Server CPUs — the processors that were all but written off two years ago — are now the component most likely to delay your organization's AI deployment, and three earnings reports released since late March confirm the shortage is structural, not cyclical. On April 24, Intel's data-center and AI revenue jumped 22 percent year-on-year to $5.1 billion in the first quarter of 2026, sending its shares up 23.6 percent in a single session — the chipmaker's largest one-day gain since October 1987. A week later, AMD reported first-quarter revenue of $10.3 billion, with its data-center segment — driven by EPYC server processors and Instinct GPUs — up 57 percent to $5.8 billion. Between those two reports, Arm abandoned 35 years of neutrality to sell its own silicon for the first time. The trigger for all three moves is identical: a new generation of AI software is consuming far more CPU capacity than anyone forecast, and supply cannot keep up.
The most immediate evidence arrived this week. Server CPU lead times have stretched from the historical norm of one to two weeks to between eight and twelve weeks for AMD parts, and up to six months for some Intel Xeon configurations, according to reports from Nikkei Asia and industry supply-chain sources. Server CPU prices have risen as much as 20 percent since March 2026. Intel has already pushed through two rounds of increases this year and is reallocating wafer production from PC chips to Xeon server processors; AMD is reportedly planning two further rounds of increases in the second and third quarters. For any organization that had planned a server infrastructure build in 2026, the era of ordering a CPU and receiving it within a fortnight is over.
Agentic AI Is Changing How Many CPUs a Data Centre Needs
Two structural shifts are driving the demand surge. The first is the rise of agentic AI — software that plans tasks, executes code, calls external tools, and validates results, often running many parallel processes in sequence. That orchestration work runs on CPUs, not GPUs. Intel CEO Lip-Bu Tan told investors on the April 24 earnings call that the CPU-to-GPU ratio in data centers has already narrowed from roughly one CPU for every eight GPUs to one CPU for every four, and that the trajectory is toward parity — one CPU per GPU — as agentic workloads scale. He called CPUs "an indispensable foundation of the AI era."
AMD CEO Lisa Su made the same argument on her May 5 earnings call: "Demand is increasing not only for accelerators, but also for the high-performance CPUs that power and orchestrate those workloads." She added that AMD has now delivered four consecutive quarters of record server CPU revenue, with cloud and enterprise customers each growing more than 50 percent year-on-year.
The second structural pressure is memory. As AI models stretch toward contexts of one million tokens or more, their key-value caches overflow GPU memory and spill into system RAM, where the CPU must manage them. Both forces push the same direction: more CPU per rack, more memory bandwidth per core, less tolerance for latency on the CPU side of the rack.
Arm Enters the Race With Its First Chip in 35 Years
The most dramatic signal of how much the CPU market has changed came on March 24, when Arm Holdings announced the Arm AGI CPU at its "Arm Everywhere" event in San Francisco — the company's first piece of production silicon in its 35-year history. Arm has always sold blueprints, not chips. That model ended with a 136-core data-center processor built on TSMC's 3nm process using Neoverse V3 cores, co-developed with Meta as the lead customer. Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom signed on as launch partners; Arm CEO Rene Haas said customer demand for the chip has already exceeded $20 billion — double the company's initial supply target — within six weeks of launch, constrained only by wafer availability.
The AGI CPU deliberately omits simultaneous multithreading and dynamic clock scaling in favor of deterministic, predictable performance at rack scale — a design trade-off that Arm's Mohamed Awad, EVP of Cloud AI, described as intentional: the team "specifically didn't want to add things that weren't going to be 100 percent utilized in the mission of this device." Arm claims a liquid-cooled 200kW rack running AGI CPUs holds more than 45,000 cores — more than twice the density of Nvidia's competing Vera CPU rack. Volume shipments are expected from the fourth quarter of 2026.
The launch has not been without controversy. By selling finished chips, Arm now competes directly with the companies that pay it licensing fees. Nvidia, which had previously been a potential acquirer of Arm, reportedly divested its remaining equity stake as of February 2026. Qualcomm has accelerated a pivot toward open-source RISC-V architecture partly to reduce dependence on Arm's IP. Haas's response: he projected that by 2030, Arm will hold the largest CPU market share by type. How much of the installed x86 enterprise software base his chip can displace is the central commercial question for the AGI CPU's first year.
Nvidia Launches Vera, Intel Spends $14.2 Billion on Its Own Factory
Nvidia moved first. On March 16, at its GTC conference in San Jose, the company launched the standalone Vera CPU, built on 88 custom Olympus cores and promising twice the energy efficiency and 50 percent higher performance than traditional rack-scale CPUs for agentic workloads. Jensen Huang described the CPU's new role bluntly: "The CPU is no longer simply supporting the model; it's driving it." Alibaba Cloud, ByteDance, Meta, Oracle Cloud Infrastructure, CoreWeave, Lambda, Nebius, and Nscale signed on as launch customers. Vera is in full production and will reach partners in the second half of 2026.
Intel responded with capital rather than silicon. On April 1, the company announced a $14.2 billion agreement to repurchase the 49 percent stake in its Fab 34 manufacturing plant in Leixlip, Ireland from Apollo Global Management — the same stake Apollo had paid $11.2 billion for in 2024. The transaction, funded with cash and $6.5 billion in new debt, gives Intel complete ownership of one of its most advanced European facilities at a $3 billion premium. CFO David Zinsner described the move as a return to balance-sheet strength — and a direct bet that CPU demand will justify full ownership of every wafer Fab 34 can produce.
Intel's manufacturing picture is improving faster than the company had previously guided. At the Q1 earnings call, Zinsner confirmed that 18A process yields are now tracking to hit year-end targets by mid-2026 — roughly two quarters ahead of the prior schedule. Yield improvement has been running at about 7 percent per month, which Intel VP John Pitzer described as the "industry average" rate for a new process node ramp. The next generation, Intel 14A, is already in customer evaluation and outpacing 18A at a comparable stage.
Two Chokepoints: Foundry Capacity and a Japanese MSG Maker
The supply crunch has two distinct origins, and neither resolves quickly. The first is advanced foundry capacity. AMD's EPYC server CPUs are manufactured by TSMC, which is prioritising higher-margin AI accelerators on its 3nm node and compressing CPU wafer allocations as a result. TSMC chairman C.C. Wei said publicly at a November 2025 industry awards ceremony that advanced-node capacity is "still not enough, not enough, not enough" and estimated it falls roughly three times short of what major customers plan to consume. TSMC's new Arizona and Kumamoto fabs are not expected to deliver high-volume advanced-node production before 2027 and beyond.
The second chokepoint is packaging. High-end CPUs and GPUs share the same advanced substrate stack, the insulating layers that connect silicon to circuit board. The critical material inside those substrates is ABF — Ajinomoto Build-up Film — a specialty insulating layer on which Japanese food and amino-acid company Ajinomoto holds more than 95 percent of the global market. ABF in turn depends on a specialty low-thermal-expansion glass cloth called T-glass, where Nittobo, a Tokyo-based fiberglass manufacturer, controls roughly 90 percent of global supply. Nittobo announced in August 2025 that it would invest approximately ¥15 billion (around $96 million) to build a new factory at its Fukushima production base and triple T-glass capacity, with the facility scheduled to begin production in the fourth quarter of 2026. The company acknowledged that even this expansion "may still fall short of meeting all market demands."
The packaging bottleneck is expected to widen rather than narrow over the next two years. Morgan Stanley has forecast a significant ABF substrate shortage in 2027, with AI chip demand growing at roughly 16 percent compounded annually through that period. Ajinomoto plans to invest at least ¥25 billion by 2030 to expand ABF output by 50 percent — a target that assumes demand growth will eventually stabilise at rates the company can plan around.
Server Buyers Are Paying More and Waiting Longer
For the organizations that actually need to procure servers — cloud operators, enterprises building internal AI infrastructure, AI-native startups — the supply crunch has direct, measurable consequences. Server CPU prices have risen by as much as 20 percent since March 2026. AMD is reportedly planning two further price increases through the rest of 2026, for a cumulative rise of 16 to 17 percent. Intel raised prices in both March and April and may raise them again by 8 to 10 percent in the second half of the year. Lead times have extended from the historical norm of one to two weeks to eight to twelve weeks for AMD parts, and up to six months for some Intel Xeon configurations.
The constraint is not simply one of money. One executive at a gaming PC brand told Nikkei Asia: "If money can solve the problem, that would be great. What we worry about is that even if we pay more, we still cannot get more." Hyperscalers absorb the impact differently from enterprise buyers: Amazon, Google, Microsoft, and Meta can pre-commit to multi-year TSMC wafer allocations and have done so. For companies without that leverage, the market resembles the GPU crunch of 2023 — except that CPUs underpin far more of the existing enterprise software stack.
The critical distinction from the 2023 GPU shortage is that CPUs anchor decades of enterprise software certification. Oracle database licensing, SAP HANA, and most Windows Server workloads are validated on x86 silicon from Intel and AMD. Arm-based alternatives — however compelling on performance-per-watt metrics — require application recompilation and, in many cases, significant refactoring. That dependency buys Intel and AMD time even as Arm and Nvidia compete for new deployments.
A Broader Semiconductor Shortage, Not an Isolated CPU Problem
The CPU crunch is the most visible symptom of a wider supply-demand dislocation across advanced semiconductors. Oxford Economics economist Makoto Tsuchiya warned in January 2026 that "the current supply–demand imbalance" across the broader AI-driven semiconductor sector is expected to "persist over the next few years, as long as the AI investment boom continues." He noted that within electronics, "the polarization between AI and non-AI players will likely intensify," with chipmakers benefiting from rising prices while downstream manufacturers absorb mounting input costs.
The same dynamic is visible in Arm's earnings. On May 6, Arm reported record fiscal fourth-quarter revenue of $1.49 billion, up 20 percent year-on-year, and CEO Rene Haas said on the earnings call that customer demand for the AGI CPU has already exceeded $20 billion — constrained not by customer appetite but by how many wafers TSMC can allocate. Haas predicted that by 2030, Arm will hold the largest CPU market share of any architecture type, measured by core count rather than chip count.
What This Means for Anyone Buying or Building AI Infrastructure
For organizations planning server infrastructure in the next 12 to 18 months, the practical consequences are specific. Procurement lead times that once took two weeks now take two to six months for advanced server CPUs. Prices will continue rising through at least 2027 as TSMC's new fabs ramp and Intel's 18A yield curve matures. Organisations with flexibility to defer purchases until TSMC's new Arizona and Kumamoto fabs reach volume production may face a more competitive supply environment — but not before late 2027 at the earliest.
The architectural choice — x86 from Intel or AMD, Arm-based from Arm or Nvidia — now carries concrete trade-offs that were theoretical one year ago. Intel's Xeon and AMD's EPYC run existing enterprise software without modification, but face the same supply constraints and rising prices. Nvidia's Vera and Arm's AGI CPU offer compelling performance-per-watt arguments for greenfield AI deployments but require software migration work for established workloads. Neither path is free.
The broader implication is structural: a component that supply-chain analysts had largely written off as mature and commoditized in 2024 is now the item most likely to set the pace — and the cost — of the next phase of AI infrastructure build-out. Intel's 23.6 percent single-day stock move on April 24 was the market's most visible acknowledgement of that reversal. The question for buyers is whether they moved early enough to be on the right side of it.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
