Agentic AI just flipped the shortage board. Server CPUs are now the bottleneck — GPUs are sitting at balanced.
Meta's Muse is the tell. Persistent VMs and orchestration live on CPUs, not accelerators. Server CPU lead times stretched to 25–30 weeks against a 16–20 week balanced baseline. Cloud providers are accelerating internal chip designs. $AMD $INTC $ARM all moved on the data.
The rest of the stack is already strained:
DRAM: Very Tight, 20 weeks vs 8 week baseline. Q4 2026 and Q1 2027 pricing talks underway as HBM and server capacity keep pulling wafer allocation.
NAND eSSD: Tight, 16 weeks.
HDD: Very Tight, 50 weeks.
ABF substrate: Very Tight, 48–56 weeks.
MLCC: Tight, 32 weeks.
GPUs: Balanced.
This is the reallocation in real time. Agents drive CPU hours, storage, and substrates first. Discrete GPUs are not the scarce part this cycle.
Micron's fiscal Q4 print Wednesday after close is the next read on whether memory still has momentum after last quarter's surge.
Meta's Muse is the tell. Persistent VMs and orchestration live on CPUs, not accelerators. Server CPU lead times stretched to 25–30 weeks against a 16–20 week balanced baseline. Cloud providers are accelerating internal chip designs. $AMD $INTC $ARM all moved on the data.
The rest of the stack is already strained:
DRAM: Very Tight, 20 weeks vs 8 week baseline. Q4 2026 and Q1 2027 pricing talks underway as HBM and server capacity keep pulling wafer allocation.
NAND eSSD: Tight, 16 weeks.
HDD: Very Tight, 50 weeks.
ABF substrate: Very Tight, 48–56 weeks.
MLCC: Tight, 32 weeks.
GPUs: Balanced.
This is the reallocation in real time. Agents drive CPU hours, storage, and substrates first. Discrete GPUs are not the scarce part this cycle.
Micron's fiscal Q4 print Wednesday after close is the next read on whether memory still has momentum after last quarter's surge.
