted actions and a set of skills, replaces many single-purpose agents.

"I don't have to build 100 specialized agents to answer questions and do work. I can build one." Leone
His examples were deliberately practical:

• A ledger application reads a close policy, learns its approval gates and tolerance thresholds, and checks every transaction against them.
• An EPM-to-execution application turns a restructuring plan into real transactions: which job requisitions to freeze, and which qualified employees could move.
• Scheduling and load planning produce rosters and shipment plans within overtime and back-to-back shift rules.

"There's no such thing as coming soon. It's available now." Leone's line from earlier in the day, quoted back to him by the session host. The new applications shipped a week before the event.

What changes for Fusion implementations?

Oracle published its implementation playbook as AI skills anyone can use. The Fusion Setup and Configuration Registry (FSCR) is a public Git repository. It describes Fusion setup in a form AI agents can read and execute. It tracks configuration dependencies, includes product managers' best practices, and runs smoke tests to validate the result. Leone said his team built the full skill set in about six months. Business rules, such as absence, time and incentive compensation rules, come next.

"Our philosophy was we wanted to democratize the mechanical work of setting up and configuring applications." Leone

The early evidence comes from customers. Lexington Health reports that Oracle's agentic delivery approach cut certain design and configuration work by 80%.

How does Oracle handle AI cost and accountability?

Oracle is designing both into the product. On cost, AI Agent Studio lets customers:

• set budgets for the whole estate, for each agent or for each agentic application;
• shut agents off when they hit a budget, or with a kill switch;
• simulate a workload to estimate its cost before going live;
• test whether a cheaper model can do the job. Oracle runs an open-weight model on OCI at no charge to customers and will recommend switching to it when accuracy holds.

On accountability, Leone was blunt: "It's an engineering problem, not a person problem." Leone, on who is accountable when an agent errs. Agents work inside a harness of delegated authority they cannot exceed. The person who defines that governance layer owns the outcome; if the harness fails, "that's on development."

What does this mean for systems integrators?

Their work moves up the value chain. Leone acknowledged that partners are unsure what this means for them. His advice was to take Oracle's registry as the foundation and build their own industry skills and best practices on top of it.
"This was going to happen no matter what. This is the natural evolution of every industry getting disintermediated and having to work on higher-value problems." Leone

OCI: from niche cloud to strategic AI capacity

OCI is now Oracle's growth engine, and its constraints are physical rather than technical. In the quarter ended August 31, 2026, OCI revenue more than doubled, up 121% to $7.4B. Remaining performance obligations, Oracle's contracted backlog, reached $664B, up $209B in a year. Oracle guided total cloud revenue growth of 64% to 70% for the following quarter.

 

OCI is growing far faster than a market that is itself accelerating. Synergy Research puts Q2 2026 cloud infrastructure spending at $143.4B, up 43%, the fastest growth in eight years. Amazon held 28% of that market, Microsoft 20% and Google 15%. Synergy groups Oracle among the fastest-growing "tier two" providers. OCI is still a fraction of the leaders' scale, but it is closing the gap faster than any of them.

"AI technology has lit a fire under the cloud market and is now driving unprecedented growth." John Dinsdale, Chief Analyst, Synergy Research Group

What limits OCI's growth?

Power, construction and public acceptance, which are the same constraints every hyperscaler faces in 2026. The infrastructure discussion at AI World made three points that apply across the industry:

• Capacity is planned against signed customer demand. At the scale of large AI campuses, build cycles run well over a year. The discipline that matters is keeping construction tied to contracts, not forecasts.

• The bottleneck keeps moving. Power, then skilled labor, then permits and local approval have each been the limiting factor at different stages of a build.

• Local communities decide. Approval is won locally: through utilities, county commissions and documented community-benefit commitments.

What should buyers watch?

Cash. Capital spending of $28.5B outran operating cash flow of $23.1B last quarter, leaving free cash flow at about –$5.4B. Oracle is funding its backlog ahead of the revenue. If the backlog converts on schedule, OCI becomes a long-term rival to the three largest clouds.

 

If large AI customers slip or renegotiate, Oracle carries the capacity risk. Either way, enterprise buyers benefit from a fourth cloud with the scale to compete on price and availability.

What can we conclude?

• SAP's lead comes partly from its own installed base. Much of SAP's 29% cloud ERP growth comes from moving existing on-premises customers to the cloud ahead of the end of core ECC maintenance in 2027. Fusion's 14% rests more on new customers and expansion, which is harder-won growth. Once SAP's migration wave peaks, the gap should narrow.

• Fusion and Workday are growing at about the same rate, with different strengths. Workday reports that AI made up more than 25% of new annual contract value, and that 5,500+ customers run at least one of its agents. Its strength is HR. Fusion's is breadth: ERP, EPM, supply chain, HCM and CX on one data model. That breadth is what Claw applications need to plan across functions.

• Oracle's AI usage is substantial. Sicilia said Fusion consumed 900 billion tokens in a single quarter. That level of production use is hard to fake.

• Oracle owns both the infrastructure and the applications on top of it. SAP and Workday lean largely on hyperscaler partners for capacity. Microsoft is the only other major vendor that owns both the apps and the infrastructure. Oracle can price, place and tune AI workloads for its own applications, and offer customers an open-weight model at no charge.

• OCI is growing nearly three times as fast as Azure, from a much smaller base. The comparison that matters is no longer growth rate but delivered capacity. Can Oracle bring capacity online as fast as it has booked it?

What is the growth trajectory?

Our estimate: Fusion holds mid-teens growth through 2027. It has room to accelerate if Claw applications drive module expansion and FSCR shortens sales cycles. OCI keeps growing much faster than the market as long as data center capacity comes online on schedule. On that path, infrastructure's share of Oracle's revenue keeps rising, and Fusion becomes the main showcase for what that infrastructure can do.

Prognosis and strategic implications

Oracle's design choices put it ahead of most rivals in enterprise AI. Its risks are financial and operational: funding the capacity, and getting customers' processes ready for agents. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear value or weak risk controls. Oracle's sessions addressed all three, at least in product design.

What this means for Oracle's competitors

1. Agents are moving from chat to computation. The next round of competition is about which agents can reconcile a billion-row ledger, re-plan a workforce or optimize a load plan inside the system of record. SAP Joule, Workday's agents and Microsoft Copilot will be judged on the same standard: does the heavy work happen next to governed data, or does the data get copied out?

2. Implementation knowledge stops being proprietary. By publishing FSCR, Oracle has turned setup expertise into free, version-controlled code. Rivals whose ecosystems profit from long, manual configuration projects will face buyers asking why theirs can't do the same. Systems integrators' value shifts to business rules, industry processes and governance design.

3. Controlling AI costs becomes a buying criterion. Per-agent budgets, kill switches, pre-production cost simulation and recommendations to switch to cheaper models address the most common reason AI projects get canceled. Expect buyers to ask every vendor for the same controls.

4. Accountability becomes something you can engineer and audit. Leone's position is that agents act only within delegated authority, and the person who sets the policy owns the outcome. That gives auditors something concrete to test, and gives rivals a standard to match.

5. Owning the full stack is becoming an advantage again. Oracle and Microsoft own infrastructure, database and applications. That control over the cost, location and performance of AI workloads is hard for application-only vendors to match.

Signals to watch
 

 

 

 

 

 

 

What's your edge?

 

For CEOs: 
1. Fund outcomes, not agent counts. Oracle builds agentic applications around a defined outcome with constraints, such as a restructuring plan carried through to individual transactions. Pick one outcome that crosses functions, like working capital or workforce cost, and fund it end to end.

2. Plan for AI compute to be scarce. Cloud infrastructure spending grew 43% in Q2 2026, and every provider is limited by power and construction. Put capacity dates in contracts, and treat OCI as a credible fourth source of capacity alongside AWS, Azure and Google.

3. Redesign roles before agents force the change. Leone's own support organization no longer measures success by calls answered. It measures how quickly each correct answer becomes a knowledge article the next customer can find. Rewrite job descriptions now so your people govern the processes agents run.

For CFOs:

1. Start Claw on reconciliation in research mode. Leone said finance teams can begin without "a heavy-handed set of standard operating procedures." Point the ledger application at your highest-volume accounts, then measure exceptions caught and hours saved before giving it authority to act.

2. Reprice your systems integrator contracts. If FSCR handles the mechanical configuration, stop paying hourly rates for it. Put that money toward the work only partners can do: business rules, industry processes and governance design. Lexington Health's 80% reduction in certain configuration work is a useful benchmark.

3. Budget AI like any other cost line. Set budgets for each agent, use the kill switch, run cost simulations before go-live, and require a test of whether a cheaper model can do the job before paying for a frontier one. Gartner's advice applies: "They must not mistake activity for impact."

For CIOs:

1. Don't copy the ledger out to AI. Design for computation that runs where governed data lives. Leone's rule of thumb: you can't move millions of records out for analysis and put them back safely. Require every AI vendor to explain where the computation actually runs.

2. Fork the FSCR repository. Treat configuration as version-controlled code. Clone Oracle's registry, add your own skills and conventions, and require partners to contribute to it instead of working from spreadsheets.

3. Set up an agent control plane before the second agent goes live. It should hold budgets, delegated-authority limits, kill switches and an inventory showing every agent, its data access and its owner. Workday reports 5,500+ customers already running agents, so agent sprawl is a real risk this year.