While Oracle is still well known for its namesake database platform, infrastructure is the foundation for expanding its AI aspirations.
At its AI World conference last week in Las Vegas, Oracle outlined a broad set of updates to its Oracle Cloud Infrastructure (OCI) platform. The updates reach across the stack.
In terms of core infrastructure, Oracle announced that it added 850MW of data center capacity in the first quarter of its 2027 fiscal year. In the compute realm, Nvidia Vera Rubin NVL72 systems and Vera CPUs are generally available, with Google TPUs and Cerebras inference planned for 2027.
On the networking front, new features in its Acceleron suite speed data transfers between servers and storage and keep AI workloads running when a network path fails.
For data management, Oracle plans to offer Snowflake as a first-party OCI service and to pursue FedRAMP Class D certification for its multicloud database services. And for AI, Oracle introduced tools built with OpenAI for business workflows, application development, security testing and small businesses.
OCI now spans more than 200 cloud services, and its footprint continues to grow.
OCI now has more than 225 live and planned regions across public cloud, multicloud and dedicated deployments, including dedicated regions that do not require a public cloud, said Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, during an AI World keynote presentation.
Compute options on OCI are expanding beyond Nvidia; upcoming OCI Compute E7 instances will use sixth-generation AMD EPYC processors, and Oracle also plans Intel-based OCI X14 offerings.
Oracle's cloud expansion also raises the question of how OCI compares against rivals.
"Oracle is one of the larger cloud providers, but it is a tier down from the big three, Amazon, Microsoft and Google," said Jeremy Roberts, senior director, research and content at Infotech. "This continued expansion is an important way to make up the gap."
Networking updates aim to keep AI clusters running
While compute and services are important for infrastructure, so too is how those technologies connect. That's where networking comes into play. One congested or failed network path can disrupt thousands of AI accelerators that need to work together.
Oracle's networking updates in its Acceleron networking and infrastructure suite last week centered on improving resilience to limit the risk of failures spreading across large AI clusters. The Oracle Collective Communications Library steers traffic across the backend network toward healthier paths and helps workloads recover when links, ports or NICs fail.
Acceleron High-Performance Fabric added RDMA (Remote Direct Memory Access) connectivity on front-end compute to inference, storage, and HPC workloads., Front-end RDMA offers up to 10 times lower latency between compute nodes and storage nodes than the previous approach, according to Thiagarajan.
This latency improvement extends beyond AI clusters, he said, and applies throughout Oracle's cloud.
Security sits in the same layer: Oracle Acceleron provides hardware-based isolation for workloads that run untrusted, AI-generated code. With its latest updates, Oracle added Zero Trust Packet Routing deny rules and a policy workbench for writing network policies, as well as short-lived sessions for agents.
Storage access also gets a boost. Oracle said new options for attaching NVMe block storage to bare metal and VM instances improve storage performance by up to 3x.
Analysts are generally positive on Oracle's networking direction. Up until now, the importance of the network in AI performance and reliability has taken a back seat to accelerators such as GPUs and TPUs, said Rob Tiffany, an analyst at IDC.
"Anticipating the next AI bottleneck, Oracle is boosting network performance with enhanced Acceleron capabilities such as front-end RDMA," Tiffany said. "AI is only as fast as the slowest component and Oracle is ensuring that component isn’t the network."
It's important to note that reliability isn't about any one component, said Dion Hinchcliffe, chief research officer at Frontier Research.
Hinchcliffe emphasized that AI reliability isn't determined solely by the network. It's a full-stack problem, from GPUs and networking to orchestration, models and apps. In his view, Oracle is headed in the right direction. He noted that Acceleron addresses network failures, while new cluster health checks and GPU repair address compute failures.
That said, Hinchcliffe emphasized that all the pieces need to work together as one reliability strategy.
"The real test is whether entire AI services keep running when things fail, not just the network," Hinchcliffe said. "CIOs seek business continuity, not simply better plumbing."
Moving AI services closer to enterprise data
Snowflake AI Data Cloud will be available as a first-party OCI service, according to another AI World update, keeping data movement inside OCI to reduce network egress costs.
Agents change who accesses data within enterprises. The next million database users will be agents, not people, and agents often get more access than intended because today's products lack the controls to limit it, Thiagarajan said.
On the application side, Oracle introduced new tools built with OpenAI. Users can build applications with Codex using App Composer that can then be deployed on OCI. Oracle AI Workspace uses ChatGPT to work with Oracle applications and data, while a new AI services practice aims to help customers put the tools to work.
Those additions will be compared with what other vendors already offer, Roberts said.
"Oracle announcing that in the future customers can use ChatGPT with enterprise applications feels a lot like Microsoft's initial Copilot pitch, so it remains to be seen if they can make a meaningful distinction in the space," Roberts said.
Roberts sees the same challenge for App Composer and the services practice, as they aren't new categories. That said, when combined with Oracle's existing customer base and penetration in key areas like database and ERP technology, it could move the needle for some customers, he said.
Whether the new additions match what enterprises need from infrastructure providers remains an open question.
"Yes, Oracle is generally responding to what enterprises need, and CIOs don't want another sprawling cloud catalog," Hinchcliffe said. "They want reliable, secure AI infrastructure that works with what they already have, without more complexity or lock-in."
Sean Michael Kerner is an IT consultant, technology enthusiast and tinkerer. He has pulled Token Ring, configured NetWare and been known to compile his own Linux kernel. He consults with industry and media organizations on technology issues
