Oracle and AMD have announced a “deep” expansion of their multigenerational collaboration to scale artificial intelligence (AI) capabilities in the cloud. Oracle Cloud Infrastructure (OCI) will serve as launch partner for the world’s first public AI supercluster based on AMD Instinct™ MI450 Series, with an initial deployment of 50,000 GPUs starting in Q3 2026 and expected growth throughout 2027 and beyond. This move comes at a time when large-scale AI capacity demand is accelerating, and next-generation models are surpassing the limits of current clusters.
The initiative is built upon AMD’s rack-scale platform “Helios”, introduced at the OCP Global Summit, and a next-gen vertical stack combining GPU Instinct MI450, AMD EPYC™ “Venice” CPUs, and advanced AMD Pensando™ “Vulcano” networking, all integrated with open factories—UALink/UALoE for scale-up among GPUs and Ethernet aligned with the Ultra Ethernet Consortium (UEC) for scale-out—, as well as liquid cooling and 72 GPU per rack density to maximize performance and energy efficiency.
“Our customers are building some of the world’s most ambitious AI applications, requiring robust, scalable, high-performance infrastructure,” said Mahesh Thiagarajan, EVP of Oracle Cloud Infrastructure. “By combining AMD’s latest innovations with OCI’s secure and flexible platform — and advanced networking with Oracle Acceleron — they can push boundaries confidently. After a decade of collaboration, we continue to deliver the best price-performance in an open, secure, and scalable cloud.”
“AMD and Oracle continue to lead AI innovation in the cloud,” said Forrest Norrod, EVP and GM of the Data Center Solutions Business Group at AMD. “With AMD Instinct, EPYC, and Pensando, Oracle’s customers gain new capabilities to train, fine-tune, and deploy next-generation AI in massive data centers.”
What’s been announced (and when)
- Public AI supercluster on OCI with AMD Instinct MI450: 50,000 GPUs starting in Q3 2026, expanding in 2027+.
- “Helios” architecture (rack-scale open) with 72 GPUs per rack, liquid-cooled, UALoE for scale-up, and UEC-aligned Ethernet for scale-out.
- AMD EPYC “Venice” CPUs as header nodes for orchestration and data, with end-to-end confidential computing.
- Converged DPU-accelerated networking with AMD Pensando “Vulcano” (up to 3 AI-NICs of 800 Gbps per GPU) offering lossless connectivity, programmability, and advanced RoCE/UEC support.
- HBM4 in MI450: up to 432 GB per GPU and 20 TB/s bandwidth, enabling training and inference of models 50% larger entirely in memory.
- Open AMD ROCm™ software and fined-grain partitioning/virtualization: SR-IOV, multi-tenancy, and secure GPU and pod sharing.
- General availability of Compute with AMD Instinct MI355X in OCI’s zettascale supercluster (up to 131,072 GPUs) offering open-source compatibility and value.
Why it matters: models surpass clusters—and the infrastructure responds
This announcement reflects a reality faced by AI teams at the cluster level: frontier models and their variants (multimodal, mixtures of experts, increasing context lengths) overwhelm current architectures in memory, bandwidth, and interconnection fabrics. The shift to HBM4 — with up to 432 GB per GPU and 20 TB/s — combined with open factories ( intra-rack UALink/UALoE and inter-rack UEC) points to three classic bottlenecks:
- Memory capacity and throughput: less forced partitioning, fewer checkpoints, and simpler pipelines.
- Efficient scaling: fewer hops and lower latency between GPUs, optimized collective operations, and observable fabrics (telemetry, congestion control).
- Sustainable operation: liquid cooling with quick disconnection, reasonable **density** per rack, serviceability, and energy efficiency.
The combined positioning of OCI (control layer, security, acceleron networking) and AMD (GPU, CPU, DPU, and open rack-scale) aims to capture this performance-scale-efficiency triangle with a public cloud that, in this wave, does more than rent compute: it delivers ready-to-work clusters with isolation controls and trustworthiness for sensitive workloads.
Under the hood: “Helios” + MI450 + Venice + Vulcano
AMD Instinct MI450 GPU (HBM4)
- Up to 432 GB HBM4 per GPU, 20 TB/s bandwidth.
- Models 50% larger than previous generation, entirely in memory (less partitioning, less synchronization overhead).
- Shapes designed for advanced LLMs, GenAI, and HPC in open environments (ROCm).
Open “Helios” rack (ORW) with scale-up/scale-out capabilities
- 72 GPUs per rack, liquid-cooled, with rapid dongles, double width (thermal and service flow).
- UALoE as the transport layer of UALink: hardware coherence and shared memory among GPUs inside the rack without passing through the CPU.
- UEC-aligned Ethernet for scale-out: high performance, multi-path, and programmability across pods and racks.
AMD EPYC “Venice” CPU (header node)
- Large-scale job orchestration, high-speed data ingestion and preprocessing.
- Confidential computing and built-in security for end-to-end flows involving sensitive data.
Converged DPU-accelerated networking with AMD Pensando “Vulcano”
- Up to 3 AI-NICs of 800 Gbps per GPU: lossless connectivity, programmable, and advanced RoCE/UEC.
- Line-rate ingestion, access control, and security policies managed by the DPU itself.
Software and multi-tenancy
- ROCm™: an open-stack including frameworks, libraries, compilers, and runtimes, designed for portability and vendor independence.
- Fine-grain GPU and pod partitioning, SR-IOV, robust multi-tenancy: secure sharing and adjusting GPU resources to the actual workload needs.
OCI Supercluster: from MI300X/MI355X to MI450
The announcement builds on a path that started in 2024 with Instinct MI300X in OCI and continued with the GA of OCI Compute with MI355X, already deployed in the zettascale supercluster (up to 131,072 GPUs). MI450 is the next step: featuring HBM4, bandwidth, and a fabric designed for trillion-scale models and long moat lengths, with less penalty for sharding.
For customers, this translates into a menu that combines value and scale: MI355X for price-performance ratio and open-source compatibility; MI450 for cutting-edge limits where memory and fabric make the difference.
What this stack solves (and for whom)
Frontier training (massive pre-training, MoE, long contexts):
- Less partitioning, more tokens/sec that pay off, more stable collectives, and fabric observability.
Fine-tuning and massive inference (RAG, agents, copilots):
- Lower latency in scale-out, enough memory for more complex orchestration without constant offloading.
HPC with mixed workflows (simulation + ML):
- “Venice” head node accelerates ingestion and orchestration; DPU offloads network and security tasks from CPU/GPU.
Regulated sectors (health, finance, public sector):
- Confidential computing and multi-tenant controls; built-in security and traceability for auditability.
Market review: open standards and smart verticalization
The strategic value is not just in the numbers (50,000 GPUs) but in the how. Oracle and AMD prioritize open standards (ORW/Helios, UALink/UALoE, UEC, ROCm) with a verticalization where each layer adds efficiency:
- Open rack (service, thermal, power) → sustainable density.
- Intra-rack factory (coherence, shared memory) → less latency and more throughput without passing through CPU.
- Inter-rack factory (Ethernet UEC) → observable scale-out and programmability.
- Header node with confidentiality → security and high utilization.
- DPU → ingestion and security posture aligned at line-rate.
- ROCm → real portability and less vendor lock-in.
Overall, the OCI supercluster aims to be more than just “GPU in the cloud”: it seeks to be the reference architecture for large-scale AI, open, and operational.
Availability and considerations
Oracle and AMD plan to start deployment in Q3 2026 (expanding in 2027+). As with any infrastructure roadmap, there are disclaimers: timelines, specifications, and prices may vary; some details (e.g., partitioning percentages, SKUs, software layers) will emerge as pilots and validations progress. The MI355X in OCI’s zettascale supercluster is already generally available for workloads requiring massive scale today.
Conclusion: a “giant leap” towards scaled AI with open standards
The expansion of the Oracle-AMD partnership culminates in a public supercluster that, on paper, pushes AI limits from the cloud: starting with 50,000 MI450, HBM4 in every GPU, UALink/UEC fabrics, EPYC “Venice”, and Pensando “Vulcano” under the rack-scale Helios umbrella. The approach is dual: power and openness. For teams facing the gap between model ambition and cluster capacity, it’s a clear signal: infrastructure is evolving to keep pace.
The real question, as always, will be in execution: actual latencies, collectives in “challenging” scenarios, total costs, and SLOs. But the direction is clear: more memory, more fabric, more efficiency… and less friction to take AI from prototypes to large-scale production.