For over a year, the debate about Artificial Intelligence (AI) infrastructure has almost always been centered around the same main player: GPUs. However, the market is beginning to reveal another side of the same storm. Server CPUs— the backbone of “traditional computing” that supports virtualization, networks, storage, databases, and daily operations of any data center— are also becoming a scarce resource. And in China, this pressure is already leading to formal notices of unusually long delivery times.
According to information published this week, Intel has informed some Chinese clients that certain 4th and 5th generation Xeon CPUs could face waits of up to 6 months. AMD, for its part, has reported delays of between 8 and 10 weeks on certain EPYC models. In a market used to planning deployments quarterly, such timelines break budgets, disrupt contracts, and force a rethinking of purchasing strategies with much less flexibility.
Why AI is Also “Eating” CPUs
The quick explanation is that AI doesn’t rely solely on accelerators. A modern AI cluster needs a complete ecosystem: servers for orchestration, storage, , security, virtualization, monitoring, and front-ends that feed pipelines. As companies move from pilot to production, the costs in “general-purpose computing” grow alongside the number of GPUs.
Adding to this is a more recent phenomenon: the advent of agentic AI systems (capable of chaining tasks, operating tools, and executing more complex workflows than a chatbot), which increases demand for server resources that, while not accelerators, remain essential for handling real work in enterprise environments.
A Direct Hit to Prices, Planning, and Negotiation Power
The subsequent effect is almost automatic: when supply tightens, prices cease to behave as a stable variable. In China, it has been reported that some processors with supply issues have seen increases of over 10%, a significant figure in volume purchases where each percentage point counts.
Additionally, Intel is reportedly rationing deliveries of certain models to manage the backlog of pending orders—a measure often used when manufacturers prefer to protect strategic relationships or distribute scarcity rather than face total bottlenecks. AMD, though with shorter timeframes, also shows tension: if the market is demanding server CPUs at a pace beyond expectations, the actual elasticity of the supply chain is demonstrated in weeks… or months.
Manufacturing Under Pressure: Yields, Capacity, and Priorities
Behind these delays are several factors. In Intel’s case, manufacturing limitations and yields (the percentage of valid chips per wafer) have been cited as part of the problem. For AMD, reliance on external foundries adds another layer: global manufacturing and packaging capacity is being reshuffled around AI-related products, with fierce competition for the same industrial resources.
Meanwhile, the market is experiencing rising costs for critical components for any server, especially memory. In fact, the pressure not only affects the “final price,” but also how agreements are signed: in memory, major manufacturers have started introducing shorter contracts and “post-liquidation” formulas, where prices can be adjusted afterward to reflect the market. This is a clear sign of a cycle where suppliers try to avoid being trapped in fixed prices if shortages persist.
The result is an environment where procurement departments lose classic leverage (long commitments, stable volume discounts) and gain variables like SKU flexibility, ability to accept alternatives, and advanced planning.
China, a Sensitive Thermometer for Intel… and a Stress Test for All
The focus on China is no coincidence. For Intel, the country is a market of enormous weight and simultaneously a hub where data center investments, urgency to scale AI, and increasingly monitored supply dynamics converge. The clients on their radar range from large server manufacturers to top cloud providers.
In this context, the key fact is equally relevant: the server CPU market was already in transition. Intel has lost market share in recent years, while AMD has gained ground, and this redistribution is happening just as overall demand is increasing. In other words: it’s not just about “who sells more,” but that total volume is growing, and any capacity deviation becomes an immediate bottleneck.
What Could Happen in 2026: From “Just in Time” to “Just in Case”
The data center industry has been refining supply chains for years to operate with tight inventories. But AI is pushing the sector toward a different logic: buying earlier, securing slots, accepting less convenient contracts, and in some cases, redesigning deployments.
In the short term, the most likely consequence is that infrastructure projects in China—and by extension, other regions—will need to recalculate timelines: not due to a lack of racks or power, but because of the most basic component of all, the CPU. In the medium term, the market may shift toward more dynamic (and more expensive) agreements, similar to what is already observed in memory: shorter contracts, more renegotiations, and increased price variability.
In an industry where “time” equals money—and each week of delay in installed capacity can mean missed opportunities—2026 is shaping up to be the year when AI infrastructure stops being measured solely in GPUs and begins to be understood as the sum of all its critical parts.
Frequently Asked Questions
What does it mean for a server CPU to have a “delivery time” of up to 6 months?
From order confirmation, the manufacturer or channel could take up to half a year to deliver specific units. In data center projects, this requires reserving capacity and planning deployments much further in advance.
Why does AI infrastructure need so many CPUs if GPUs are the key?
Because GPUs accelerate training and inference, but the entire system requires CPUs to orchestrate tasks, move data, virtualize services, operate storage, and perform auxiliary tasks that keep the cluster running.
How does rising memory cost affect server CPU availability?
Memory is a critical component of servers. When prices rise or supplies shrink, many clients accelerate purchases to secure supply, which can lead to increased orders for complete platforms (CPU + memory + motherboard + network) and strain inventories.
How can companies and sysadmins prepare for delays in Xeon or EPYC?
By broadening procurement windows, standardizing alternatives (different tiers or configurations), designing more flexible architectures, and avoiding reliance on a single critical SKU for entire deployments.