Geoffrey Seiler: Buy AMD, Sell Arm Holdings

Seiler's split verdict rests on the idea that agentic AI shifts the hardware mix in data centers. As AI agents handle sequential logic and connect to tools and external APIs, CPUs become more important, and the GPU-to-CPU ratio is projected to go from 8:1 for LLM training to 4:1 for inference to 1:1 for AI agents. AMD has already built purpose-built CPUs for agentic AI and projects the data center CPU market will reach $220 billion, with the company aiming for 50% share. It has also signed two $100 billion deals with OpenAI and Meta Platforms, plus a sizable deal with Anthropic.

Arm Holdings, by contrast, has long licensed its simpler instruction set architecture, which uses less power and generates less heat, making it dominant in smartphones. Its architecture is also behind custom data center CPUs from Nvidia, Alphabet, and Amazon. This year, Arm decided to build its own chips and enter the server CPU market. In March, it projected 15% share of what could become a $100 billion market by 2031, $25 billion in revenue including $15 billion from server CPUs, and $9 in EPS. AMD and Nvidia have since projected the market will exceed Arm's initial target.

Seiler's sell case is largely about valuation and margin structure. Arm trades at a forward P/E of nearly 140 times and faces declining smartphone volumes. Its premium may have been justified as an IP company, but as it becomes a chipmaker, gross margins will compress. The outcome hinges on whether Arm's server CPU push can justify its multiple and whether AMD can turn its projections into results.