Lisa Su on AMD's Helios launch, the Anthropic partnership, and why heterogeneous compute wins

Lisa Su on Helios, Anthropic, and the heterogeneous compute bet

AMD's biggest AI event of the year produced two announcements that matter: the Helios rack-scale AI system, powered by the MI 450 accelerator, and a partnership with Anthropic to deploy up to two gigawatts of AMD Instinct MI 450s in a deal valued at $5 billion. Su describes the Anthropic win as something AMD had been working toward for a long time, waiting for the right intersection of AMD's technology and where Anthropic's Claude platform had reached.

The detail that travelled furthest on the day came from Anthropic's Tom Brown, who described on stage how Anthropic had been running workloads on MI 355s and optimising for the hardware largely without AMD's help. Su confirms the story. A few months ago her team flagged that Anthropic appeared to be doing serious work on the 355s. When AMD offered to assist, the answer was essentially: they're already handling it, using Claude to do the optimisation work themselves. Su calls it a striking example of AI as a force multiplier on hardware bring-up — and a direct answer to the long-running question about whether AMD's software ecosystem can close the gap with Nvidia's CUDA.

“I have to say, we have very much wanted Anthropic on AMD for a long time. And we just had to find the right intersection point. The highlight for me was the MI 355 story — a few months ago, my guys were like, 'Anthropic is on 355s and they don't really want our help. They're able to do it with Claude.' And I'm like, wow. That's incredible. That's a lot about how the ecosystem has evolved over time.”

Helios and the heterogeneous argument

Su's consistent strategic line is that no single chip wins every workload. CPUs, GPUs, FPGAs, and the networking that ties them together each serve different parts of the pipeline, and customers who try to run everything on one architecture leave performance and efficiency on the table. Helios is AMD's attempt to make that argument concrete at rack scale, pulling together chiplet design, networking, and CPU-GPU integration into a single system solution. She traces the ideas behind it to decisions made three to four years ago and says the roadmap already extends to MI 500, MI 600, and beyond.

The Xilinx acquisition brought physical AI and FPGA capability into the portfolio. Pensando and ZT Systems filled out the networking and systems layers. Anush Elangovan, who joined via acquisition, is specifically credited with expanding ROCm support for developers who want to run on AMD hardware.

Software ecosystem

Su acknowledges the ecosystem question directly without over-claiming. The Anthropic story matters partly because it suggests that as AI models get better at code optimisation, the cost of porting workloads to non-Nvidia hardware falls. That flywheel — better models making software bring-up easier, which attracts more partners, which produces more model-driven optimisation — is what AMD is betting tightens over the next few years. The partnership with Cerebras on disaggregated inference, pairing different compute engines to prefill and decode phases separately, is another piece of the same thesis: match the right silicon to each part of the pipeline rather than forcing everything through a single architecture.

Consumer to cloud

Su argues the talent and adoption pipeline matters as much as the enterprise deal flow. Developers who start on Ryzen AI Max or Radeon consumer hardware and move up to cloud instances are the same people making infrastructure decisions at frontier labs. AMD's investment in AI PCs and small robotics form factors is partly about lowering the barrier to entry so that AMD is the first hardware a new researcher touches. She also flags local AI inference as underrated — not every token needs to travel to a hyperscaler data center, and on-device compute is a long-term bet she is clearly not backing away from.

On R&D cadence, Su is direct: the products launching today were ideas that barely existed three to five years ago, and the decisions being made now will define what ships in 2029 and 2030. The pressure she feels is not existential anxiety but a constant push to compress timelines. Her summary of every customer conversation is the same two words: more compute, faster.