Posted in Cloud & Infrastructure

Oracle Cloud Infrastructure’s GPU Supercluster Strategy: How OCI Is Positioning Against AWS and Azure for AI Workloads in 2026

Oracle entered the hyperscale AI infrastructure race later than AWS, Azure, and Google Cloud, and its strategy for closing that gap has been consistent: lean hard into bare-metal GPU access, undercut the competition on price, and build networking and data-center campuses sized for the largest AI training runs anyone is attempting. By late 2025, that strategy had a name — OCI Zettascale10 — and a scale that Oracle says reaches hundreds of thousands of GPUs across gigawatt-class data-center campuses. Here’s what’s verifiable about how OCI’s GPU infrastructure is actually built, and how its public positioning against AWS and Azure holds up.

At a Glance

DimensionOCI detail
Max GPUs per SuperclusterUp to 131,072 (B200/GB200); 65,536 (H200); 16,384 (H100); 16,384 (AMD MI300X)
Zettascale10 target scaleUp to 800,000 Nvidia GPUs across multiple gigawatt-scale campuses
Zettascale10 peak performanceUp to 16 zettaFLOPS (Oracle says 10x its first Zettascale cluster, baseline not restated)
Cluster networkingRDMA, up to 3,200 Gb/sec bandwidth, microsecond-level latency (bare metal)
GPU access modelBare metal (no hypervisor overhead) for NVIDIA and AMD instances — a claimed OCI differentiator
Example on-demand priceAMD MI300X at $6/GPU-hour, same price in every region, no long-term commitment required
Flagship deploymentStargate campus, Abilene, Texas, built with OpenAI
Zettascale10 availabilityOrders open now; general availability targeted for H2 2026

The Core Bet: Bare Metal Access as a Differentiator

OCI’s own GPU and AI infrastructure page frames its core differentiator plainly: Oracle says it is the only major cloud provider offering bare-metal instances with both Nvidia and AMD GPUs, free of the virtualization overhead that comes with running on a hypervisor. That matters for CUDA-heavy training workloads where every bit of memory bandwidth and interconnect latency affects wall-clock training time, and it’s also the mechanism behind Oracle’s price claims — Oracle’s own page states that other cloud providers’ comparable GPU instances can run up to 220% more expensive, based on on-demand pricing it dated to June 2024. That comparison is now well over a year old by late 2026 and both AWS and Azure have since adjusted GPU pricing and instance options, so treat the specific “220%” figure as dated context rather than a current, up-to-the-minute comparison.

Oracle’s GPU lineup spans current and prior Nvidia generations — Blackwell, H200, H100, L40S, A100, and A10 — plus GB200 Grace Blackwell Superchips used in its largest superclusters, and AMD’s Instinct MI300X (192 GB of memory per GPU) with MI355X also listed as becoming available. On the scale side, Oracle’s published Supercluster maximums are tiered by GPU generation: up to 131,072 GPUs for B200/GB200 configurations, 65,536 for H200, and 16,384 each for H100 and AMD MI300X, all connected over RDMA networking rated at up to 3,200 Gb/sec with microsecond-level latency on bare-metal instances.

Zettascale10: Scaling Past 100,000 GPUs

Oracle’s October 2025 announcement of OCI Zettascale10 pushed the strategy further: a cluster architecture connecting “hundreds of thousands” of Nvidia GPUs across multiple data centers, with initial deployments targeting up to 800,000 GPUs and up to 16 zettaFLOPS of peak performance — a figure Oracle described as 10 times the output of its first Zettascale cluster from September 2024, though the release didn’t restate that earlier cluster’s absolute numbers for a direct comparison. The networking layer behind this is Oracle’s own Acceleron RoCE (RoCEv2) architecture, which uses switching built into GPU network interface cards to route traffic across multiple isolated network planes — so a failure in one plane doesn’t take down the cluster — combined with a flatter two-tier network design (down from three tiers) intended to make GPU-to-GPU latency more uniform across a large cluster, and Linear Pluggable and Linear Receiver Optics to cut power and cooling costs at 400G/800G networking speeds.

Physically, Oracle is building these clusters inside gigawatt-scale data-center campuses, with multiple campuses sited within a two-kilometer radius of each other to keep latency low across what is effectively one giant logical cluster. The flagship example is the Stargate campus in Abilene, Texas, built jointly with OpenAI; OpenAI’s own VP of Infrastructure, Peter Hoeschele, is quoted in Oracle’s release saying the custom RoCE design “maximizes fabric-wide performance at gigawatt scale.” Oracle is taking orders for Zettascale10 capacity now, with general availability targeted for the second half of 2026.

How This Actually Stacks Up Against AWS and Azure

It’s worth being direct about a gap in Oracle’s own materials: Oracle’s Zettascale10 press release does not mention AWS or Azure by name at all — the competitive comparison exists mainly on OCI’s separate GPU product page, in the form of the pricing claims above, not in Oracle’s biggest infrastructure announcements. Oracle’s actual competitive argument, pieced together from its own public materials, rests on three legs: lower claimed GPU-instance pricing (OCI’s own page cites MI300X at a flat $6/GPU-hour with no regional variation or commitment requirement), bare-metal access that avoids virtualization overhead AWS and Azure’s standard GPU instances typically carry, and raw scale-out ambition through Zettascale10’s multi-campus design. What Oracle’s materials don’t provide is an apples-to-apples benchmark against a specific AWS or Azure supercluster offering, so the “better than AWS and Azure” framing should be read as Oracle’s own pricing and architecture claims rather than an independently verified head-to-head result.

Bottom Line

Oracle’s GPU strategy is coherent and aggressive: bare-metal access, a specific and public pricing undercut argument, custom RoCE networking engineered for massive, failure-tolerant GPU fabrics, and gigawatt-scale campuses built in partnership with anchor customers like OpenAI. Whether that adds up to a genuine lead over AWS or Azure for a given AI workload depends heavily on specifics — region, GPU generation, commitment terms — that Oracle’s own public materials don’t fully spell out in a current, side-by-side comparison. For organizations evaluating OCI specifically for large-scale GPU training, the Zettascale10 roadmap and bare-metal architecture are real and substantive; the specific cost-savings percentages Oracle advertises are worth re-verifying against current, same-day pricing from all three providers rather than taking at face value.

FAQ

What is OCI Zettascale10?
It’s Oracle’s next-generation AI supercluster architecture, announced in October 2025, designed to connect hundreds of thousands of Nvidia GPUs — up to 800,000 in initial deployments — across multiple gigawatt-scale data-center campuses, reaching up to 16 zettaFLOPS of peak performance per Oracle’s own figures.

What does “bare metal” GPU access actually mean, and why does Oracle emphasize it?
It means the GPU instance runs without a virtualization/hypervisor layer in between, which Oracle argues removes performance overhead; Oracle positions this as a key differentiator versus AWS and Azure, whose standard GPU offerings more commonly run virtualized.

Is Oracle’s GPU pricing actually cheaper than AWS and Azure?
Oracle’s own page claims competitors can be up to 220% more expensive, but that figure is based on pricing data dated to June 2024, so it should be re-verified against current pricing rather than treated as an up-to-date comparison.

When will Zettascale10 be generally available?
Oracle is taking orders now (as of its October 2025 announcement), with general availability targeted for the second half of 2026.

Does OCI only support Nvidia GPUs?
No. OCI also offers AMD Instinct MI300X (and lists MI355X as becoming available) alongside its Nvidia lineup, which spans Blackwell, H200, H100, L40S, A100, and A10.

Sources

Unknown's avatar

Author:

Innovations, logic & design thinking