Oracle to offer cloud services using AMD's AI chips

Oracle and AMD have announced plans to expand their partnership by launching a new AI supercluster on Oracle Cloud Infrastructure (OCI) using AMD’s Instinct MI450 Series GPUs.

The deployment will start with 50,000 GPUs in Q3 2026, with further expansion planned for 2027.

The supercluster will use AMD’s “Helios” rack design, which includes next-generation EPYC CPUs and Pensando networking technology.

Oracle, a major cloud services provider, and AMD, a US-based chipmaker, have worked together for years to deliver GPU platforms for AI computing on OCI.

Oracle also announced that its cloud service now offers AMD Instinct MI355X GPUs, available in its high-capacity OCI supercluster.

The companies said these moves are intended to meet rising demand for large-scale AI workloads and provide customers with more options for training and deploying advanced AI models.

🔗 Source: AMD

🧠 Food for thought

Implications, context, and why it matters.

AMD MI450 faces ROCm maturity and training cost tests

Oracle Cloud Infrastructure (OCI) plans a 50,000-GPU MI450 rollout in Q3 2026, but no independent checks exist for performance or cost-per-training vs Nvidia’s Blackwell (its current flagship GPU architecture) and Rubin (the next-generation successor) 1.
AMD lists 432 GB of High Bandwidth Memory (HBM) 4 per GPU and liquid-cooled 72-GPU racks. Bigger memory allows 50% larger models, yet Radeon Open Compute (ROCm) still trails CUDA in enterprise use, with uneven Python support and lagging collective communication libraries (software that coordinates data exchange across large GPU clusters) 1 2.
SemiAnalysis (a semiconductor industry research firm) says AMD runs limited, bursty internal GPU clusters. Nvidia keeps persistent multi-year setups. That hurts trust in AMD’s ability to validate and tune software at OCI scale 3.

CUDA-to-ROCm migration demand rises with OCI MI450

The OCI deployment opens work for systems integrators (IT services firms that build complex infrastructure) plus tool vendors selling CUDA-to-ROCm ports, tuning, and compatibility layers. Pay gaps between AMD and Nvidia for AI software engineers let vendors hire talent and build ROCm skills 1 3.
AMD’s hipify-perl and hipify-clang convert 90 to 99 percent of smaller CUDA codebases to Heterogeneous-Compute Interface for Portability (HIP) for ROCm. Complex production AI apps still need manual tuning, which drives specialist consulting 4.
Teams adopting MI450 will need tools for Python kernel domain-specific languages (DSLs), faster ROCm Communication Collectives Library (RCCL), plus Non-Volatile Memory express (NVMe) key-value (KV) cache tiering. That invites new products from infrastructure software vendors 3.