Oracle will deploy 50,000 AMD "Altair" MI450 GPU sockets across 700 Helios racks beginning in the third quarter of 2026. The all-in deal is estimated at $3.5 billion to $4 billion, including storage and networking. Oracle and AMD disclosed the commitment at Oracle's AI World conference this week, though neither company filed a formal press release and the dollar figure comes from industry estimates.

The cluster will draw roughly 200 megawatts, according to the two companies. Oracle co-founder and chief technology officer Larry Ellison announced the agreement during his keynote. Details surfaced through Oracle's online materials, not a press statement.

What capacity is Oracle actually committing to?

The headline figure of 50,000 sockets matches 50,400 sockets distributed across 700 racks at 72 GPUs each. Oracle says deployments start in Q3 2026 and expand through 2027. The cluster serves Oracle's general OCI infrastructure, not OpenAI contracts. Customers will rent time on the MI450 fleet the same way they rent time on Oracle's existing 30,000-MI355X cluster.

That 30,000-MI355X deal was struck in March. The MI355X chips became generally available this week. An eight-way MI355X node holds 2.3 TB of HBM3E memory and delivers 74 petaflops at FP6 and FP4 precision.

Which Altair variant will Oracle use?

Oracle confirmed it will deploy the MI450, not the MI450X. Two MI450 variants exist:

• Standalone MI450: targets traditional eight-way nodes, 432 GB HBM4 per GPU
• MI450X: powers double-wide AI racks co-developed with Meta Platforms and OpenAI

Oracle picks the variant with the most HBM capacity. Standard MI450X parts are expected to ship with at least 288 GB of HBM4. Oracle's stated preference suggests a custom configuration, though the company has not disclosed exact memory counts.

Each MI450 chiplet is etched on TSMC's 2 nanometer process. Peak compute reaches roughly 40 petaflops at FP4 precision. Memory bandwidth measures 19.6 TB/sec per GPU. An eight-way system board aggregates to 3.2 exaflops at FP4, 3.4 TB of HBM4, and 156.8 TB/sec of bandwidth.

What does a Helios rack deliver?

AMD's published rack numbers:

• 1.45 exaflops at FP8 precision
• 2.9 exaflops at FP4 precision
• 31 TB aggregate HBM4 memory
• 1.4 PB/sec aggregate bandwidth

The Helios rack holds 72 GPUs. The MI450X form factor scales to 64 or 128 GPUs per rack, while Oracle's 72-GPU configuration implies a customized footprint that the two partners have not publicly detailed. Networking relies on UALink over Ethernet.

The scale-up network delivers 260 TB/sec of aggregate bandwidth. Scale-out across 700 racks reaches 43 TB/sec. Each GPU can host up to three "Vulcano" Pensando DPUs at 800 Gb/sec apiece.

How does this position Oracle against Nvidia?

The Helios line targets Nvidia's "Oberon" rackscale systems. Oberon pairs Grace CG100 Arm CPUs with current Blackwell B200 and B300 GPUs, with future Vera CPUs and Rubin GPUs in the pipeline. AMD's choice of TSMC 2nm for MI450 matches the leading-edge process tier Nvidia relies on.

Oracle continues to buy Nvidia hardware. The same OCI datacenters will host both AMD and Nvidia GPU clusters. Oracle has not named the Ethernet ASIC supplier for this build. Broadcom's Nvidia alignment and Cisco Systems' or Marvell's presence leave the field narrow, with AMD's Pensando DPUs also a candidate.

What comes next?

Oracle has not confirmed per-socket pricing, full rack composition, or networking supplier. Watch for bill-of-materials disclosure and HBM4 allocation signals before Q3 2026 deliveries begin. The market will soon learn whether Oracle's MI450 capacity signals a genuine competitive shift, or simply marks another fleet addition to an OCI roster still dominated by Nvidia silicon.