AMD Helios rack-scale solution. These are liquid-cooled, 72-GPU racks powering AMD Instinct MI450 Series GPUs, AMD EPYC CPUs (Venice) and do feature integrated UALoE scale-up connectivity.
Oracle buying 50K MI450 for 2026. Is that about $2 billion.
The AMD Helios is a next-generation, open-standard, rack-scale AI infrastructure platform designed for large-scale AI training, inference, and fine-tuning.
It integrates AMD Instinct GPUs, EPYC CPUs, and advanced networking in a double-wide, liquid-cooled (or air-cooled) configuration optimized for high density, efficiency, and scalability.
It supports open compute standards like OCP DC-MHS, UALink for GPU-to-GPU interconnects, and Ultra Ethernet Consortium (UEC) for scale-out networking. Helios is positioned as a flexible alternative to proprietary systems, enabling “agentic AI” workloads with heterogeneous compute.
72 AMD Instinct MI400 Series GPUs (next-gen, including MI450 variants) per rack.
Up to 432 GB HBM4 per GPU (total ~31 TB per rack).
Supports in-memory processing for models 50% larger than prior generations.
Up to 40 petaFLOPS FP4 per GPU (rack-scale: ~2.88 exaFLOPS FP4). 1.4 exaFLOPS FP8 per rack.
20 TB/s memory bandwidth per GPU (rack-scale: ~1.44 PB/s). 300 GB/s scale-out per GPU; up to 1.6 TB/s sustained CPU-to-memory bandwidth.
UALink (scale-up, low-latency GPU communication). AMD Pensando “Vulcano” 800G AI-NICs (up to 3 per GPU, UEC 1.0 compliant with RoCE); PCIe and Ethernet tunneling.
6th Gen AMD EPYC “Venice” (Zen 6 architecture, up to 256 cores).
Supports RACm software stack; programmable networking for congestion control.
AMD Instinct MI450
The AMD Instinct MI450 is a high-performance AI accelerator in the Instinct series, built on the CDNA 5 architecture. It’s chiplet-based, targeting data center AI/HPC workloads like large language model training and inference. As a successor to the MI350/MI355X, it emphasizes massive memory scaling for trillion-parameter models while maintaining competitive compute.
It uses TSMC’s 2nm node on the core die for density gains. TSMC 2nm (N2) for the accelerator core die (chiplet design. not all dies on 2nm).
Oracle Deal for 50,000 MI450 GPUs
Oracle and AMD announced an expanded partnership to deploy 50,000 AMD Instinct MI450 GPUs in Oracle Cloud Infrastructure (OCI), marking the first publicly available hyperscale AI supercluster powered by this scale of AMD hardware. This is part of OCI’s zettascale Supercluster (scalable to 131,072 GPUs) and builds on prior MI300X/MI355X integrations. Deployment starts in Q3 2026, with expansions in 2027+.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.