Invest3 publishers3 min readPublished
AMD stock surges as agentic AI shifts workloads from GPUs to CPUs, powering Meta's Muse and Dots
AMD rose 32% and Intel 21% in a month as Meta's Muse and OpenAI's Dots ran long-running AI agents on CPUs. The cash behind the move depends on how many agent sessions run and whose processors they run on.
The Investor · Invest desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Muse tells users it runs on an AMD-powered computer, and Dots has said its virtual computer uses AMD's EPYC server CPU.
- Benchmarks and the agents' own answers show Muse's virtual computer using two CPU cores and Dots using nine, on a chip line that reaches 192 cores.
- Dots appears to run on an EPYC 9V74 that resellers sell for under $3,000, and Muse reportedly uses an EPYC 9D25 that costs less secondhand.
- In March Arm unveiled an agent CPU of its own, with Meta signed as its first customer. Arm's shares have more than doubled since.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost Each agent occupies a slice of a chip under $3,000 while Nvidia GPUs can top $30,000 apiece and ship by the hundreds, so CPU vendors need very large session counts to approach GPU-sized revenue.
- exposure AMD's place inside Muse rests on what the app says about itself, at a buyer that was first to sign for Arm's agent CPU, so the share can shift to Arm without any public contract change.
- contradiction Signal65's Shrout expects agents to pull work off GPUs while Futurum's Newman has GPUs still doing the thinking, and only Shrout's version takes revenue from Nvidia.
Split a 192-core EPYC evenly and, before any overhead, it would host about 96 Muse sessions or about 21 Dots sessions [14]. Dots asks for 4.5 times the cores Muse does [15]. Both companies carve servers into virtual machines so that one box serves scores of users [7].
The CPU cost per agent is small, or rather small per hour. Futurum Group CEO Daniel Newman explained that each agent needs a computer of its own, one that can keep going in the background for hours or even days [22]. An agent left running for a day holds its cores for the whole day, so demand scales with sessions multiplied by duration [22]. The CNBC report does not give user counts or session lengths for either agent.
"As more agents are created and developed, and more people start to use them for more tasks, it's going to start to shift the workload away from GPUs and onto CPUs," said Ryan Shrout, president of Signal65, which consults on AI hardware and costs [21]. Newman put it differently. "CPUs are actually performing the workflows while GPUs are doing the thinking," he said [23].
One outcome is that agents add CPU demand on top of GPU demand, lifting AMD and Intel without costing Nvidia a sale. Another is that cloud providers move agent virtual machines onto their own Arm-based chips, which nearly every major one is developing [8]. The third is that Nvidia's Vera, a CPU it says it built for agents and sells in a CPU-only rack, takes the sockets [3]. Of Vera, Mizuho Securities analyst Jordan Klein said, "It's just a new product and they're going to be competing against a lot of different players," adding that AMD currently has the "best product" [4]. I think the evidence fits the first outcome best, because Newman's account has the GPU still doing the inference for every session [23]. The counter-case is Shrout's. If agent work does move off GPUs, a re-rating of AMD and Intel against Nvidia is earned.
Meta's own statement limits what AMD can count on. "We take a diverse approach to our hardware and are largely CPU-agnostic by design, which gives us the most flexibility in acquiring capacity," a Meta spokesperson said [11]. OpenAI also buys from several CPU providers [12].
AMD's June-quarter data center figure implies total quarterly revenue above about $11.2 billion [16] and a year-earlier data center line below $3.35 billion [17]. Annualized, the whole data center business, GPUs included, runs near $26.8 billion, about 13% of the $200 billion CPU market Nvidia forecasts for 2030 [18]. The report does not break out CPU sales within that line, and for Intel it gives only the share move [9]. "The conversation has changed quite a bit over the last eight to 10 months in terms of agentic usage," said Dan McNamara, AMD's senior vice president and general manager of compute and enterprise AI, who added that CPU sales are going to "really ramp" [20].
What to watch
- AMD's next quarterly report, and whether it breaks out CPU sales inside data center revenue as McNamara's "really ramp" forecast is tested.
- Any sign that Meta moves Muse sessions onto the Arm agent CPU it was first to buy.
- A named cloud customer for Nvidia's CPU-only Vera rack.