AMD's new Helios rack takes direct aim at Nvidia

For the past two years, hyperscalers hunting for rack-scale AI muscle basically had to take whatever Nvidia offered at whatever price Nvidia asked.

AMD is officially crashing that party. At its packed Advancing AI conference in San Francisco, CEO Lisa Su took the wraps off Helios, a massive rack-scale system built to go toe-to-toe with Nvidia's flagship infrastructure.

Packing 72 MI455X GPUs into a single frame

Helios isn't just a collection of servers; it's a co-designed supercomputing unit combining 72 Instinct MI455X GPUs, EPYC Venice CPUs, Pensando networking gear, and AMD's ROCm software stack.

The raw specs are staggering: 2.9 exaflops of FP4 compute, 31 TB of next-gen HBM4 memory, and 260 TB/s of scale-up bandwidth.

AMD claims Helios outperforms Nvidia's Vera Rubin NVL72 across several critical baselines, offering up to 15% more AI compute, 50% more HBM memory capacity, and 50% more scale-out bandwidth.

AMD also teased its next data center processor, the Venice-X CPU, slated for release in 2027.

Big tech lines up by the gigawatt

First revealed in 2025 and showcased at CES earlier this year, Helios is scheduled to begin shipping later this year, and the customer lineup is stacked.

OpenAI, Meta, Oracle, Anthropic, and Microsoft have all committed to deploying the rack system. Microsoft CEO Satya Nadella confirmed Azure will expand its fleet with Helios, while Anthropic announced a multi-gigawatt partnership to deploy up to 2 gigawatts of GPU capacity on the platform.

Su attributed this exploding compute appetite to agentic AI, arguing that autonomous agents requiring multi-step reasoning and endless tool calls are driving a structural leap in GPU consumption.

AMD now projects the AI accelerator market will hit $1.4 trillion by 2030, roughly rivaling the size of today's entire semiconductor industry.

Why it matters

For frontier AI labs and enterprise cloud providers, Helios provides the first credible, full-scale alternative to Nvidia's rack monopoly.

Having 50% more HBM capacity is a huge deal for long-context inference and multi-step agent reasoning, where memory footprint often becomes a harsh bottleneck before raw compute does.

For the broader industry, the battleground has definitively shifted from standalone silicon to integrated rack-scale systems.

We think AMD played this smartly by packaging networking, CPUs, and software into an open ecosystem while locking in headline-grabbing commitments from Anthropic and Microsoft right out of the gate.

That said, the real test won't be benchmark slides. It will come down to cluster reliability under grueling training runs and whether AMD's ROCm software can provide the turnkey developer experience that CUDA has spent two decades refining.