AMD Stock and Meta Muse: Is the AI Agent CPU Boom Real?

Muse is Meta’s personal AI agent. Meta says each Muse user gets a dedicated virtual machine, or VM: a private workspace carved from a physical server. It can keep files, run tools and work in the background. That changes the hardware question. A chatbot can answer a prompt; a persistent agent also needs a computer to coordinate the work between prompts.

The CPU, a general-purpose processor, runs that workspace and tools such as a browser or code interpreter. GPUs elsewhere do much of the AI model calculation. A VM can be assigned two virtual CPUs without owning two physical chips, and an idle VM uses less processor time than one running several tasks. The useful numbers are active agents, concurrent tasks and CPU time per task. Meta has not published them.

Our hypothesis: cheaper inference—running the AI model—could encourage agents to do more work and use more CPU time. For AMD, the opportunity depends on how often agents run, how much processor time each job takes, and how much of that work AMD wins.

AMD says Meta already runs millions of EPYC processors and will be a lead customer for its next Venice and Verano server CPUs. The same agreement plans shipments for the first gigawatt of custom MI450-based GPUs in the second half of 2026. That gigawatt is a GPU deployment milestone, not a measure of Muse CPU demand.

AMD’s data center revenue reached $6.7 billion in the June quarter, up 107% from a year earlier. The segment combines EPYC CPUs and Instinct GPUs, so the figure cannot isolate either Muse or the CPU opportunity. But it shows that AMD is converting broader AI demand into reported sales. The next test is whether the new systems ship at scale and preserve operating profit.

There is another reason to follow the contract carefully: Meta received a warrant—the right to acquire up to 160 million AMD shares—with portions tied chiefly to GPU purchase and other milestones. As of June 27, according to AMD’s quarterly filing, none had vested. If earned and exercised, those shares would dilute existing holders’ ownership. That cost belongs in the stock assessment.

Meta owns Muse and its distribution. AMD sells part of the infrastructure. Those are different investment cases. Meta reported $60.8 billion in second-quarter revenue, but also $31.1 billion in capital spending, including finance-lease principal payments, and $784 million in free cash flow for that quarter. Muse-specific revenue and costs are not disclosed. Popularity alone cannot establish a return on that spending.

Arm has two routes into this buildout. Meta is co-developing Arm’s AGI data center CPU, while Amazon says its Meta agreement starts with tens of millions of Arm-based Graviton cores for agentic AI. Those cores are a stated deployment plan across Meta workloads, not a disclosed Muse chip count.

Arm sells its own AGI silicon and earns royalties when customers ship processors using its designs. Its data center royalties more than doubled year over year in the June quarter; Arm also reported more than $2 billion of AGI CPU customer demand across fiscal 2027 and 2028. Demand is not booked revenue. AMD has the clearer reported installed EPYC base at Meta today, while Arm has more ways to participate if Meta diversifies.

Agent adoption raises electricity demand if growth in work outpaces efficiency gains. The likely constraint is delivered power at the right site and price, not just the number of processors a supplier can make. A higher data center bill does not automatically become a higher household bill: location, grid investment and utility cost-allocation rules decide who pays. The US Energy Information Administration models larger wholesale-price effects in some high-demand regions; regulators are examining how large loads should bear their costs.