The AI hardware crunch: CPUs join the chip shortage

A version of this article originally appeared in Quartz’s AI & Tech newsletter. Sign up here to get the latest AI & tech news, analysis and insights straight to your inbox.

For three years, the defining hardware story of artificial intelligence has been a GPU shortage. Companies hoarded Nvidia $NVDA +1.16%'s graphics processors like gold bars, willing to pay almost anything for the chips that power AI training. Entire business strategies revolved around who could get access and who couldn’t.

Now a second shortage is emerging alongside it, and this one has caught the industry genuinely off guard. The humble CPU, the generalist processor that has powered computers for decades, is suddenly in desperately short supply.

Intel $INTC +2.63% warned Chinese customers in recent weeks that delivery lead times for some server CPUs have stretched to six months. AMD +0.27% has pushed its own lead times to eight to ten weeks. Server CPU prices in China have jumped more than 10 percent. In the US and Europe, PC prices are creeping upward too, as chipmakers divert manufacturing capacity from consumer products toward data centers hungry for more processors.

Intel’s CFO David Zinsner admitted during the company’s January earnings call that demand had blindsided them. Six months ago, he said, every major cloud customer was signaling they would need more powerful CPUs but not necessarily more of them. That forecast turned out to be wrong. Unit demand surged through the second half of 2025, and Intel now finds itself running its fabs “hand to mouth,” shipping processors as fast as they come off the line.

The agent problem

The explanation sits at the intersection of two trends that converged faster than anyone in the semiconductor supply chain anticipated.

The first is straightforward. Microsoft $MSFT -0.89% ended support for Windows 10 last October, triggering a wave of PC upgrades. Many of those buyers opted for cheaper machines running older Intel chips rather than the pricier AI-enabled PCs that Intel and Microsoft had been pushing. That created unexpected demand for processors Intel had been winding down.

The second trend is more structurally interesting. The AI industry is shifting from building chatbots to deploying autonomous software agents, and this shift is fundamentally changing the ratio of hardware that data centers need.

When you ask ChatGPT a question, the CPU does very little. It converts your text into tokens, hands them to the GPU for processing, and converts the answer back. The GPU does perhaps 90 percent of the work.

But agentic AI systems behave differently. They plan, execute multi-step tasks, call APIs, query databases, write and run code, coordinate dozens of sub-processes, and evaluate whether they succeeded before starting over. All of that work happens on CPUs.

AMD CEO Lisa Su put a finer point on it during her earnings call. The more autonomous AI agents become, she said, the more they depend on the oldest, least glamorous chip in the server rack.

Translation: the server CPU market is about to have a very good year. Su predicted strong double-digit growth in 2026.

Squeezing the wrong chips

Intel and AMD are short for entirely different reasons, which makes the problem harder to solve.

Intel has been struggling with manufacturing yields at its own fabrication plants, limiting how many usable chips it can produce from each silicon wafer. The company is investing in new tools and reallocating capacity from PC chips to server chips, but the improvements won’t arrive until later this year at the earliest.

AMD doesn’t make its own chips. It relies on TSMC $TSM -0.46% in Taiwan, the world’s most advanced contract manufacturer. But TSMC is prioritizing its most advanced production lines for higher-margin AI accelerators and GPUs, leaving less room for CPU orders. TSMC’s chairman has acknowledged that the company can only produce about a third of what its biggest customers want.

Meanwhile, a global memory chip shortage is making everything worse. When memory prices started climbing late last year, customers rushed to lock in CPU purchases too, hoping to assemble complete server systems before costs spiraled further. That panic buying deepened the CPU backlog.

Nvidia, sensing opportunity, is pushing aggressively into the market. Its server CPUs are already deployed at scale inside Meta $META +1.03%'s data centers for workloads that don’t require a GPU at all, and a next-generation chip designed for agentic reasoning arrives next year. Jensen Huang said in January he sees Nvidia becoming a major CPU producer.

The irony is that the technology most likely to be affected by the CPU shortage is AI itself. Companies racing to deploy agents may find that the bottleneck isn’t the expensive GPU they fought so hard to secure but the cheap, unglamorous processor they assumed would always be available.

📬 Sign up for the Daily Brief

Our free, fast and fun briefing on the global economy, delivered every weekday morning.

Sign me up