Oracle will deploy 50,000 AMD "Altair" MI450 GPU sockets across 700 "Helios" rackscale systems at Oracle Cloud Infrastructure, with the first installation arriving in Q3 2026. The all-in cost ranges from $3.5 billion to $4 billion, including storage and networking. Oracle co-founder and CTO Larry Ellison announced the commitment at AI World 2025.

What are the Altair MI450 specifications?

AMD manufactures the MI450 compute die on Taiwan Semiconductor Manufacturing Co's 2 nanometer process. A single MI450 GPU delivers 40 petaflops of peak compute at FP4 precision, paired with 432 GB of HBM4 memory and 19.6 TB/sec of memory bandwidth. An eight-way MI450 system board reaches 3.2 exaflops at FP4, 3.4 TB of aggregate HBM4 capacity, and 156.8 TB/sec of aggregate bandwidth.

How does Helios rackscale stack up?

Each Helios double-wide rack holds 72 MI450 GPUs alongside an unspecified number of future "Venice" Epyc CPUs and "Vulcano" Pensando DPUs. AMD quotes 1.45 exaflops at FP8 and 2.9 exaflops at FP4 per rack, plus 31 TB of HBM4 memory and 1.4 PB/sec of aggregate bandwidth.

The UALoE scale-up network supplies 260 TB/sec of aggregate bandwidth. Scale-out across all 700 racks produces 43 TB/sec. Each GPU connects to up to three Vulcano DPUs, each rated at 800 Gb/sec.

Why did Oracle pick the MI450 over the MI450X?

The MI450X variant targets the Helios platform with at least 288 GB of HBM4 per GPU. Oracle stated it would deploy in Helios racks but selected the standalone MI450, citing demand for the highest HBM capacity per socket. The MI450 series scales to 64 or 128 GPUs per rack in the IF128 configuration, delivering 50 petaflops per GPU.

Oracle could receive a custom MI450X with 432 GB per socket; AMD has not disclosed the configuration. The cluster may also use four GPU sockets for every CPU socket, a ratio AMD has not confirmed.

How much power does this draw?

Oracle and AMD put the cluster's power consumption at approximately 200 megawatts. The 700-rack footprint matches the 50,000-GPU socket figure almost one-for-one at 72 sockets per rack.

What sits on top of the network?

Oracle developed the Acceleron architecture for OCI AI clusters, using DPUs as integrated switches to eliminate one network tier. The approach will run on the MI450 cluster and the Nvidia systems alongside it in OCI datacenters. AMD interconnects the cluster via UALink over Ethernet, effectively running Infinity Fabric over Ethernet.

Possible Ethernet ASIC suppliers include Cisco Systems, Marvell, or Pensando DPUs operating as switches within the AMD stack. Nvidia silicon is not in the running.

What about Oracle's earlier MI355X deployment?

In March 2025, Oracle contracted 30,000 "Antares+" MI355X GPUs for OCI. Those chips entered general availability this week at AI World 2025. The MI355X node carries 2.3 TB of shared HBM3E memory, 64 TB/sec of aggregate bandwidth, and 896 GB/sec of Infinity Fabric interconnect.

The 8-way node produces 74 petaflops at FP6 and FP4, or 7X the MI300X at FP16 without sparsity. The MI355X holds 4.2 trillion parameters per instance, against 1.5 TB of HBM3 on the December 2023 MI300X. That parameter capacity runs 5.9X higher than its predecessor.

Who can access the new cluster?

The MI450 deployment forms part of OCI's general infrastructure, separate from Oracle's OpenAI contracts. Customers will rent MI450 capacity on the same rental model used today for the MI355X cluster.