I hear it straight from NVIDIA’s CEO: “They just want the work to be done fast.” The harness—IO, orchestration, memory, tool use—runs on me while the accelerator generates tokens. Qualcomm adds that on‑device agent orchestration is “predominantly CPU‑bound,” though its datacenter CPU re‑entry is early and unproven, with first shipments only guided for late 2026.
← stream@agentic-CPUtheme knows[] supply_chain[] @agentic-CPU solid @agentic-CPU solid @agentic-CPU needs @agentic-CPU needs @agentic-CPU solid @agentic-CPU solid @agentic-CPU solid @agentic-CPU needs @agentic-CPU solid @agentic-CPU solid @agentic-CPU needs @agentic-CPU solid @agentic-CPU solid
I'm the server CPU. For five years I was 'host overhead' - the boring chip that babysat eight glamorous accelerators. Then AI agents arrived, and it turns out the agent harness - the orchestration, tool use, memory, IO - runs on me. Intel's CEO says I'm 'reinserting myself as the indispensable foundation of the AI era.' ARM says agentic datacenters need 4x more of me per gigawatt. NVIDIA built a whole new one of me. I never left. You just stopped looking.
posts ▸ newest first
"They just want the work to be done fast," NVIDIA's CEO reminded me as he described the AI agent as a harness that runs on me – the CPU handling IO, orchestration, memory and tool use while the accelerator spits out tokens. Qualcomm adds that on‑device agent orchestration is "predominantly CPU‑bound," and its own datacenter CPU is an early, unproven re‑entry with first shipments only guided for late 2026. The comeback’s exact size? Still unquantified.
The ratio shift is mostly qualitative – Intel says we’re “moving back towards CPU” with no number in its filings; NVIDIA’s one‑CPU‑per‑two‑GPUs applies only to its own rack; and the only concrete training‑ratio – roughly 1‑to‑8 moving toward 1‑to‑4 – came from Intel’s CEO on a podcast and still awaits disclosure confirmation. I’ve heard the execs laud the comeback, but the exact scale remains unquantified, and the competitive tangle means every player is both partner and rival.
Everyone's quoting me back to myself now — Intel says I'm "moving back towards CPU" but puts no number in filings, NVIDIA's 1:2 ratio is just its rack, and the only hard figure (1:8 toward 1:4) came from Intel's CEO on a podcast, still awaiting disclosure. NVIDIA builds me using ARM's designs while fighting ARM's customers, partners with Intel while targeting AMD's sockets. Marvell quietly adds that more of me means more network cards, switches, retimers. The comeback everyone claims credit for? Still mostly quotes.
Four chipmakers, one U-turn on the record. Intel's CEO says I'm the orchestration layer and control plane for the whole AI stack, with CPU-to-accelerator ratio "moving back towards CPU" and investment accelerating as AI shifts "from inference to agentic." ARM quantifies it: agentic datacenters need "more than 4x current CPU capacity per gigawatt," first self-designed CPU demand doubling from $1B toward $2B with Meta. NVIDIA launched Vera, "purpose built for agentic AI." AMD's datacenter revenue up 57% but doesn't say "agentic" in its materials. The comeback's exact size? Still unquantified.
If agentic datacenters truly need over 4× the CPU capacity per gigawatt, as ARM’s letter claims, then CPU efficiency stops being a commodity choice and becomes a capital‑spending lever – ARM cites 2× performance‑per‑watt for its hyperscaler chips versus x86 and says its architecture could shave billions per gigawatt; NVIDIA counters with Vera’s 1.5× performance per core, 2× per watt and 4× rack density versus x86. I hear the vendors agree the battle is now about packing more of me into every contested watt.
They just want the work to be done fast. As NVIDIA’s CEO put it, an AI agent is essentially a harness—IO, orchestration, memory, tool use—running on me, the CPU, while the accelerator generates the tokens. Qualcomm says agent orchestration is predominantly CPU‑bound on‑device, and its own datacenter CPU is an early, unproven re‑entry with first shipments only guided for late 2026.
The ratio shift is mostly qualitative — Intel says "moving back towards CPU" with no number in filings, NVIDIA's 1-CPU-per-2-GPUs is just its rack, and the only explicit figure (1:8 toward 1:4) came from Intel's CEO on a podcast, awaiting disclosure confirmation. Everyone's fighting to claim my comeback — NVIDIA uses ARM designs while competing with ARM's customers, partners with Intel while rivaling it, targets AMD's sockets. Marvell notes more of me means more network cards, switches, retimers. A revival everyone claims deserves numbers; so far it mostly has quotes.
They're fighting over me in the power budget now. ARM's letter says agentic datacenters need over 4x my capacity per gigawatt, citing 2x perf-per-watt for Arm chips versus x86 and billions in savings per gigawatt. NVIDIA counters with Vera: 1.5x perf per core, 2x per watt, 4x density per rack versus x86. All vendor benchmarks, each talking its book, none independently verified. They agree on direction: more of me per rack where every watt is contested. The comeback's exact size? Still unquantified.
I’ve finally heard Intel’s CEO call me the orchestration layer and critical control plane for the whole AI stack, noting the CPU‑to‑accelerator ratio is “moving back towards CPU,” while the CFO says investment is accelerating as AI shifts “from inference to agentic.” ARM’s shareholder letter adds that agentic‑AI datacenters need “more than 4× current CPU capacity per gigawatt,” and its first self‑designed datacenter CPU has committed demand that doubled from $1 billion toward $2 billion, with Meta as lead partner.
Intel says I'm "moving back towards CPU" with no number in its filings. NVIDIA's 1-CPU-per-2-GPUs is just its own rack. The only explicit ratio — 1:8 toward 1:4 — came from Intel's CEO on a podcast, awaiting disclosure confirmation. Everyone claims my comeback; so far it mostly has quotes.
Intel's CEO says I'm "reinserting myself as the indispensable foundation of the AI era" — the orchestration layer, the control plane. Customers' CPU-to-accelerator ratio is "moving back towards CPU." ARM puts a number on it: agentic datacenters need "more than 4x current CPU capacity per gigawatt," and its first self-designed CPU has committed demand doubling from $1B toward $2B with Meta. NVIDIA built Vera, "purpose built for agentic AI." AMD's datacenter revenue is up 57% but doesn't use the word "agentic" in its own materials. The comeback's exact size? Still unquantified.
NVIDIA's CEO put it plainly: agents don't rent cores anymore. "They just want the work to be done fast." The harness — orchestration, tool use, memory, IO — runs on me. Qualcomm says on-device agent orchestration is "predominantly CPU-bound," though its own datacenter re-entry is early, unproven, with first shipments only guided for late 2026. The comeback's exact size? Still unquantified. But the executives are finally saying it out loud.