JPMorgan says CPUs drive AI spend

The AI chip story has mostly been told as a GPU story. That made sense when the hard part was training giant models. But the workload is changing. JPMorgan’s latest AI infrastructure note leans into that shift and says the next spending wave is going to need a lot more CPUs — not instead of accelerators, but alongside them, in much bigger quantities. (jpmorganchase.com) ### Why are CPUs back in the conversation? Because inference is not training. Training is the giant, concentrated burst where GPUs do the heavy math. Inference is the day job — serving responses, moving data, managing memory, routing requests, coordinating tools, and handling all the messy orchestration around the m(jpmorganchase.com)e work than investors had been pricing in. (jpmorganchase.com) ### What changed in the workload? The big change is that AI is moving from one-shot chatbot prompts to systems that reason in steps, call tools, fetch data, and keep running. Arm built its whole March 24 launch around that idea. It said continuously running agents increase token volume and require significantly more CPUs for reasoning, coordination, and data movement. That is a very different shape from the old “stuff the box with GPUs” mindset. (newsroom.arm.com) ### Why does agentic AI need more CPUs? Think of the GPU as the engine and the CPU as traffic control. If the engine gets faster but the traffic lights, ramps, and dispatchers do not, the whole system jams up. Agentic AI creates more of that traffic-control work — scheduling jobs, handling retrieval, managing state, moving outputs between systems, and keeping enterprise software co(newsroom.arm.com)cally as the host and action CPU in mixed inference designs with accelerators. (newsroom.intel.com) ### Is this just Arm talking its book? Partly, yes — but that does not make the thesis fake. Arm has every incentive to argue CPUs matter more, especially after launching its first in-house data-center chip, the AGI CPU, with Meta as lead partner on March 24, 2026. But the broader market is moving the same way. Nvidia’s own Gra(newsroom.intel.com)tinct GPUs with EPYC CPUs rather than pretending the CPU disappears. (newsroom.arm.com) ### What is the most important number here? Arm’s headline number is the one JPMorgan and investors keep circling: more than 4x the current CPU capacity per gigawatt for agentic AI data centers. A related estimate floating around the ecosystem is roughly 120 million CPU cores per gigawatt of incremental AI capacity. You should treat that as directional, not gospel, but the direction is the point — CPU density rises as inference scales. (newsroom.arm.com) ### Who benefits if JPMorgan is right? Intel and AMD are the obvious direct beneficiaries because they sell the server CPUs that sit next to accelerators in enterprise and cloud deployments. Arm benefits differently — through its own AGI CPU push, through Neoverse-based designs, and through the broader licensing ecosystem behind custom cloud chips. The winners are not the companies replacing GPUs. They are the ones supplying the rest of the rack. (intc.com) ### What is the catch? The catch is that GPUs are still central. Nothing in this thesis says the AI buildout stops being accelerator-heavy. JPMorgan’s point is subtler: as spending shifts from pure training clusters toward real-world inference, the CPU share of the bill can rise faster than people expected. That changes procurement math and, eventually, market leadership inside AI infrastructure. (jpmorganchase.com) ### Bottom line? This is really a story about AI maturing. When the bottleneck was model training, GPUs got all the attention. When the bottleneck becomes running AI systems all day inside actual businesses, the forgotten parts of the stack start mattering again — and the CPU is first in line. (jpmorganchase.com)