Intel and AMD are negotiating CPU supply contracts for servers of one or multiple years with major Chinese clients, at a time when AI-related demand exceeds available capacity. Some processors are already costing over 40% more than at the beginning of 2026 in China, while buyers are willing to commit to volumes without knowing the final price upfront.
The key points of server CPU shortages in 30 seconds
- Intel and AMD seek one- or two-year purchase commitments with Chinese clients.
- These contracts secure volumes but generally do not lock in the processor prices.
- Some server CPUs have increased more than 40% since January.
- Certain Xeon models face waits of up to six months, and some EPYC models are delayed by eight to ten weeks.
- AI data center demand is also stressing memory, networking, and storage components.
This situation confirms that AI pressure is no longer solely focused on NVIDIA GPUs, AMD GPUs, or accelerators developed by major cloud providers. Each new cluster also requires central processors, memory, storage, high-speed networking, and complete power and cooling systems.
CPUs continue to perform an essential part of the workload. They run the operating system, prepare and distribute data, manage storage, coordinate accelerators, and support numerous services that do not benefit from GPUs. The growth of inference and AI agents adds workloads that heavily utilize conventional processors.
Supply contracts that guarantee units but not prices
Intel and AMD are leveraging their bargaining power to ask certain Chinese customers for longer-term commitments. According to Reuters, these agreements typically last about a year, although both companies have discussed contracts of two years or more.
The main feature is that these contracts commit to a volume of CPUs the customer will buy, but they usually do not lock in the price. The provider reserves a portion of production, giving the buyer some predictability of deliveries, yet they remain exposed to future price increases.
This is a significant difference from traditional supply agreements with fixed quantities and costs. A company can guarantee it will receive thousands of CPUs over upcoming quarters, but it will pay according to the market conditions and its relationship with the manufacturer at the time of delivery.
This model benefits Intel and AMD as long as demand exceeds supply. Some processors have increased by over 10% month-to-month in China, and certain models have gained more than 40% since January 2026.
The percentages do not mean that the entire Xeon and EPYC ranges have increased equally. Actual prices depend on the model, order volume, delivery date, distributor, and negotiated discounts. Manufacturers often publish reference tariffs, but major cloud providers and data center operators typically work with their own contracts.
Waiting six months foreshadowed the problem
Early signs of shortages appeared in early February. Intel informed several Chinese clients that certain 4th and 5th generation Xeon Scalable processors could have lead times of up to six months. Similarly, AMD’s lead times for some EPYC CPUs ranged between eight and ten weeks.
At that time, Intel’s server processors had already risen by over 10% on average for some buyers in China. Five months later, the pressure persists, and the market has shifted from managing isolated delays to negotiating supply a year or two in advance.
Intel has acknowledged that demand for Xeon remains above its delivery capacity. The company also increased its capital expenditure forecast for 2026 from $18 billion to $20 billion after reporting quarterly results driven by data center CPU growth.
AMD relies on TSMC to manufacture its EPYC processors. TSMC must also meet demand for AI accelerators, networking chips, consumer processors, and other products produced on its advanced nodes. The limit is not necessarily just in one part of the chip; encapsulation, substrates, memory, and assembly capacity can all influence the number of finished systems.
The shortage resembles what was observed with DRAM memory and especially high-bandwidth memory (HBM). Major buyers have entered multi-year agreements to secure capacity, though this strategy reduces the availability for smaller clients and makes price drops less rapid.
China faces an additional challenge in expanding its AI infrastructure
Price increases predominantly impact Chinese cloud service providers, internet companies, and operators building AI data centers. These organizations need U.S.-made CPUs even when using their own accelerators or chips developed locally.
Furthermore, China’s access to the most advanced U.S. GPUs remains restricted due to U.S. export controls. The reduced availability of CPUs adds another hurdle to cluster deployment, although these CPUs are not necessarily subject to the same restrictions as high-performance accelerators.
Buyers can turn to CPUs designed in China, ARM-based architectures, or processors developed for specific applications. However, switching platform is not immediate, as data centers have accumulated software, management tools, operating systems, and applications optimized over years for Intel and AMD x86 processors.
A 40% increase in CPU price does not mean the total cost of a server rises proportionally. The CPU is just one component; its impact depends on the configuration. In machines with large memory, NVMe storage, or multiple GPUs, other parts may constitute a more substantial share of the investment.
The impact will be noticeable in projects with thousands of nodes. A difference of several hundred or thousand euros per processor can alter the overall budget, delay expansion plans, or lead to longer use of older equipment.
The consumer market may not face the same shortages
The bottleneck in data centers has sparked debate about whether Intel and AMD could cut production of PC CPUs to focus on more profitable server products. Server CPUs are typically sold at much higher prices and via volume contracts, but it’s not necessarily true that each unit meant for a data center directly replaces a consumer processor.
Intel and AMD use different designs, validation processes, packaging, and product supply chains. In AMD’s case, several processors share blocks manufactured by TSMC, which can create competition for certain wafers or encapsulation capacities. Intel manufactures most of its CPUs in-house but also allocates resources across different families and nodes.
This pressure on shared components can spill over into other markets. Substrates, memory, storage, and advanced manufacturing capacity are used in numerous devices. When data centers are willing to pay more and sign extended contracts, suppliers are incentivized to prioritize those orders.
However, there is no immediate proven direct link between Chinese agreements and a future generalized shortage of consumer CPUs. The most immediate effect is on server CPUs, lead times, and the cost of expanding enterprise infrastructure.
AI has thus transformed a market that had previously experienced relatively predictable supply. GPUs were the first to become scarce, followed by HBM pressure, and now CPUs are also ordered months or years in advance. For data center operators, securing electrical power is no longer enough: they must also reserve every component that converts energy into compute capacity.
Frequently Asked Questions
How much have server CPUs increased in China?
Some models have risen more than 40% since early 2026. The increase varies by CPU, product, customer, and contract.
How long do Intel and AMD take to deliver processors?
Intel has reported waits up to six months for certain 4th and 5th generation Xeon processors. AMD’s lead times for some EPYC CPUs range between eight and ten weeks.
Do long-term contracts lock in prices?
Generally not. These agreements primarily specify the volume the customer will buy, with prices determined later based on market conditions.
Will the shortages affect consumer processors?
There is no evidence of widespread shortages directly caused by these contracts. The pressure might impact manufacturing capacities and shared components, but product lines are not fully interchangeable.
