AMD shipped Helios on July 20 â its first rack-scale AI system â and landed Microsoft Azure as a headline customer alongside Meta, OpenAI, and Oracle. One rack: 72 MI455X GPUs, 31 TB of HBM4, 2.9 exaFLOPS of FP4 compute. Thatâs roughly twice Nvidiaâs GB200 NVL72 on both performance and memory. Three new Azure VM families are coming later this year. None are available yet, and pricing hasnât been announced. But the customer list alone signals this isnât vaporware.
What Helios Actually Is #
Helios is a fully integrated rack â not just a GPU. AMD bundles 18 compute trays into a single OCP Open Rack Wide chassis. Each tray holds four MI455X accelerators and one sixth-generation EPYC âVeniceâ CPU. The networking is handled by AMDâs own Pensando NICs, and the scale-up interconnect uses UALink â an open standard â rather than Nvidiaâs proprietary NVLink.
That open-standards bet is deliberate. Enterprise teams buying multi-vendor infrastructure can integrate Helios without committing to a proprietary fabric. Itâs a meaningful differentiator for hyperscalers that have grown increasingly vocal about vendor lock-in.
The Specs That Matter #
The MI455X is AMDâs CDNA 5 flagship: 432 GB of HBM4 per GPU, 19.6 TB/s of bandwidth per chip, and 40 PFLOPS of peak FP4 compute. At rack scale, those numbers stack into 2.9 exaFLOPS FP4 and 31 TB of total HBM4 memory â with 1.4 PB/s of aggregate bandwidth.
| Spec | AMD Helios (72x MI455X) | Nvidia GB200 NVL72 |
|---|---|---|
| FP4 Compute | 2.9 EFLOPS | 1.44 EFLOPS |
| Total Memory | 31 TB HBM4 | 13.5 TB HBM3e |
| Memory Bandwidth | 1.4 PB/s aggregate | ~0.96 PB/s aggregate |
| Scale-Up Interconnect | UALink (open standard) | NVLink (proprietary) |
| Form Factor | OCP Open Rack Wide | Proprietary |
The memory gap is the headline. Large model inference is memory-bound at scale â keeping trillion-parameter weights in fast memory requires capacity, not just throughput. Helios has 2.3Ã the HBM of Nvidiaâs flagship rack at this generation. Whether real-world inference throughput matches paper specs is a separate question, and one that wonât be answered until production hardware ships in 2027.
Microsoft Azure: Whatâs Actually Coming #
Microsoft will deploy Helios on Azure in H2 2026. Three new VM families were announced with the partnership:
ND MI455X v7â AI inference workloads** HDv2â agentic AI workloads HXv2**â HPC and chip design
None of these are provisionable today. Pricing hasnât been announced. Regional availability hasnât been announced. âH2 2026â is the only public commitment â Helios engineering samples are shipping to customers including Microsoft now, with mass production beginning Q2 2027. Watch for Azure preview signup announcements; they typically open before GA, as with H100 and GB200 preview access. For full technical details, see Tomâs Hardwareâs Azure Helios coverage.
The Software Side: ROCm Is No Longer the Excuse #
The standard knock on AMD for AI workloads has been ROCm. That excuse is wearing thin. ROCm 7.x supports PyTorch 2.7, JAX, vLLM, SGLang, and TensorFlow natively. ROCm 7.2 is the version where Ollama, llama.cpp, and vLLM reached CUDA-comparable behavior without manual patching.
For teams with significant CUDA codebases, AMD provides hipify-perl
and hipify-clang
â tools that automate most of the syntax translation from CUDA to HIP (AMDâs portability layer). The approach that works: port incrementally on a CUDA machine first, run HIP and CUDA side by side, then validate on AMD hardware. Itâs still work, but itâs tractable work. The official HIP porting guide is a solid starting point for scoping migration effort before committing.
The Customer Signal #
AMD landed Microsoft, Meta, OpenAI, Oracle, HPE, and the US Department of Energy as Helios customers. Metaâs deployment targets 1 gigawatt of capacity using MI450-series GPUs. Oracle committed to 50,000 GPUs starting Q3 2026. AMD is projecting tens of billions in datacenter AI revenue from Helios by 2027. Thatâs not a niche experiment â the hyperscalers are in with production commitments, which is what AMD has consistently lacked in the AI chip market until now.
What Developers Should Do Now #
For most teams, the practical timeline is 2027. But a few actions are worth taking now: Evaluate ROCm compatibility for your existing PyTorch or vLLM stack. ROCm 7.x support for major frameworks is solid enough to test against today â see theROCm documentation.Watch for Azure preview signups for ND MI455X v7 instances. When they open, early access will give you real benchmark data to compare against Nvidia alternatives.Factor open standards into 2027-28 infrastructure decisions. UALink and OCP rack designs avoid the NVLink dependency â relevant if youâre planning multi-vendor GPU deployments.
Helios is AMDâs most credible system-level challenge to Nvidia yet. The specs are compelling, the customer list is real, and ROCm has matured enough to take seriously. Production lands in 2027 â but the planning window is now. Full announcement details at CNBC and the TNW specs breakdown.