Oracle stacks racks, networking and inference partners into one AI infrastructure pitch

The rack at the centre

Oracle and AMD have put a rack-scale platform at the centre of the pitch. AMD Helios is described by Oracle as an open, co-designed AI infrastructure platform, and it integrates 72 AMD Instinct MI455X GPUs per rack. The same platform provides up to 31 TB of HBM4 memory per 72-GPU rack and up to 1.7 PB/s of aggregate peak memory bandwidth per rack, alongside 6th Gen AMD EPYC processors with up to 256 cores per socket. m:…1333

That rack is only the top layer. Oracle says Positron Atlas, now generally available on OCI, is a purpose-built inference platform for transformer-based workloads, and it is expanding the OCI AI compute portfolio with it. Each Positron Atlas instance includes 8 Positron AI accelerators, 144-core Intel Xeon 6 CPUs, 1.5 TB of system memory, 30 TB of local NVMe storage and 200 Gbps networking. Oracle says the platform has a benchmarked memory-bandwidth efficiency of 93% and that Positron accelerators offer a 3-5x total cost of ownership advantage for targeted workloads, and the company says it supports cost-efficient deployment of models like GPT-OSS, Qwen 3.5 and Gemma 4. Treat those efficiency and cost figures as vendor claims: they come from Oracle and Positron, not from an independent test in the announcements. m:…3940

The network bet

Oracle is also selling the fabric between the racks, not just the racks themselves. Its Acceleron SmartNICs use hierarchical token buckets to track bandwidth consumption for RDMA and non-RDMA traffic, and OCI uses One-Way-Delay Congestion Control in those SmartNICs to manage congestion within the fabric. When a SmartNIC detects elevated delay for a flow, Oracle says it probes three alternate paths and moves the flow if it finds a better one. Standard equal-cost load balancing, by contrast, leads to uneven link utilisation because it pins flows without knowing the bandwidth of flows, according to Oracle. The point of the design is reliability under load: the OCCL plugin allows customers to connect to multi-planar fabrics for higher workload reliability, and it supports port-level and NIC-level traffic failover, steering collective communication to another available port or NIC after a failure. m:…7948

Partners fill the gaps

No single vendor covers every inference shape, so Oracle has lined up partners to cover the rest. Cerebras and Oracle announced plans to make Cerebras fast inference available on OCI, with availability planned in calendar year 2027, and Cerebras says its Wafer-Scale Engine integrates compute and memory on a single wafer-scale processor to reduce data movement. SambaNova and Oracle are collaborating to deploy SambaRack SN50 inference chips in OCI data centres, and SambaNova says SN50 delivers up to 5x the maximum speed of competitive chips with 3x lower total cost of ownership, built for disaggregated inference where GPUs handle prefill and SN50 handles decode. Oracle plans to host the systems in its data centres and provide the capacity to customers, and says it will offer the capacity to AI model builders. m:…0141 m:…7575

Intel and Snowflake extend the same pattern into enterprise workloads. Intel and Oracle announced a multiyear collaboration to expand Intel Xeon-based computing on OCI, including planned OCI X14 offerings based on Intel Diamond Rapids processors, with Advanced Matrix Extensions supporting FP8, FP16 and BF16 for inference. Snowflake and Oracle are expanding their partnership to launch Snowflake on OCI as a native, first-party offering, with planned integrations for Oracle Fusion Applications and NetSuite, and customers able to purchase it directly through Oracle using Oracle Universal Credits. Snowflake says the offering on OCI could reduce cross-cloud data movement and egress costs. The pattern across all four partners is the same: Oracle provides the data centre and the billing relationship, and the partner provides a specialised engine. m:…1389 m:…0105

What to watch

For agent workloads, the most practical piece may be the routing and sandbox layer rather than the raw rack numbers. Oracle launched Smart Model Router and Agent Sandbox for OCI Enterprise AI, and says OCI Enterprise AI is expanding to 30 additional commercial regions. Smart Model Router routes inference requests to GPU-backed regions, customers define routing profiles to manage model destination regions, and Oracle says the router eliminates the need for local model hosting. Agent Sandbox is now generally available, provides a managed execution layer for agentic applications, lets agents create temporary Linux workspaces, test generated code and run shell commands, and lets organisations configure compute resources, duration and network access. That is a deliberate answer to the question every enterprise agent team asks first: where does the agent run, and who controls what it can touch. m:…4307

This suggests Oracle is competing on completeness rather than on any single accelerator. A buyer who wants AMD racks, a specialised inference chip, a governed data cloud and a sandboxed agent runtime can now, on paper, get all of them through one OCI relationship. The gaps are timing and proof. Cerebras on OCI is not planned until calendar year 2027, Positron says its next-generation silicon Asimov tapes out in late 2026, and the headline efficiency and cost figures in these announcements come from the vendors themselves. The sensible test for an executive is therefore narrow: run a real agent workload, measure latency and cost per completed task, and compare that against the incumbent cloud before treating the stack as a single purchase decision. m:…0141 m:…3940