AMD expects the datacenter AI accelerator market to grow at more than a 45 percent compound annual rate between 2025 and 2030, reaching over $1.4 trillion. CEO Lisa Su disclosed the revised total addressable market figures at the company's annual Advancing AI event in Silicon Valley this week.

The projections anchor a roadmap that includes the "Altair" MI400 architecture — specifically the MI455X variant — which debuts in AMD's "Helios" rackscale server designs later this year through multiple OEM and ODM partners. The "Venice" Epyc 9006 processors are also expected before year-end.

How fast is the inference shift happening?

Su said AI training workloads have increased their compute needs by a factor of 5X every year since 2020, with no sign of this abating. The larger driver, she argued, is the shift from training at a handful of companies to inference, where enterprises rent GenAI models through APIs or buy systems to run licensed or open source models themselves.

The split within the AI accelerator market was roughly 50-50 between inference and training last year, roughly 40-60 in favor of training back in 2024, and Su expects it to reach 60 percent inference and 40 percent training this year.

Token consumption data supports the trend. Citing the State of the AI Economy report from Exponential View, Su showed worldwide GenAI token consumption or generation reached an estimated 35 quadrillion tokens per month in February. Given the exponential curve, the figure should be well past 50 quadrillion tokens per month by the end of July.

Demand for AI processing still exceeds supply, which is why pricing for CPUs, GPUs, DRAM, flash, switch ASICs and optical components is rising along a similar curve.

What do the TAM numbers show?

AMD's broader compute TAM — spanning datacenter, client and embedded devices — is projected to grow at around a 40 percent CAGR between 2025 and 2030, to more than $2 trillion. These are AMD's internal projections.

Su broke the datacenter CPU market into three buckets:

• General purpose CPUs for back-office applications, databases, web infrastructure and data analytics
• Host processors for AI cluster nodes
• Agentic AI clusters running sandboxes where models generate and execute code

The agentic AI CPU segment did not exist last year. It will be almost as large as the general purpose CPU TAM in 2026, and by 2030 it will be 5.2X larger.

"When I think about where AI is today, the biggest change that we see is we are no longer talking about what might be possible," Su said. "We are actually seeing how AI can have real and significant impact across every industry and every part of our personal lives."

She added: "At AMD, what we are focused on is building the technology, the roadmaps, and the partnerships. And there has never been a more exciting moment than today for our 30,000 plus engineers."

What does the spending mix look like?

The cumulative datacenter compute TAM across 2025 through 2030 reaches $6.45 trillion. Datacenter compute is 8.3X larger in aggregate than embedded and client compute combined over those six years, and 10X larger in 2030 alone.

AI accelerators will far outstrip CPU spending despite a roughly 1:1 device ratio in the long run. The ratio of datacenter GPU/XPU spending to datacenter CPU spending was 7.9X in 2025, is expected to be 6.8X in 2026, and averages 6.5X across the six years.

Extracting training from inference across GPUs, XPUs and CPUs yields a cumulative $1.42 trillion for training-side compute over six years, versus $3.82 trillion for the inference side — including inference GPUs/XPUs, their host CPUs, and all CPUs deployed in agentic AI clusters. That puts the inference-to-training spending ratio at 2.7X over the period, up from 1:1 in 2025 and 1.6X expected in 2026, reaching 3.2X by 2030.

That 3:1 ratio matches what many analysts projected for the earlier machine learning era.

What happens to AMD's own forecast?

AMD has not updated its own revenue expectations. With roughly $600 billion in expected TAM in 2030 alone and upward pressure across the intervening years, the company's 2026 revenue forecast may follow suit — and AMD could plausibly publish a CAGR out to 2030, given the visibility into customer spending it holds while demand exceeds supply for everything it sells into the datacenter.