AMD projects the datacenter AI accelerator market will grow at a 45 percent compound annual rate from 2025 to 2030, surpassing $1.4 trillion, with inference pulling the bulk of the revenue mix. CEO Lisa Su laid out the revised forecast at AMD's Advancing AI event in Silicon Valley this week.

The company also lifted its broader compute total addressable market to more than $2 trillion across datacenter, client, and embedded silicon, a 40 percent CAGR over the same six-year window.

Su framed the shift as structural, not cyclical. "The biggest change that we see is we are no longer talking about what might be possible," she said. "We are actually seeing how AI can have real and significant impact across every industry and every part of our personal lives."

Where does inference sit in the mix?

Su split her accelerator TAM between training and inference, and the inference wedge has grown meaningfully. Last year's mix was roughly 50-50; this year she expects about 60 percent inference and 40 percent training. By 2030, AMD's figures imply a 3.2-to-1 split for inference over training.

The full six-year split lands at 2.7-to-1, with about $3.82 trillion in cumulative inference spending against $1.42 trillion for training. Token consumption, per Exponential View's State of the AI Economy report, hit 35 quadrillion per month in February and likely crosses 50 quadrillion by month-end.

Su also flagged that AI training compute has grown by roughly a factor of five every year since 2020, a curve that has yet to flatten. The only question that matters is whether demand continues to justify the hundreds of billions of dollars in datacenter capacity now under construction.

What about the CPU side?

AMD broke datacenter CPU revenue into three buckets: general-purpose silicon for databases and web stacks, host processors paired with accelerator nodes, and a new agentic AI category for sandbox clusters running Python-heavy inference work. The agentic segment was effectively zero a year ago.

By 2026 it nearly matches the general-purpose bucket. By 2030 AMD expects it to be 5.2 times larger, reviving server CPU revenue in a fashion the industry has not seen.

The silicon itself is not exotic. Agentic servers demand different feature balances, not fresh architectures.

Across the 2025-to-2030 window, datacenter compute spending tops $6.45 trillion, against roughly $780 billion for embedded and client silicon stacked. Datacenter finishes the period at ten times the size of those two other segments combined.

The ratio of accelerator spending to CPU spending was 7.9-to-1 in 2025, eases to 6.8-to-1 in 2026, and settles near 6.4-to-1 for the back half of the forecast, in AMD's reckoning.

Will AMD's revenue track?

AMD has not refreshed its own revenue guidance alongside the new TAM charts. With $600 billion of incremental TAM by 2030 and demand running ahead of supply across CPUs, GPUs, DRAM, flash, switch ASICs, and optical components, an update looks overdue.

Coming products include the "Altair" MI400 architecture in the MI455X variant, deployed through AMD's "Helios" rackscale platform with OEM and ODM partners, along with "Venice" Epyc 9006 server processors. Both families land before year-end.

If customers continue to outbid each other for capacity, AMD will have little reason to underwrite its own forecast below the trajectory it is showing the rest of the industry.