Nvidia, AMD, ARM and Qualcomm spent the last three years selling investors the same idea: buy more GPUs. In the summer of 2026, all four quietly started selling a second one. The video above walks through the shift chip by chip, and the short version is that agentic AI does not run on GPUs alone.
AMD put a number on it in July, around its Advancing AI 2026 event. In the chatbot era, AMD says data centers ran roughly one CPU for every four to eight GPUs. In the agentic era, where software agents plan tasks, call tools, query databases and check their own output in loops, that ratio is heading toward one CPU for every GPU, sometimes more. AMD closed at $482.93 on August 12, a roughly $788 billion market cap, up more than 180% over the past year.
The CPU Stops Being the Side Character
AMD’s own framing is blunt: “the CPU isn’t a supporting actor to the GPU story anymore.” The company backed that line with hardware. On July 23, it launched the 6th Gen EPYC 9006 Series (“Venice”), built specifically around agentic workloads: up to 256 cores and 512 threads, PCIe Gen 6, and 16 channels of MRDIMM memory. Instead of one general-purpose CPU line, AMD split it into a three-tier model: chips for agent sandboxing, chips for feeding GPU clusters, and chips for ordinary enterprise workloads.
Dan McNamara, AMD’s SVP of Compute and Enterprise AI, put the pitch this way: 6th Gen EPYC exists for “running more agents and feeding GPUs,” not one or the other. AMD is also projecting that its server CPU market alone could grow at over 35% a year, up from an 18% estimate before agentic workloads entered the forecast, and pass $120 billion by 2030.
AMD Isn’t the Only One at This Table
The video’s chapter list is really a map of who else is placing the same bet:
- Nvidia pairs its GPUs with its own Grace and upcoming Vera CPUs, rather than leaving the CPU slot to Intel or AMD.
- ARM is licensing CPU designs aimed squarely at agentic orchestration, not just phones and laptops.
- Qualcomm is pushing into the server rack with its new C-1000 chip, a market it has never seriously competed in before.
Read together, this isn’t four companies reacting to the same trend by coincidence. It’s every accelerator vendor racing to own both halves of the rack before someone else’s CPU becomes the default one sitting next to their GPU.
Where the Thesis Runs Into Alphabet
This is also where the story gets more complicated for the CPU vendors than the ratio chart suggests. Alphabet is not shopping for either half of this bundle. It already designs its own Arm-based server CPUs (Axion) and its own AI accelerators, the TPU v6 and v7 silicon documented in our audit of Alphabet’s hardware moat. A hyperscaler that vertically integrates both chips is a customer nobody in this new CPU race gets to sell to.
That matters for sizing AMD’s $120 billion forecast: it assumes a buyer pool that keeps shrinking every time a hyperscaler the size of Google Cloud, Amazon or Microsoft moves another workload onto silicon it built itself.
What to watch next: whether AMD starts breaking out agent-sandbox CPU revenue separately when it reports Q3 2026 earnings, how fast Nvidia ships Vera against its own Grace roadmap, and whether the hyperscalers building in-house silicon keep growing their share of total AI infrastructure spend relative to the merchant-silicon buyers this ratio is supposed to help.
Analyst Note: A ratio chart is not a revenue guarantee. AMD’s CPU story only works at scale if the buyers still need to buy, and the three companies with the deepest pockets in AI are increasingly building their own. This is The Frequency.
