TrendForce's latest report indicates that CPU lead times have extended from the usual 16–20 weeks to 25–30 weeks, with this shift directly attributed to surging demand for CPUs driven by AI accelerators such as Muse. This marks the first time the market has obtained concrete evidence at the supply-chain level of AI accelerators exerting pressure on CPU demand. Previously, discussions around Muse largely focused on user‑base growth and conceptual speculation; now, the pronounced lengthening of lead times suggests that the squeeze on hardware infrastructure caused by AI accelerators has spread from GPUs to CPUs.
$Meta Platforms (META.US)$The AI agent Muse is shifting CPUs from a supporting role in AI‑driven narratives to center stage—according to TrendForce's latest report, CPU lead times have extended from the usual 16–20 weeks to 25–30 weeks, with this shift directly attributed to the surge in CPU demand driven by AI agents like Muse.
This marks the first time the market has obtained direct evidence that AI agents are exerting pressure on CPUs at the supply-chain level. Previously, discussions about Muse largely focused on user growth and theoretical projections; now, the pronounced lengthening of delivery cycles indicates that the squeeze AI agents are placing on hardware infrastructure has spread from GPUs to CPUs.
However, a significant gap remains between current demand and expectations. According to Freda Duan, a partner at Altimeter, the figure of 12.5 million core instances is closer to the theoretical upper limit; Meta may not necessarily deliver a dedicated virtual machine for every user—there is still a validation gap between bullish expectations and actual order fulfillment.
Lead times have lengthened, and the bottleneck stages remain unresolved.
TrendForce's report identifies a clear uptick: CPU lead times have extended from 16–20 weeks to 25–30 weeks, with AI accelerators such as the MUSE chip cited as a key driver of this surge in demand. Over the past two years, discussions around AI infrastructure have revolved almost exclusively around GPUs; now, for the first time, the pressure exerted by AI accelerators on hardware is beginning to ripple through to CPUs.
However, the report does not specify at which stage delivery delays are occurring. The CPU supply chain comprises three major segments—wafer fabrication, packaging, and testing—and bottlenecks in each segment correspond to distinct supply-and-demand dynamics and varying resolution timelines. At present, CPU manufacturers, server OEMs, and cloud service providers have all refrained from publicly identifying the root cause; this information gap itself constitutes a significant market variable.
12.5 million cores: a theoretical ceiling, not an actual order quantity.
Altimeter partner Freda Duan's calculations provide a quantifiable benchmark for market discussion. She assumes Muse achieves 100 million daily active users, with each user averaging two hours of usage, which would require roughly 8.33 million virtual machines to handle the average load. At peak times—when demand is 2.5 times the average—the maximum number of virtual machines would rise to 20.8 million. Factoring in the physical CPU cores required per virtual machine—using DeepSeek as a reference, though Meta's operational efficiency falls short of DeepSeek's—she concludes that, under the assumption of 25 million active virtual machines, Meta might need approximately 12.5 million CPU cores.
Duan herself adopts a cautious stance toward this figure. She points out that AI‑agent sandboxes and virtual machines typically suffer from resource overcommitment, and that when a VM remains running, its actual CPU utilization often hovers near zero. Consequently, Meta does not need to strictly deliver the promised hardware for every user—12.5 million cores represent more of a theoretical ceiling than an actual procurement requirement. She also notes that as inference speeds improve, the number of physical cores allocated to each VM will shrink further; however, gains in technical efficiency are often accompanied by expanding demand, so overall CPU demand may not necessarily decelerate as a result.
An unexpected variable for Intel and AMD
Meta has pledged to provide each Muse user with a dedicated virtual machine equipped with two virtual CPUs, 8 GB of memory, and a 100 GB solid-state drive. Within ten days of its launch, Muse's user base reached 700,000.
Preliminary calculations suggest that as Muse's daily active user base expands, it could drive demand for millions of additional CPUs. Market participants therefore view this as a structural tailwind for CPU manufacturers; however, this expectation has yet to be reflected in concrete revenue guidance or order‑book data.
Delivery cycle data is the first hard signal. Going forward, it will be important to watch Intel and AMD's commentary on their data center CPU businesses in their next-quarter earnings reports, as well as changes in the CPU-to-GPU spending mix among cloud providers.
Whether the lengthening delivery cycles can translate into revenue recognition on the financial statements is the key test for this round of CPU‑related narratives.
Editor/Deng