Lisa Su on Why Inference Is AI's Real Inflection Point
AMD's Lisa Su argues AI has crossed from training to inference — the everyday running of models — and that meeting accelerating demand will require abundant compute, open ecosystems, and capacity committed 12 to 24 months ahead.
Training Was the Build-Out. Inference Is the Payoff.
AI's real inflection isn't building bigger models — it's inference, the everyday running of them, which Su pegs as a $1.4 trillion accelerator market by 2030.
And inference is really the way we turn, you know, AI from a technology to something that really changes the way we do business, the way we do research, the way we do healthcare, all of those aspects of it.
The Curve Got Steeper Than Anyone Planned For
Su's biggest surprise is pace: markets she used to re-forecast yearly are now shifting monthly and quarterly, and agents have made the adoption curve steeper than anyone in the industry expected.
I've never seen a technology adoption curve like what we're seeing with AI right now, and agents are just taking it to a whole new level.
One Employee, a Thousand Agents Each
AI is a tool that multiplies people — give each of ten employees ten, a hundred, or a thousand agents and throughput explodes, while the human stays the one who decides right from wrong.
But if each of those 10 employees had, you know, 10 agents or 100 agents or even 1,000 agents, they can be much much more productive.
The Compute Flywheel: More Compute Is More Intelligence
More compute buys more intelligence, which solves more problems and pulls in still more compute — a self-reinforcing flywheel that Helios spins faster with a claimed 30x jump in performance.
The more AI can do, the more compute you need, the more compute you have, the more intelligence you have, the more problems that you can solve.
No One Chip — or One Company — Can Do It All
Because the world's problems are heterogeneous, Su argues no single chip or company can do everything — AMD's strategy pairs an end-to-end lineup (CPU, GPU, FPGA, ASIC) with an open developer ecosystem.
We're all different, our companies are different, our problem sets are different, what we're trying to solve is different.
Read the Long Arc, Not the Quarter
To investors spooked by hyperscaler CapEx, Su's answer is to read the long arc: compute must be committed 12 to 24 months ahead, ROI is already showing up inside AMD, and every big customer is asking how to go faster.
I mean every conversation I have with every large customer is how can we go faster?
We're All in This Together
The up-to-$5 billion Anthropic investment is Su's proof that AI is built by an intertwined ecosystem — hardware, software, models, and data centers aligned — and she argues that entanglement is a feature, not a risk.
We believe in Anthropic. We are very happy to be making a strategic investment of up to 5 billion as we work together and this is one of the things about the AI ecosystem is we're all in this together.