Chipping Away at Nvidia's Lead: Lisa Su's AMD Plan, Two Years On

In December 2024, Bloomberg's Emily Chang visited AMD's labs in Austin to ask Lisa Su, ten years into her time as CEO, how she planned to chip away at Nvidia's lead in AI. Almost two years on, AMD's own results let us check the plan against what happened, and ask what it means for UK organisations buying AI infrastructure today.

What Lisa Su said

The episode is part company profile, part race story, and Su's lines are worth hearing in full. The points that matter for anyone buying AI hardware:

• The pace. She called AI "the most transformational technology that I've seen in my career", ahead of the internet, PCs, mobile and cloud, and said the industry had made more progress in the previous 18 months than in the ten years before.
• The product. She showed the Instinct MI300, already in production and running work such as GPT-4o on Microsoft Azure and Meta's Llama models.
• The underdog question. Asked whether being seen as the next-best option to Nvidia was good enough, she said AMD has always "punched well above our weight" in technology. Bloomberg's chip reporter Ian King described AMD as "a solid number two".
• Custom chips. On hyperscalers designing their own silicon: "This is a huge market... There's like more than enough chips for everybody to build."
• Export controls and supply. She accepted there is "a national security element" to limits on AI chip exports, while wanting "as open a market as possible", and stressed resilience in a supply chain concentrated in Taiwan.

The rest covers her ten years turning AMD around (the Zen architecture was "the most important moment"), the Mercedes Formula 1 partnership, a Porsche collection named after AMD products, and her advice to young engineers: "run towards problems".

Two years on: the scorecard

AMD's third-quarter 2026 results, published on 3 November, are the clearest check. Data Center revenue reached a record $8.5bn, up 97% year on year, driven by EPYC server processors and the launch of the Instinct MI450 Series. Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure are running models on the first Helios racks, AMD's answer to Nvidia's rack-scale systems. AMD expects around $17bn of total revenue in the fourth quarter.

Helios packs 72 Instinct MI455X GPUs into one rack, each with 432 GB of the newest HBM4 memory, for 31 TB per rack. AMD claims it beats Nvidia's Vera Rubin NVL72 on compute, memory and tokens per dollar. Those are AMD's numbers, and Nvidia remains far larger, with the deeper software ecosystem. But "solid number two" now means a second full platform that the biggest AI companies run in production, not a niche alternative.

What it means if you are buying AI infrastructure in the UK

1. You have a real choice. For many workloads, especially serving large language models with long contexts, AMD's larger memory per GPU can mean fewer GPUs per model. That is worth testing, not assuming.
2. Software is the deciding question. Code written in PyTorch, or served with vLLM or SGLang, usually runs on AMD's ROCm software with little change. Custom CUDA code and Nvidia-only tools are where the work is.
3. Two vendors give you leverage. Teams that can run on both negotiate better prices and are less exposed when one vendor's lead times stretch.
4. Supply risk has not gone away. As Su said, the industry depends on a concentrated supply chain. Plan lead times into AI projects whichever vendor you choose.

For the full engineering detail, with the platforms compared layer by layer, memory arithmetic, the CUDA-to-ROCm software map, a porting workflow with commands, running both vendors on one Kubernetes platform and a fair benchmarking method, read our technote: AMD Instinct vs NVIDIA for AI infrastructure: evaluate, port and run both.

How Workstation helps

We design and run GPU infrastructure for UK organisations and help teams evaluate AMD alongside Nvidia: a like-for-like benchmark of your own model on both, a check of your code for CUDA-only dependencies, and a Kubernetes platform that runs either. See our GPU private cloud guide, which includes AMD's official Helios video, compare leasing and buying at gpuleasing.co.uk, or talk to us.

Quotes are from the Bloomberg Originals video. Company figures are from AMD's announcements and are vendor claims, not independent benchmarks. This post is independent and not endorsed by Bloomberg, AMD or Nvidia.

Acknowledgement of rights: AMD, Instinct, EPYC, ROCm, Helios and Zen are trademarks of Advanced Micro Devices, Inc. NVIDIA and Vera Rubin are trademarks of NVIDIA Corporation. Bloomberg is a trademark of Bloomberg Finance L.P. Other names are the property of their respective owners.