"AMD's Helios AI Rack Packs 72 GPUs and 31 Terabytes of Memory Into One Cabinet"

At the Advancing AI 2026 event in San Francisco this week, AMD CEO Lisa Su walked onto a keynote stage and laid out the company's most aggressive bid yet to reshape the AI infrastructure market. The centerpiece: Helios, a rack-scale system that crams 72 Instinct MI455X accelerators, 256-core EPYC Venice CPUs, and 31 terabytes of HBM4 memory into a single cabinet. Pricing starts at roughly $5 to $5.5 million per rack, with volume deployments expected in the second half of this year. It is, by any measure, an audacious piece of engineering — and one that signals AMD is done playing catch-up.

The MI455X at the heart of Helios is built on AMD's CDNA 5 architecture, fabricated on TSMC's cutting-edge 2-nanometer process. Each rack interconnects 72 of these accelerators into a unified coherent memory fabric, meaning any GPU in the rack can address the full 31 TB of HBM4 as if it were local. Four MI455X GPUs pair with a single Venice CPU per node, and the full rack delivers roughly 9 exaflops of FP4 performance. Those numbers put Helios in direct competition with Nvidia's Vera Rubin NVL72 — and on the spec sheet, AMD is arguably ahead on memory capacity and bandwidth per rack.

The EPYC Venice processor deserves its own moment in the spotlight. It's the first x86 server CPU built on TSMC's 2nm node, a manufacturing milestone that gives AMD a process lead over Intel's server roadmap and puts it on equal footing with the silicon that powers the latest mobile SoCs. Pairing a leading-edge CPU with a leading-edge GPU inside a single rack architecture isn't just about bragging rights — it means the host processor can feed the accelerators without becoming a bottleneck, which has been a persistent pain point in heterogeneous AI systems.

Lisa Su didn't present Helios alone. She brought OpenAI and Anthropic executives onto the stage to publicly endorse the platform, a move that carries real weight. These are the two labs at the frontier of large language model development, and their presence on an AMD stage — rather than an Nvidia one — signals that the GPU monopoly in AI training is genuinely cracking. OpenAI and Meta together have committed to 12 gigawatts of AMD accelerator capacity, and Microsoft Azure and Oracle are already named as early Helios customers deploying the racks in their cloud regions.

The Anthropic relationship runs deeper than a customer contract. AMD is investing up to $5 billion in the AI lab, a deal that blurs the line between hardware vendor and strategic partner. This kind of vertical integration — a chipmaker owning a meaningful stake in a frontier model developer — creates a tight feedback loop that purely transactional relationships can't replicate. Anthropic's engineers can tell AMD exactly what their next-generation models need from silicon, down to the instruction level, and AMD can design for those workloads before they're public. In an industry where the gap between announcing a GPU and having optimized software for real workloads can stretch into years, that feedback loop is worth more than the investment dollars.

The software story is where AMD has historically struggled, and the company knows it. ROCm, AMD's open-source GPU compute platform, has come a long way from its early days as a CUDA also-ran, but the gap remains real. At Advancing AI 2026, AMD paired the Helios hardware announcement with updates to the ROCm stack and the introduction of Pensando Vulcano, a networking fabric designed to keep 72 GPUs fed with data without the interconnect becoming the bottleneck. Getting the hardware specs right is table stakes; getting the software to extract those specs in production is the actual test, and it's one AMD still needs to pass at scale.

The pricing tells its own story. At $5 to $5.5 million, a Helios rack costs roughly what a single high-end Nvidia DGX system went for not that many years ago — but Helios delivers an order of magnitude more compute and memory. AMD appears to be pricing aggressively to gain footprint in cloud data centers, betting that once customers build their inference and training pipelines around ROCm and the MI455X, switching costs will do the retention work. It's the same playbook AMD used to break Intel's lock on the server CPU market with EPYC a decade ago, and it worked then.

There's a deeper architectural philosophy at work here that's worth appreciating. The industry spent years comparing individual GPUs — an A100 versus an MI250, an H100 versus an MI300 — as if you could understand a system's capabilities by looking at one component. Helios represents the maturation of a different idea: the rack is the computer. Memory, compute, networking, and cooling are designed as an integrated whole, not assembled from off-the-shelf parts. Hyperscalers think this way by necessity; Google's TPU pods and Nvidia's DGX SuperPOD are built on the same principle. AMD bringing this philosophy to a commercially available, partner-shippable rack design lowers the barrier for cloud providers and enterprises that want turnkey AI infrastructure without the integration tax.

One detail that's easy to overlook: 31 terabytes of HBM4 in a single coherent memory space. The largest frontier models today — the ones with hundreds of billions of parameters — need dozens of terabytes just to hold their weights in memory during training. Sharding a model across multiple racks introduces communication overhead that eats into utilization. A single Helios rack can hold an entire frontier-scale model in one coherent memory pool, which means less time spent waiting on cross-rack network transfers and more time spent actually computing. For labs training the next generation of models, that's the difference between a training run that takes three months and one that takes two.

What makes this moment particularly interesting is the constellation of customers AMD has lined up before the hardware is even shipping. OpenAI and Anthropic are the two most prominent independent AI labs; Meta is the largest company running open-weight models at planetary scale; Microsoft Azure and Oracle represent two of the biggest cloud platforms. That's not a beta program — that's a launch coalition. When all of them deploy Helios racks simultaneously in the second half of 2026, the aggregate installed base will create enough demand for ROCm tooling and optimization that the software ecosystem will have to catch up, fast. Network effects work for platforms, not just for social networks.

None of this means Nvidia is in trouble. The company's software moat — CUDA, cuDNN, TensorRT, the entire NVIDIA AI Enterprise stack — is deep and wide, and it will take years for AMD to match it. But competition in AI silicon is no longer theoretical. AMD has the architecture, the manufacturing node, the customer commitments, and the pricing to be a genuine second source for the hyperscalers. In a market where demand for AI compute is still outstripping supply by a wide margin, a credible second source isn't just good for AMD's shareholders — it's good for the entire industry. More competition means more supply, lower prices, and faster innovation cycles. After years of watching Nvidia print money with 70%+ gross margins on data center GPUs, the Helios announcement is the strongest signal yet that the AI hardware market is becoming a real market — with real competitors, real choices, and real consequences for whoever falls behind.

The Verge covered the Advancing AI keynote live (theverge.com). CNBC reported on the AMD-Anthropic $5 billion investment deal (cnbc.com). Wccftech and TechPowerUp have detailed spec breakdowns of the MI455X and Helios architecture (wccftech.com, techpowerup.com).