AMD CEO Lisa Su on AI and Nvidia: "I'm not a believer in moats

amdnvidiagpuai-chipssemiconductortsmcmi300h100rocmsupply-chain

Abstraction: Lisa Su interview on AMD competing with Nvidia in AI chip market

Key points:

AMD's MI300 chip targets LLM training and especially inference workloads; Lisa Su claims it is competitive with Nvidia's H100 and designed with inference as the larger future market
Su rejects the idea that Nvidia's CUDA moat is durable: "I'm not a believer in moats when the market is moving as fast as it is"; AMD's strategy is open software (ROCm) and PyTorch 2.0 hardware-agnostic compatibility
GPU supply is tight globally due to AI demand; supply/demand broadly balanced except at the high-end GPU tier powering LLM training/inference
CHIPS and Science Act is a "fantastic" step but semiconductor fab buildout takes years; AMD manufacturing diversification underway but still relies on TSMC for leading-edge nodes
AMD uses AI internally for chip design, test generation, and manufacturing diagnostics — "copilot functions" easily adopted, pure generative AI still requires human verification
On regulation: not the chip alone but chip + software + model safeguards; AMD willing to participate in that ecosystem; export controls already exist for highest-performing GPUs

Connections: Amd · Lisa Su · Nvidia · Tsmc · AI Chips · Semiconductor Industry · GPU Market