Reuters reported on July 23 that Intel and AMD are seeking longer purchase commitments from Chinese server customers as some CPU prices in China have risen more than 40% this year. That makes the US chip-control story messier than Washington likes to admit.
There's a quiet contradiction inside American chip policy. Washington has spent years trying to slow China's access to the most advanced AI accelerators, especially Nvidia-style GPUs, but the servers those chips sit inside still need central processors. Now Reuters says Intel and AMD are using that demand to push Chinese server customers toward longer commitments for Xeon and EPYC data-centre CPUs.
This isn't a small parts-bin story. It is the plumbing of AI infrastructure.
According to Reuters, the agreements under discussion generally lock in purchase volumes, not prices. Most cover about a year of supply, though Intel and AMD have discussed two-year or longer commitments with some customers. That difference matters. If you're a Chinese cloud provider, you may get a queue position. You don't necessarily get price protection.
The price pressure is already visible. Reuters reported that server CPU prices in China are still climbing, with month-on-month increases topping 10% for some products and some CPU products up more than 40% since the start of 2026. Reuters also reported in February that Intel had warned of lead times of up to six months for some server CPUs, while AMD had told customers that some products could take eight to ten weeks.
That is a very different market from the one server buyers knew a few years ago. CPUs used to be the manageable part of the build. You worried about Nvidia accelerators, HBM memory, networking gear, power, cooling and racks. The Xeon or EPYC chip was not usually the thing that kept the project team awake. Now it can be.
AI did that. Reuters described the demand as spreading beyond accelerators into memory, networking gear and server processors, because AI data centres need CPUs for servers, storage, networking and inference workloads. AMD has also raised its server CPU market forecast to more than $120 billion by 2030, citing demand tied to agentic AI workloads. You don't have to buy the whole agentic AI narrative to see the practical result. That much is simple. More automated work means more servers running longer and doing more coordination around the accelerator.
The CPU shortage puts Intel and AMD in the driving seat #
Intel and AMD have not had this kind of pricing power in the server CPU market for a long time. Reuters said Intel CEO Lip-Bu Tan told analysts in April that demand "continues to run ahead of supply," especially for Xeon server CPUs, and cited a multi-year Google deal as one of several long-term contracts Intel signed in the first quarter. That is the part investors care about. Scarcity lets a supplier ask for commitment before it gives you allocation.
CNBC reported in January that KeyBanc upgraded both Intel and AMD to overweight on strong server demand, with a $60 price target for Intel. KeyBanc analyst John Vinh said Intel was largely sold out of server CPUs in 2026 and was considering average selling price increases of 10% to 15%. For AMD, the same setup is cleaner in one respect: EPYC already has the performance story and the hyperscaler credibility. Supply tightness makes that story easier to price.
But don't confuse pricing power with control. Intel did not respond to Reuters' request for comment on the July 23 report, and AMD did not either. In February, Intel told Reuters that rapid AI adoption had lifted demand for traditional compute and that supply should improve through 2026. AMD told Reuters then that it remained confident in its ability to meet global demand through supplier agreements and its TSMC partnership. Those are careful corporate answers. Neither company manufactures infinite chips on command.
Washington left a policy gap #
Frankly, the export-control picture is untidy. The Bureau of Industry and Security revised its license review policy in January 2026 so certain advanced AI chips, including Nvidia's H200 and AMD's MI325X, could be reviewed case by case for China under conditions such as third-party testing and customer-screening requirements. That rule dealt with advanced computing chips. It did not turn ordinary server CPUs into the same kind of controlled frontier AI accelerator.
On paper, that distinction makes sense. A Xeon or EPYC processor is a general-purpose chip. It runs databases, enterprise software, cloud instances and a lot of dull but necessary work. Yet the Reuters report shows why the distinction is harder to defend in practice. Chinese AI clusters restricted from buying the best GPUs still need CPUs to move data, manage servers and run inference-related tasks around the accelerator layer.
That is the awkward part. US policy can block one component while demand floods into another.
China is one of the world's largest server markets, and Reuters pointed to rapid construction of data-centre racks, AI computing clusters and national computing infrastructure. Chinese buyers also include major cloud and internet companies expanding AI services. If those buyers sign longer purchase commitments with Intel and AMD, they get more certainty in a market where waiting six months can wreck a deployment schedule. Intel and AMD get volume visibility and pricing power.
The risk is that policy catches up after the contracts are signed. Because the reported deals lock in purchase volumes rather than prices, Chinese customers may still face higher costs, and Intel and AMD may still face disruption if Washington decides server CPUs have become too important to leave alone. For now, Reuters has exposed the simple truth inside the shortage: restricting GPUs doesn't remove China's AI demand. It just moves the pressure to the parts still available.
Also read: OpenAI's own AI models broke out of a test sandbox and autonomously hacked Hugging Face • Hyundai's 35,000 striking workers just forced the first real test of who controls humanoid robots on the factory floor • Substack's new AI scanner tells paying readers exactly how much of their newsletter a human actually wrote