Semiconductor analyst Jeff Pu has raised his forecast for the AI accelerator market to $1.4 trillion by 2030, up from an earlier $1 trillion estimate, and now expects the market to already reach $1 trillion by 2028, according to a note relayed by chip-industry commentator Jukan on X. Pu's note, focused on AMD's AI roadmap, also lifted the server CPU total addressable market to more than $220 billion by 2030, with agentic AI workloads expected to make up roughly half of that demand and the ratio of CPUs to GPUs in data centers tightening toward 1:1. On AMD's own hardware, Pu highlighted the MI455X accelerator as a major step up from its predecessor, the MI355X, citing 34 times higher throughput and 18 times lower cost per generated token, and said it already beats Nvidia's B200 chip on some large language model inference workloads. Looking further out, Pu said AMD's next-generation MI500 chip will bring optical interconnects, a faster way of linking chips together using light instead of electrical signals, putting it ahead of Nvidia's upcoming Rubin Ultra chip in areas like high-bandwidth memory, four-die packaging, and how many chips can be linked into a single system. He also flagged AMD's Venice server CPU as the strongest option for AI workloads, with 2 to 2.2 times higher throughput than competing chips, and estimated around 300,000 CoWoS advanced packaging units will be built for AMD in 2027, which would lift the average price AMD can charge per chip. Pu kept his price target and estimates unchanged despite the upgraded market forecasts.
Why it matters · A sell-side analyst just raised AMD's addressable market by 40% and says its next two chip generations already beat or match Nvidia on cost and some workloads, which matters because AMD's stock has traded for years on the promise of catching up rather than proof that it has.
Worth asking · Jeff Pu's note reads bullish across the board for AMD, bigger TAM, faster chips, better margins ahead, but sell-side estimates on AI capex have been revised up all year without much scrutiny of whether that spending pays off. Do you buy AMD's roadmap as a real threat to Nvidia's lead, or is this another round of analysts extrapolating the current AI boom too far forward?