AMD unveiled Helios — a rack-scale?rack-scale: An approach where the product is a whole server rack — with its networking, power and cooling — rather than a single chip. system for training and running large AI models — at its Advancing AI conference in San Francisco on July 23, 2026. CEO Lisa Su called it the industry's highest-performance AI rack and said it will ship later in 2026, taking direct aim at Nvidia's grip on the data center.
Key takeaways
- Helios ships later in 2026, built on Instinct MI450 series GPUs.
- Lisa Su claims Helios beats Nvidia's Vera Rubin on several performance metrics (per The Register).
- Anthropic will deploy up to 2 GW of MI450 compute in Helios racks, with the first gigawatt in H1 2027.
- AMD will invest up to $5 billion in Anthropic (per CNBC).
- The new Venice-X CPU (2027): 96 cores, 1,152 MB of 3D V-Cache, 5.15 GHz boost.
- Su projects the AI accelerator market at roughly $1.4 trillion by 2030.
Helios: the rack as a single compute unit
Helios is not a single chip but an entire server rack engineered as one coherent machine for training AI models at gigawatt?gigawatt: A gigawatt is a billion watts of power — here a measure of a data center's energy draw, not a chip count. scale. AMD first revealed it in 2025 and showed working hardware at CES 2026. The July 23 Advancing AI conference was the moment the company moved from teaser to a concrete shipping timeline within this year.
The rack-scale approach is a direct answer to Nvidia's strategy, which since the NVL72 systems sells not bare GPUs but pre-integrated, densely networked cabinets. In that model the unit of sale and the measure of performance is no longer a single chip but the whole rack, with its networking, power and cooling.
Su claims Helios edges out Nvidia's upcoming Vera Rubin platform on several performance metrics, as The Register reports. That is a vendor claim ahead of independent testing, so it should be read with caution.
MI450 and the strategic Anthropic deal
At the heart of Helios sit Instinct MI450 series GPUs. The largest confirmed customer is Anthropic, which — per AMD's announcement — will deploy up to 2 gigawatts of MI450-based compute in Helios racks.
The first gigawatt is set to come online in the first half of 2027. The scale is telling: capacity is measured in gigawatts rather than card counts, a sign that power, not silicon, has become the binding constraint for AI data centers.
The deal has a financial dimension too. According to CNBC, AMD will invest up to $5 billion in Anthropic, tying the chip supplier to one of the leading language-model labs.
It echoes the Nvidia–OpenAI playbook: the hardware maker takes an equity stake in the company buying its accelerators. Beyond Anthropic, AMD lists Microsoft, OpenAI, Meta and Oracle as customers — precisely the hyperscalers?hyperscaler: One of the largest cloud and data-center operators, e.g. Microsoft, Google or Amazon. Nvidia is competing for today.
Venice-X and the race for $1.4 trillion
Alongside the GPUs, AMD announced the Venice-X data center processor, planned for 2027. It is set to offer 96 cores, a striking 1,152 MB of 3D V-Cache?3D V-Cache: An AMD technology that stacks extra cache vertically above the CPU cores. and a 5.15 GHz boost clock.
Such a large cache is aimed squarely at AI workloads where the CPU handles agentic inference — multi-step, tool-calling tasks that Su described as dozens of steps for a single query.
Su framed the whole event with a market forecast: by 2030 the AI accelerator market should reach roughly $1.4 trillion, approaching the size of today's entire semiconductor industry.
That number explains AMD's determination — even second place in a market that large means revenue in the hundreds of billions. The open question is whether Helios delivers real performance and availability at the volumes where Nvidia currently sets the terms.
Why it matters
AMD's problem was never raw chip performance but software and the availability of ready, integrated systems. Helios attacks that second front: by selling the whole rack, AMD removes the burden of assembling networking, power and cooling — the very integration that gave Nvidia its edge for years.
The Anthropic deal matters more than the spec sheet, because it shows a top AI lab is willing to bet two gigawatts of production capacity on AMD hardware. For the market that means a real, not merely stated, alternative to Nvidia — and for customers, potential pricing pressure and less single-vendor lock-in.
The decisive factors remain independent performance testing and whether AMD can ship racks in the promised quantities and timeframe. A claim of the highest-performance rack, without external benchmarks, is still marketing rather than a verdict.
What's next?
- Helios is due to ship later in 2026, when the first independent benchmarks against Nvidia's Vera Rubin should appear.
- Anthropic will bring the first gigawatt of MI450 compute online in the first half of 2027.
- The Venice-X processor arrives in data centers in 2027, adding a CPU layer to the platform.
