AMD · 2026-07-23 · major
AMD Helios — 72-GPU rack-scale AI system aimed straight at Nvidia's NVL72
AMD launched Helios, a rack-scale AI system with 72 Instinct MI455X GPUs, 18 EPYC Venice CPUs, 31 TB of shared HBM4, and 2.9 exaflops of FP4 inference. Microsoft, OpenAI, Meta, Oracle, and Anthropic are named launch customers.
AMD's first rack-scale answer to Nvidia's NVL72: 72 MI455X GPUs, 31 TB of HBM4, and hyperscaler commitments already signed.
Key specs
| Gpus per rack | 72 |
|---|---|
| Cpus per rack | 18 |
| Hbm4 memory tb | 31 |
| Fp4 inference exaflops | 2.9 |
Quick facts
| Maker | AMD |
|---|---|
| Product | Helios rack-scale AI system |
| GPUs per rack | 72× Instinct MI455X |
| CPUs per rack | 18× 6th-gen EPYC Venice |
| Shared memory | 31 TB HBM4 |
| Peak inference | 2.9 exaflops FP4 |
| Networking | AMD Pensando |
| Named customers | Microsoft, OpenAI, Meta, Oracle, Anthropic |
What is it?
AMD Helios is a full rack of AI compute AMD sells as one system. Each rack combines 72 Instinct MI455X GPUs, 18 sixth-generation EPYC Venice CPUs, 31 TB of shared HBM4 memory, and AMD Pensando networking, delivering 2.9 exaflops of FP4 inference for large model training and serving.
How does it work?
Helios treats a rack as a single accelerator. The 72 MI455X GPUs share their HBM4 through a high-bandwidth fabric so one training or inference job can see all 31 TB as one pool, while Venice CPUs manage scheduling and agentic tool loops. AMD also ships the ROCm stack, Pensando data-plane, and pre-integrated firmware so customers plug the rack into an existing data center instead of composing it themselves.
Why does it matter?
Nvidia's GB200 NVL72 has been the only rack customers could buy for frontier training runs, which is why the shortage keeps pushing GPU prices up. AMD Helios is the first credible alternative with signed hyperscaler commitments, and Microsoft, OpenAI, Meta, Oracle, and Anthropic all took slots at launch. If deliveries hold, buyers finally have a second source for rack-scale AI compute.
Who is it for?
Hyperscalers, AI labs, and sovereign cloud operators planning the next generation of frontier training clusters
Frequently asked questions
- How does AMD Helios compare to Nvidia's NVL72?
- AMD Helios packs 72 Instinct MI455X GPUs, 31 TB of shared HBM4, and 2.9 exaflops of FP4 inference into one rack, directly targeting Nvidia's NVL72 (GB200) tier. AMD leads on HBM capacity per rack; Nvidia still leads on shipping volume and software maturity. Real workload numbers will come when customer deployments start in Q4 2026.
- When can I buy an AMD Helios rack?
- AMD said Helios is in production now. Microsoft is the first named launch customer for H2 2026, OpenAI brings its Helios racks online starting Q4 2026, and Meta, Oracle, and Anthropic have signed multi-gigawatt commitments that ramp through 2027. Broader availability follows those hyperscaler deliveries.
- What chips are inside AMD Helios?
- Each AMD Helios rack pairs 72 Instinct MI455X accelerators with 18 sixth-generation EPYC Venice CPUs and AMD Pensando networking. The MI455X is AMD's new HBM4-based GPU with 431 GB per package; the Venice CPUs handle host duties and agentic orchestration.
- How much does one AMD Helios rack cost?
- Press coverage of the Advancing AI 2026 launch put a single AMD Helios rack at roughly 5.0 to 5.25 million dollars, in line with Nvidia GB200 NVL72 pricing. AMD has not published an official public list price, so the number will vary with configuration, networking, and hyperscaler volume discounts.
Try it
See the full spec sheet in AMD's Advancing AI 2026 press release