Datacenter and HPC

Constraint

Agentic AI Made the Server CPU Scarce Again

Why CPUs became a constraint in 2026

For three years the data center story was GPUs. In 2026 the shortage moved to the CPU. AI agents call tools, search, run code and pass data between models, and that work runs on CPUs. Intel's CFO said in April 2026 that server deployments had gone from about one CPU per eight GPUs to one per four, and could reach one to one. Intel said its unmet server CPU demand "starts with a B", Xeon lead times reached about six months, and Intel and AMD told customers in China to expect long waits (Reuters, February 2026).

Prices followed. Server CPU prices rose 10% to 20% from March 2026, with another 8% to 10% expected in the second half (Commercial Times, via Tom's Hardware). Intel has been moving wafer capacity from PC chips to data center chips, and its Data Center and AI group grew 22% in the first quarter of 2026 to $5.1 billion while PC unit volumes were expected to fall by low double digits.

Supply

Intel Makes the Only US-Built Server CPUs in Volume

Server CPU families and where they are made

Mercury Research put AMD at 34.5% of x86 server CPUs in the second quarter of 2026 and Intel at 65.5%. Arm CPUs designed by AWS, Google, Microsoft and NVIDIA add to that outside the x86 count. Every server CPU not made by Intel is made by TSMC, almost all in Taiwan. Intel launched Xeon 6+ (Clearwater Forest) on its 18A process in 2026, made at Fab 52 in Arizona, which makes the US-built share of server CPUs mostly an Intel question.

• Made in US
• Some US volume
• Made abroad

The x86 and Arm families are covered on CPUs; the Arizona fabs on Leading-Edge Logic Fabs and Intel Foundry.

Bill of materials

A General-Purpose Server Is Built Mostly Abroad

Chip suppliers for a general-purpose cloud server

An AI rack is covered on AI Data Center. The general-purpose servers around it, which run storage, databases, web services and now agent orchestration, carry a different chip list: two CPUs, a dozen or more DDR5 memory modules, SSDs, a network card, a board management controller and power stages.

• Made in US
• Some US volume
• Made abroad

Memory

Server Memory Prices Rise, Except for the Largest Buyers

Server DRAM and NAND prices in 2026

Memory is now the largest single chip cost in many servers. TrendForce expected server DRAM contract prices to rise 13% to 18% in the third quarter of 2026 alone, driven by buyers without long-term contracts: several US cloud companies hold multi-year agreements that stop suppliers from raising their prices. Enterprises, smaller clouds and server makers absorb the increase. HBM for AI accelerators takes the wafers first; the effect on DRAM supply is on DRAM and DRAM Fabs.

SSD prices moved the same way: Gartner forecast NAND revenue up 371.9% in 2026. Enterprise SSD demand for AI data storage keeps NAND fabs, covered on 3D NAND Fabs, at full use.

Network and power

Network and Power Chips Scale With the Rack

Networking and power semiconductors in the data center

Each GPU or CPU added to a data center pulls network and power chips with it: switch ASICs and NICs (Broadcom, NVIDIA, Marvell, all made at TSMC), optical lasers and DSPs for the links between racks, and power stages that step 48 or 800 volts down to under one volt at the processor. Power chips are one of the few data center categories with US fabs, at Texas Instruments, onsemi, Microchip and Infineon's US sites, and GaN and SiC parts are moving into 800 volt rack power.

Facility-side coverage of networking fabrics and rack power distribution is on the sister site DatacentersX, which also traces rack suppliers in its supplier rings and shows how the accelerator sets server power, cooling and building size on its AI Chip Layer page.

China

China Builds Servers on Domestic CPUs and Memory

China's domestic server chips

China's cloud companies buy Intel and AMD server CPUs where export rules allow, and they faced the longest waits in 2026. Domestic alternatives come from Huawei (Kunpeng), Hygon (x86 under an old AMD license) and Phytium, on SMIC and older stockpiled wafers. CXMT and YMTC supply a growing share of China's server DRAM and SSDs. The split is covered on China Bifurcation.

Related coverage

Related Coverage

• CPUs
• DRAM
• NAND
• AI Data Center
• AI and Machine Learning
• Intel Foundry
• AMD
• DatacentersX supplier rings