Advanced Micro Devices CEO Lisa Su used a high-profile appearance in Taipei on Thursday to reframe the terms of competition in artificial intelligence infrastructure — arguing that Nvidia's dominance is most vulnerable not in the GPU arena it has long controlled, but in the economics of AI inference.
Speaking at the Commonwealth Magazine 45th Anniversary Flagship Forum at the Taipei Marriott Hotel, Su predicted that the global data center AI infrastructure market will surpass $1 trillion within three to four years, but said the real battleground will be fought on cost efficiency, CPU supply, and system-level architecture — areas where she believes AMD holds a structural edge.
"Training is necessary, but that's not where you see the return on investment," Su said. The commercial payoff, she argued, comes from inference — the moment users actually interact with AI systems — and that shift is reordering the entire competitive landscape.
Why Inference Changes Everything for AMD
For much of the past two years, the AI industry's attention has been almost entirely fixed on GPUs, driven by the massive parallel compute demands of training large language models. That focus benefited Nvidia enormously. But Su argued that the industry is now entering a different phase — one defined by inference workloads, agentic AI systems, and a far broader mix of chip requirements.
In that environment, customers need CPUs, GPUs, and ASICs — not a single dominant solution. AMD's ability to supply all three simultaneously, she said, is a structural advantage that Nvidia cannot easily replicate.
The key performance metric in this new phase, Su said, is "tokens per dollar" — how much AI output a customer can generate per unit of cost. AMD's GPU roadmap has been optimized for inference, with particular emphasis on memory capacity and bandwidth. She cited those characteristics as the primary reason major clients including OpenAI and Meta have adopted AMD hardware for their inference workloads.
A CPU Shortage Nobody Saw Coming
One of Su's most pointed observations concerned a corner of the market that had been largely overlooked: server CPUs.
Six to twelve months ago, she noted, almost no one was talking about a CPU shortage. That has changed dramatically. As inference workloads scale and agentic AI applications proliferate, CPUs have moved back to the center of AI infrastructure planning — a domain where AMD has long been competitive.
Su said current CPU demand has already exceeded projections made just a year ago, with supply running tight across the industry. AMD is ramping production every quarter this year and has committed to further capacity expansions through 2027 and beyond.
The numbers she cited were striking. Over the past three to four years, the CPU market grew at only around 3 to 4 percent annually, sidelined by the GPU boom. Su now projects annual CPU market growth exceeding 35 percent over the next five years, driven entirely by the inference and agentic AI wave that is only beginning to build.
AMD's Position on ASICs: Competition or Opportunity?
As custom application-specific integrated circuits — ASICs — gain traction among hyperscalers looking to cut costs, AMD has faced questions about whether that trend poses a competitive threat. Su's answer was measured but revealing.
She acknowledged that ASICs can reduce costs in targeted use cases and represent a legitimate technical option. But she framed AMD's core philosophy as matching the right chip to the right application — a mix-and-match approach rather than a winner-take-all competition.
More broadly, Su characterized the rise of ASICs as a net positive for AMD. Overall AI demand is large enough, she argued, to support multiple competing technology approaches at the same time. The real constraint on growth is not rivalry among chip types — it is the pace at which the industry can scale infrastructure to meet demand.
The Next Frontier: Physical AI and Real-Time Computing
Looking beyond data centers, Su identified physical AI — robotics and edge infrastructure — as the next major growth wave, and one that plays to AMD's particular strengths.
These applications demand not just raw compute power but real-time processing with strict latency limits. AMD is addressing this by combining CPU capabilities with dedicated AI engines, building on assets acquired through its 2022 purchase of Xilinx. That deal gave AMD a portfolio of FPGAs and adaptive computing technology already deployed in medical devices, automotive systems, aerospace, and defense applications — a foundation Su indicated will be central to AMD's physical AI push.
What AMD's Strategy Signals for the Broader Market
Su's remarks amount to a clear statement of competitive intent: AMD is not simply trying to close the gap with Nvidia on GPU benchmarks. It is positioning itself for an architectural shift in which the companies that prosper will be those that can serve the full stack of AI infrastructure needs — inference efficiency, CPU capacity, system-level integration, and an open software ecosystem.
Whether AMD can execute on that vision at scale remains to be seen. But the timing of Su's pitch is deliberate. As AI investment matures from the training phase into real-world deployment, the industry's cost and performance priorities are shifting in ways that AMD is betting will favor a diversified chip portfolio over any single dominant platform. （Related： Opinion | At the Xi-Trump Summit, the CEOs Upstaged the Presidents ｜ Latest ）
Original Article in Chinese
