Oracle AI World 2026 (October 25–28, Las Vegas) marked a shift from AI that assists people to AI that executes governed business work. Across Fusion Applications, APEX, Oracle AI Database, and OCI, almost every announcement answered the same question: how do you let AI agents act on enterprise data without losing control?
The announcements also arrived at very different stages of readiness. Some are generally available today, some are in limited release, and some are planned for 2027 or within the next 12 months. Each takeaway below notes where things stand, and a full availability table closes the post.
Status labels used throughout:
- Available now: generally available or shipping today
- Limited availability: preview, beta, or limited release
- Announced: coming, with no firm date or a "within 12 months" window
- 2027: Oracle has committed to a 2027 start
1. Fusion moves from AI assistance to AI execution
Oracle's biggest applications message: Fusion is becoming a "system of outcomes," where people set objectives and AI agents carry out the work within enterprise controls. Four pieces make up that story.
Fusion Claw (Available now). Fusion Claw is a governed runtime for agents inside Fusion. The model plans and reasons, while ordinary deterministic code does the calculations and posts the transactions. Claw runs in an isolated environment, so the model never updates Fusion business objects directly
- Controls: customers set objectives, permissions, risk thresholds, and approval rules in an "Enterprise Operating Envelope." Every completed task produces an "Outcome Receipt" listing the policies applied, decisions made, and transactions executed.
- Applications: 25 Claw-powered agentic applications span ERP, HCM, SCM, and CX, bringing Oracle's agentic application portfolio to 75. Examples include Ledger (close readiness and accounting exceptions), Workforce Staffing (staffing plans within labor rules), and Procurement Spend (finding off-contract spend).
- Models and cost: at launch Claw runs on Google Gemini and OpenAI models. The runtime is included with Oracle's agentic applications product, which is purchased separately from a standard Fusion subscription, and work is metered in AI units.
Fusion Companion and Fusion Agentic Home (Announced, Release 26D). Companion is an intelligence layer that lets employees work with Fusion through supported AI tools such as ChatGPT, or customer- and partner-built interfaces. It picks and coordinates the right transactions, workflows, approvals, and agents, and can act across other enterprise systems within existing permissions (Oracle). Agentic Home is the native version inside Fusion: users state objectives, delegate work, and track it to completion. Both are planned for Release 26D, and Oracle has not published pricing (CIO).
Fusion Setup and Configuration Registry (Available now). FSCR publishes setup and configuration knowledge as machine-readable AI skills in a public GitHub repository. Agents can use it to plan, execute, test, and validate configuration, which could change how Fusion implementations are staffed and scoped.
For implementation teams: pilot one Claw application in a sandbox with human review on every run, and study the Operating Envelope and Outcome Receipts before widening automation. Start exploring FSCR for repetitive setup and validation tasks.
For leaders: confirm whether your contract includes the agentic applications product, and get AI-unit pricing for the two or three use cases with the clearest payback. Note that Claw supports only Gemini and OpenAI models today, which matters if your company has standardized on another provider.
2. Oracle and OpenAI bring frontier AI into enterprise systems
The headline partnership, with Sam Altman on stage alongside Oracle CEO Clay Magouyrk, pairs OpenAI models, ChatGPT, and Codex with Oracle's applications, data, and cloud. Most of it is still ahead, so it is worth separating the parts.
| Offering | Status | What it does |
| Oracle AI Workspace | Announced, no date | ChatGPT plugins that let employees access Fusion, NetSuite, Oracle AI Database, and OCI, and complete approved workflows |
| Oracle App Composer | Limited availability | Employees build apps with Codex; Oracle deploys and runs them on OCI with identity, permissions, audit, and cost controls set by IT. Supports APEXlang. |
| Oracle AI SMB Bundle | Available, eligibility-restricted | $12,000 a year in Oracle cloud credits for up to three years, plus ChatGPT Business, for independent, privately held US businesses |
| OpenAI Jalapeño accelerators on OCI | Announced | Inference accelerators developed with Broadcom, deployed on OCI for OpenAI capacity |
| AI Services Practice | Announced | New Oracle practice, plus plans to train and certify one million professionals |
The SMB Bundle excludes existing Oracle customers that have received or consumed more than $250,000 in Universal Credits, so most established Oracle shops will not qualify.
For implementation teams: if you have a backlog of small internal apps, App Composer's limited-availability program is worth a low-risk pilot. Test the export path early, since apps targeting APEXlang and OCI are Oracle-specific.
For leaders: do not plan ChatGPT or Copilot renewals around AI Workspace yet. Oracle has not said which ChatGPT tiers the plugins require, who bills for them, or when they ship.
3. APEX lets AI generate applications you can actually govern
APEX 26.1 with the AI Application Generator is generally available on Autonomous AI Database Serverless in OCI. Its key idea is a different model for AI-assisted development: instead of generating thousands of lines of code that someone must secure and maintain, AI generates an application definition that APEX validates and runs in its managed runtime.
- APEXlang: an open, declarative, human-readable language that describes a whole APEX app, including pages, navigation, forms, reports, dashboards, and business logic. Apps export as structured text files that can be diffed, merged, and reviewed in standard version control.
- Three ways to build, one definition: the low-code App Builder, APEX's native AI agent, or external AI coding agents, all producing the same APEXlang definition.
- AI Agents and AI Tools: developers add agents that can only call tools the app explicitly exposes, and tool execution stays inside the application's boundary.
- AI Interactive Reports: users reshape reports in plain language, such as "pivot sales by region" or "sort by customer count."
- Long-lived apps: because apps run on the managed runtime, even older applications pick up the latest patches and updates.
APEX remains included at no additional cost with all supported Oracle Database editions. It is also one of the four layers of Oracle's Deep Trust AI Architecture (takeaway 4), and Oracle's new App Composer supports APEXlang.
For implementation teams: put APEXlang definitions under source control from day one, and add AI Interactive Reports to an existing app as a low-risk first step.
For leaders: APEX is a governed path for AI-built and citizen-built apps. Set the rules now: who can generate apps, which data they can reach, and who reviews definitions before production.
4. Trust is enforced in the platform, not left to each agent
If agents are going to take action, someone must guarantee what they can see and change. Oracle's answer is to enforce security, privacy, and business rules in the underlying stack rather than relying on every agent or app to behave.
Deep Trust AI Architecture. Introduced on the opening day, it spans four layers: software platforms, data, data access, and application development. It combines capabilities across Oracle AI Database, Autonomous AI Database, and the APEX AI Application Generator, including Deep Trust Data APIs that limit how agents interact with business information.
Deep Data Security (Available now). This is the most concrete piece, and it ships in Oracle AI Database 26ai.
- User and agent identity, plus runtime context, are passed to the database. Declarative SQL policies decide which rows, columns, or cells each request may see and which actions it may take.
- Access is denied unless a policy grants it, and the database applies the same rules even when an agent generates an unexpected query.
- It builds on Virtual Private Database and Real Application Security, modernized into a SQL-native model.
- Release Update 23.26.3 extends it to remote databases, object stores, and lakehouse platforms, and adds MFA for end users connecting directly.
Security for the agent era (Announced). Oracle AI Security Mesh, powered by OpenAI's cyber models, finds vulnerabilities across on-premises systems, OCI, and other clouds, provides evidence such as reproduction steps, and applies fixes only after customers approve. It is complemented by OCI Identity Assurance in OCI IAM, Zero Trust Packet Routing, and a planned F5 collaboration for app and API protection. Oracle has not given Security Mesh an availability date.
For implementation teams: plan Release Update 23.26.3 and run the Deep Data Security LiveLabs against a non-production copy of a real schema.
For leaders: make database-enforced authorization a standard for any agent touching production data, not a per-project choice.
5. The database becomes the agent's memory
Agents that work over days or weeks need to remember earlier decisions, preferences, and changed assumptions. Oracle's position, featured in Juan Loaiza's database spotlight, is that this memory belongs in the database, under the same governance as the rest of your data.
Oracle AI Agent Memory (Available now). It is a database-backed memory layer that stores messages, records, durable memories, summaries, and prompt-ready context in Oracle AI Database. It is initially available as a Python library and is integrated with Oracle AI Database Private Agent Factory.
- Short- and long-term memory: working context for the current task, plus durable facts and preferences extracted over time.
- Hybrid retrieval: semantic search combined with exact matching of identifiers and phrases, with embeddings generated in the database.
- Graph-aware memory: related memories are connected, including facts that refine or supersede earlier ones. If a planning assumption changes after a management review, the agent can find the current version and see how it relates to the old one.
- Enterprise operations: memory inherits backup, replication, high availability, encryption, fine-grained access control, and audit from Oracle AI Database.
- Framework friendly: it can serve as the storage layer for external memory frameworks, and Oracle's LangChain and LangGraph integrations use the same database.
This pairs naturally with Deep Data Security: one controls what an agent may access, the other what it remembers, and both live in the same governed engine.
For implementation teams: prototype one agent with AI Agent Memory on a 26ai instance or the free container, and define memory scope per user from the start.
For leaders: treat agent memory as enterprise data. Decide retention, isolation, and audit rules now, before memory stores multiply across teams and tools.
6. Business context becomes the new data layer
Access to data is no longer enough. An agent asked about "margin" needs to know which revenue and cost measures, business unit, period, and scenario apply. Oracle announced two complementary ways to give agents that shared business meaning, both still on the roadmap.
Autonomous AI Lakehouse: Business Semantic Layer and AI Data Catalog (Announced, within 12 months). The Semantic Layer builds reusable models of enterprise data from catalog annotations, SQL annotations imported from Oracle AI Database, and AI-suggested analytic views. Oracle also announced expanded Apache Iceberg support, centralized governance, and a Business Simulation Engine that brings Essbase into the lakehouse for what-if planning, write-back, and allocations. Oracle has given no date for the Simulation Engine or the Iceberg updates.
Oracle AI Data Platform: business ontologies and Semantic Reasoning Engine (Announced, within 12 months). An AI Discovery agent drafts ontologies by analyzing database schemas, semantic models, application and API documentation, and unstructured documents. Domain experts then review and approve them.
- The Semantic Reasoning Engine assembles business-ready context before it reaches the model: it retrieves across federated systems, resolves business relationships, and applies governance controls.
- Reusing context and earlier findings across queries is meant to cut repetitive reasoning and token usage.
- Deep Insight Agents use the ontologies to investigate complex business questions, and new governance features track agent ownership, dependencies, permissions, and lifecycle to limit agent sprawl.
For implementation teams: start annotating key tables and analytic views in 26ai now. That metadata feeds the Semantic Layer when it arrives.
For leaders: the hard part is organizational, not technical. Agreeing on one definition of each core metric across finance, sales, and operations is work you can start today, and it transfers to any platform.
7. OCI goes multi-silicon
OCI now spans more than 200 services, and its infrastructure revenue reached $7.4 billion in fiscal Q1 FY27 (June–August 2026), up 121 percent year over year, with 850MW of data center capacity added in the quarter. The infrastructure theme was "fungibility": training, inference, and the tools agents run need different hardware, so OCI is offering many architectures in one cloud.
| Compute option | Status | Notes |
| NVIDIA Vera Rubin NVL72 | Available now | Rack-scale systems, with Superclusters planned to expand to multigigawatt scale |
| NVIDIA Vera CPUs | Available now | For the CPU-heavy work agents do between model calls; initial deployment of 150,000 CPUs, rolling out by region |
| NVIDIA HGX Rubin NVL8 | Announced | Eight Rubin GPUs with two 64-core Intel Xeon 6 CPUs, 4TB DDR5, and 6.4Tbps RDMA networking |
| OCI Compute E7 (AMD EPYC, 6th gen) | Announced | AMD Pensando DPUs supporting Oracle Acceleron NICs |
| OCI X14 (Intel Diamond Rapids) | Announced | Part of a planned multiyear Intel collaboration |
| Google TPUs | 2027 | First hyperscaler outside Google to offer TPUs, ramping to gigawatt scale over multiple years |
| Cerebras wafer-scale inference | 2027 | Fast inference |
| SambaNova SN50 | Announced | No date given |
| OpenAI Jalapeño (with Broadcom) | Announced | Inference capacity for OpenAI |
| Quantinuum Helios | Limited availability | Quantum AI Compute beta program, alongside OCI GPU and HPC |
OCI's existing NVIDIA estate includes GB200 and GB300 NVL72 systems, and OpenAI trained and served GPT-6 Astra on an OCI Supercluster of more than 100,000 Blackwell GPUs. Although Vera is announced as generally available, Oracle's detailed Vera announcement describes a select-region rollout this quarter, so check shape visibility and quota in your tenancy.
For implementation teams: benchmark Vera CPU shapes for agent sandboxes, code execution, and tool calls, where single-thread speed matters most.
For leaders: treat TPUs, Cerebras, and other new silicon as 2027 capacity. Write them into multi-year commitments as dated expectations, not as capacity you can draw down this year.
8. The network and Kubernetes keep all that compute productive
An accelerator waiting on data is capacity you pay for but cannot use. Oracle's answer is to keep extending Oracle Acceleron, its networking suite introduced in 2025, and to make OKE a stronger home for AI workloads.
Oracle Acceleron (Announced). New capabilities target data movement, resilience, and secure multi-tenant AI clusters :
- High-Performance Fabric: RDMA connectivity on OCI front-end compute for inference, storage, and HPC workloads.
- Expanded NVMe block storage: more attachment options across Acceleron bare metal and VM shapes, with up to 3x better performance for AI and database workloads.
- Oracle Collective Communications Library: keeps distributed training running when a network path fails, by spreading collective traffic across OCI's backend multiplanar network.
- Front-end multiplanar networking: automatic path selection, failover, and congestion control.
OCI Kubernetes Engine (Available now, mostly). NVIDIA Dynamo is coming to OKE for disaggregated inference. OKE also shipped a steady run of AI-focused updates through 2026:
| Update | Released |
| GPU maintenance event repair for bare metal GPU nodes | Oct 21, 2026 |
| Kubernetes 1.37 (preview) | Oct 8, 2026 |
| GPU host repair for managed bare metal GPU nodes | Oct 2, 2026 |
| Node rollback to an earlier Kubernetes version | Aug 14, 2026 |
| AMD GPU Operator add-on | Aug 3, 2026 |
| NVIDIA GPU Operator and NVIDIA Network Operator add-ons | Jul 14, 2026 |
| Up to 20,000 nodes per enhanced cluster | Jul 8, 2026 |
| Zero Trust Packet Routing support | May 27, 2026 |
| Karpenter Provider for OCI | Apr 9, 2026 |
Two related notes. Oracle recommends the eBPF-based Cilium CNI for clusters above 5,000 nodes, but today it is self-installed rather than a managed option, And the Oracle AI Database Operator for Kubernetes (OraOperator) reached v2.2.0 in August 2026, adding Autonomous AI Database lifecycle automation, point-in-time restore, Data Guard Fast-Start Failover management, and GPU-backed private AI services.
For implementation teams: enable the GPU Operator add-ons and GPU host repair before scaling AI node pools. If you rely on Kubernetes NetworkPolicies with the VCN-native CNI, confirm they are actually enforced.
For leaders: Kubernetes is becoming the operating layer for both AI inference and databases on OCI. Decide which workloads belong there and which stay on Exadata or Autonomous services.
9. Data stays put, and stays up
Rather than asking customers to move data to Oracle, the announcements bring more services to where the data already sits, and raise the availability bar for the databases agents depend on.
Snowflake on OCI (Announced). Snowflake AI Data Cloud will run as a native, first-party OCI service. Oracle will handle pricing, discounts, and billing, and customers can apply Oracle Universal Credits (Oracle).
- Supported sources include Fusion Applications, NetSuite, Autonomous AI Database, MySQL HeatWave, and Oracle NoSQL Database, with native transfer meant to replace custom pipelines.
- Oracle says it will reduce cross-cloud data movement and egress costs. The release does not promise zero egress, and it gives no date, regions, or pricing.
Multicloud database (Mixed). Oracle AI Database@Google Cloud is now generally available in Mexico, with Exadata, Autonomous AI Database, and Autonomous AI Lakehouse. Oracle also plans to pursue FedRAMP Class D (formerly FedRAMP High) for its database services in the AWS, Azure, and Google Cloud government environments, and Microsoft outlined plans to bring Oracle AI Database@Azure to Azure Government Cloud with new sovereign controls.
Storage and delivery (Announced). The OCI NetApp Storage Service brings managed ONTAP to OCI. OCI Global CDN Connect links Object Storage directly to CDN providers, including Cloudflare and Fastly, with no OCI egress charges on eligible transfers.
Availability tiers (Available now). Oracle AI Database 26ai on Exadata now offers Platinum-tier availability, with disaster failover typically under 30 seconds for Active Data Guard and RAC, without application changes. Diamond tier targets failover under three seconds with zero data loss, using active-active clusters built on GoldenGate or Globally Distributed AI Database. Both were introduced earlier in 2026, but they underpin Oracle's case for running agents on live operational data.
For implementation teams: inventory the pipelines you run today from Oracle sources into Snowflake or other clouds. These are the candidates for native integrations once dates are published.
For leaders: Snowflake on OCI is mainly a procurement and data-gravity decision. If you hold Universal Credits and run Snowflake elsewhere, ask both account teams about terms, residency, and migration, and get dates in writing.
10. Oracle enters media with Oracle Media Fabric
The most unexpected launch was a new industry product. Oracle Media Fabric joins AI-powered content creation with next-generation video delivery, built around open standards, and is planned to be available within the next 12 months. Clay Magouyrk framed it as bringing content creation and streaming, historically separate processes, together on one AI platform.
| Component | What it does |
| Oracle Creative Studio | A model-agnostic production suite that picks Oracle or third-party models for each creative task based on security, provenance, cost, and IP protection |
| Oracle AI Media Models | Video and image creation and editing that help studios move faster while keeping creative control |
| Oracle Video Edge | Delivery infrastructure for high-quality streaming |
| Oracle Adaptive Media Platform | Connects creation and delivery for audience-specific advertising, personalized programming, and shared viewing |
Oracle positioned this alongside its growing work with Skydance and Paramount, and paired it with cheaper delivery: Netskrt's planned participation in OCI Global CDN Connect adds dedicated connections with no OCI egress charges for eligible transfers.
Who should care: media, entertainment, and marketing teams first. The broader signal is that Oracle is now building complete industry products on OCI, and that model choice governed by provenance and IP rules is becoming a standard requirement for any generative media tool.
Availability at a glance
Ten of the 26 items below can be used today. The rest are in limited release, undated, or scheduled for 2027.
| Announcement | Area | Status |
| Fusion Claw and 25 Claw-powered agentic applications | Fusion | Available now |
| Fusion Setup and Configuration Registry | Fusion | Available now |
| APEX 26.1 AI Application Generator and APEXlang | APEX | Available now |
| Deep Data Security (incl. RU 23.26.3) | Database | Available now |
| Oracle AI Agent Memory | Database | Available now |
| Platinum and Diamond availability tiers | Database | Available now |
| Oracle AI Database@Google Cloud, Mexico | Database | Available now |
| NVIDIA Vera Rubin NVL72 and Vera CPUs | OCI | Available now (regional rollout) |
| OKE AI updates (20,000 nodes, GPU operators, GPU repair) | OCI | Available now |
| Oracle AI SMB Bundle | OpenAI | Available, eligibility-restricted |
| Oracle App Composer | OpenAI | Limited availability |
| Quantinuum Helios (Quantum AI Compute beta) | OCI | Limited availability |
| Fusion Companion and Fusion Agentic Home | Fusion | Announced (Release 26D) |
| Oracle AI Workspace | OpenAI | Announced, no date |
| Oracle AI Security Mesh | Security | Announced, no date |
| Snowflake on OCI | Data | Announced, no date |
| Oracle Acceleron updates | OCI | Announced |
| NVIDIA Dynamo on OKE | OCI | Announced |
| HGX Rubin NVL8, AMD E7, Intel X14, SambaNova SN50 | OCI | Announced |
| OCI NetApp Storage Service, Global CDN Connect | OCI | Announced |
| FedRAMP Class D for multicloud database services | Database | Announced |
| Lakehouse Business Semantic Layer and AI Data Catalog | Database | Within 12 months |
| AI Data Platform ontologies and Semantic Reasoning Engine | Data | Within 12 months |
| Oracle Media Fabric | Media | Within 12 months |
| Google TPUs on OCI | OCI | 2027 |
| Cerebras inference on OCI | OCI | 2027 |
What to do next
The practical move after AI World is to separate what you can act on now from what you should track. A simple rule helps: budget for what has a date, and get the rest in writing.
Implementation teams
- Apply Release Update 23.26.3 in test and pilot Deep Data Security on one real schema
- Build one agent with AI Agent Memory and per-user memory scope
- Put APEX apps under source control with APEXlang, and try AI Interactive Reports
- Pilot one Fusion Claw application in a sandbox with human review on every run
- Enable OKE GPU Operator add-ons and GPU host repair; verify NetworkPolicy enforcement
- Benchmark Vera CPU shapes for agent and tool-execution workload
