There was only one character in the AI hardware story for the majority of the previous three years. GPUs from Nvidia. The chips were leased, traded, hoarded, and even allegedly smuggled across international borders. The ability of their founders to get a call back from someone at a cloud provider was crucial to the survival of entire startups. The story is still ongoing. However, it is no longer the sole one.
A second shortage has surfaced somewhere quieter, and this one has the industry looking a little ashamed. Suddenly, the CPU—the workhorse processor that has been humming inside servers for decades—is in short supply. A few weeks ago, Intel informed Chinese clients that some server chips now have lead times of up to six months. Wait times at AMD have risen to eight or ten weeks. In China, the cost of server CPUs has increased by over 10%. Because chipmakers are covertly shifting capacity from consumer goods to data centers that are short on silicon, PC prices in the US and Europe are also gradually rising.
The startling thing is that Intel didn’t anticipate this. CFO David Zinsner essentially acknowledged this during the company’s January earnings call, stating that six months prior, cloud customers had informed him that they would prefer more powerful CPUs rather than just more of them. He proved to be mistaken. According to him, the fabs are currently struggling to make ends meet.
Here, two stories are intertwined. The first is unremarkable. Last October, Microsoft discontinued support for Windows 10, which caused a surge in purchases at computer stores. Many opted for less expensive laptops with older chips that Intel was already retiring instead of the more expensive AI-enabled devices. That alone led to an unanticipated surge in demand for processors that Intel had no intention of continuing to produce in large quantities.
In the long run, the second story is more significant. AI is moving away from chatbots and toward agents, which are programs that plan, carry out, call APIs, query databases, and evaluate their own work continuously. When you ask ChatGPT a question, the GPU does about 90% of the work and the CPU hardly does any work at all. However, an agent? A CPU makes up the majority of an agent. Organizing subtasks, calling tools, branching, and planning. On a graphics card, none of that functions well. Lisa Su of AMD recently advised investors to anticipate robust double-digit growth in server CPUs through 2026, possibly sensing where demand is headed.
In the meantime, the supply side is caught up in its own web. There are too many chips coming off the line that can’t be sold, which is causing Intel to struggle with yield issues at its factories. AMD relies on Taiwan-based TSMC, which has been focusing on higher-margin AI accelerators and freely acknowledging that it can only provide roughly one-third of what its largest clients require. AMD does not manufacture any chips of its own. The system becomes brittle when the global memory shortage is added on top.
A side story is worth seeing. With its server chips already operating inside Meta’s data centers for workloads that don’t require a GPU at all, Nvidia—never one to overlook a margin—has been pushing into CPUs. In January, Jensen Huang stated that he anticipates Nvidia becoming a significant CPU manufacturer. It’s difficult not to interpret that as a jab at Intel.
Some contend that more hardware isn’t the true solution. The only long-term solution, according to an increasing number of people, primarily enterprise IT leaders, is smarter software, such as memory tiering, oversubscription, and optimization layers that maximize what is already racked. It’s genuinely unclear if this pragmatism takes off or if the industry just keeps investing in silicon.
The AI boom has finally outgrown the cozy belief that hardware would always be available whenever someone needed it, as can be seen when passing any contemporary data center humming behind its chain-link fence.
