Home » Blog » Server CPU prices are still climbing in mid-2026, and more hikes are already signaled Server CPU prices are still climbing in mid-2026, and more hikes are already signaled

For two years the data-center story was all about GPUs and the memory feeding them. CPUs were the boring line item: you bought one or two per node to babysit the accelerators and moved on. That assumption broke this spring. On Intel’s Q1 2026 earnings call, Intel’s CFO said the ratio of CPUs to GPUs deployed in data centers has already tightened from roughly 1:8 to 1:4 — and that as workloads keep shifting toward inference and agentic AI, it could converge to 1:1 or tilt further toward CPUs. His summary of the moment, delivered at a Morgan Stanley conference a month earlier: the CPU is “cool again.”

AMD’s quarter told the same story from the other side of the aisle. Data-center revenue hit a record $5.8 billion, up 57% year over year, and the company raised its server CPU market growth forecast from roughly 18% a year to more than 35%, projecting a market above $120 billion by 2030. On the earnings call, AMD guided to server CPU revenue growth of more than 70% year over year for Q2, with the CEO pointing to the “orchestration, data movement and parallel execution” that inference and agentic workloads demand from CPUs. Forecasts that aggressive deserve some skepticism — every vendor is talking its book right now — but the supply-side facts beneath them are checkable, and they’re striking.

The server CPU shortage itself isn’t news anymore — we covered why AI demand set it off when it took hold, and prices have been moving since March. The question worth asking in June is the same one we just worked through for DRAM and NAND: where do prices and supply actually stand, and is there any sign of relief? Short version: prices are still rising with more hikes already signaled, lead times haven’t improved, and institutional investors expect the market to stay constrained into 2027. The CPU market is running the memory playbook a few months behind — demand shock, allocation, price hikes, stretched lead times, and a secondary market that suddenly matters a lot more.

Why agents need CPUs, not just GPUs

We unpacked the mechanics of this shift in our piece on NVIDIA’s Vera launch, so the short version: an agent doesn’t just run a model forward — it plans, calls tools, queries databases, and manages state across long sessions, and nearly all of that glue work runs on CPU cores. That’s what’s behind Intel’s ratio math: at 1:8, a rack of eight GPUs needed one server CPU; at 1:4 it needs two; at 1:1, eight. Whether the ratio actually lands at 1:1 across the industry is an open question — one frames it as a structural rebalancing rather than a fixed endpoint — but the direction isn’t really in dispute. Both vendors are reporting the demand in their actual shipment numbers, not just their forecasts.

And it isn’t only inference. frontier AI labs are now competing directly with cloud providers for commodity x86 servers, because reinforcement-learning training environments — the code compilation, verification, and tool execution that happen alongside the model — need large CPU clusters sitting next to the GPU clusters. A demand source that barely existed two years ago is now bidding against everyone else for the same Xeon and EPYC allocation. (The same dynamics are pulling Arm server silicon into the fight — Graviton, Grace, Axion — which we covered separately in our Arm vs x86 piece.)

What the supply side looks like right now

The shortage is no longer hypothetical. Here’s the state of play, pulled together from the past few months of channel reporting and vendor disclosures:

A few notes on what’s behind those rows:

• The forecasts keep getting overtaken. Omdia’s February estimate of 11–15% for the full year read as aggressive at the time; the market blew past it by April.
• The single-distributor figures check out against vendor statements. The allocation and 30-week numbers come from Fusion Worldwide’s view of the open market, but they square with what Intel itself is saying — Zinsner put the company’s unmet Xeon demand in the billions of dollars.
• Client CPUs are paying the bill. Intel has been shifting wafer capacity away from consumer chips since at least last October to keep server customers supplied, and now expects full-year PC unit volumes to decline by low double digits.
• It stacks on top of the memory crunch. Omdia flagged back in February that CPU constraints were arriving on top of DRAM prices that nearly doubled in a quarter and NAND up more than 30%. A server isn’t one component; when CPUs, memory, and enterprise SSDs all inflate at once, the cost of a configured node moves much faster than any single line item suggests.

One honest caveat: unlike memory, where TrendForce publishes finalized contract surveys, server CPU pricing has no public weekly tracker. It surfaces through channel reporting and vendor disclosures, and the freshest hard data points as of this writing are late April and early May (AMD’s May 5 results). The picture is consistent across multiple independent sources, but the numbers are softer than DRAM contract data, and we’d treat any single percentage as directional. The next real reads arrive with the H2 price-adjustment announcements and July’s earnings calls.

If you build your own machines, this is your problem too

This reads like an enterprise story, but the wafer math reaches the desktop. Every wafer Intel moves to Xeon is a wafer that doesn’t become a Core i5, which is why consumer chips are running above MSRP with more increases expected — you’re competing with hyperscalers for fab capacity whether you know it or not. The consolation prize flows the other way: as enterprises refresh early to lock in supply, their retired hardware lands on the secondary market, and a pulled Xeon or EPYC with a couple dozen cores often costs less than a current mid-range desktop chip. Homelabbers and power users have quietly built on decommissioned enterprise gear for years; a shortage that pushes new parts out of reach makes that channel more attractive, not less.

What this does to the used CPU market

Here’s the part that matters if you’re holding hardware rather than buying chips.

Every previous shortage cycle has taught the same lesson: when new parts go on allocation, demand doesn’t disappear — it moves. Buyers who can’t get new Xeon or EPYC inside their project timeline start looking at the previous generation, and at pulled processors from decommissioned systems. Refurbished and renewed server gear typically runs well below new pricing, and hosting providers are openly advising customers to lock in hardware now and consider alternatives rather than wait out 2026’s increases. A buyer facing a 30-week lead time on a new EPYC has a strong incentive to pay good money for a tested, working part available this week.

That flows directly into resale values. Secondary-market pricing tracks the new-part market with a lag — we saw it with DDR4 and DDR5 over the past year, and the same dynamic is now setting up for server CPUs. High-core-count parts from recent generations are the obvious beneficiaries, but the effect reaches further down the stack than people expect: when the front of the line is sold out through year-end, even two- and three-generation-old silicon gets a second look from buyers who just need working cores.

Two practical implications, depending on which side of the trade you’re on:

• If you’re decommissioning or refreshing: processors pulled from retired servers are worth meaningfully more in a shortage than in a glut, and shortage windows close. If you have Xeon or EPYC inventory sitting in a drawer waiting for someone to deal with it, this is a better-than-usual time to sell used CPUs rather than letting it depreciate through the cycle.
• If you’re trying to buy: budget for the H2 price increases analysts are flagging, get orders in early, and don’t dismiss previous-generation parts. The performance gap between generations is real, but so is the gap between hardware that arrives this month and hardware that arrives in Q1 2027.

The thing to watch from here is whether the CPU:GPU ratio keeps tightening the way Intel projects, or stabilizes around 1:4 as the first wave of agentic deployments matures. Vendor earnings calls in July and the H2 price adjustments — Intel’s expected 8–10%, AMD’s rumored Q3 round — will tell us which way it’s breaking. Either way, the era of the server CPU as an afterthought line item is over for a while — and so is the era of retired processors being worth pocket change.

---

This is the third piece in our coverage of the AI-driven CPU squeeze: why AI set off the global CPU shortage explains how we got here, and our piece on NVIDIA’s Vera launch covers what agent workloads actually demand from CPUs. We’ll update this one when the H2 price adjustments land.

Market figures here reflect reporting as of early June 2026 and will change — this market is moving quickly. For current trade-in values on specific processors, contact us directly rather than relying on any snapshot.