Register and share your invite link to earn from video plays and referrals.

Carrio
@carrioresearch
EE background, covering photonics from InP up and the rest of the AI supply chain|News, traced to the bottleneck|Deep research on Substack. NFA
Joined January 2026
93 Following    3.5K Followers
$INTC $AMD $MU $SNDK $STX $WDC WOW! GOOGLE'S AI INFRASTRUCTURE CHIEF SAYS AGENTS ARE SENDING CPU AND STORAGE DEMAND "THROUGH THE ROOF," RIGHT ALONGSIDE TPUs AI agents are pushing up Google's demand for CPUs, memory and storage along with its demand for TPUs, Amin Vahdat, the company's chief technologist for AI infrastructure, said in an interview Sequoia Capital released on Oct. 6. "The demand for accelerated compute is going up, but the demand for CPU and networking and storage... is also going through the roof at the same time," he said. - Time between requests: "seconds, maybe tens of seconds" when a human reads each answer, "perhaps milliseconds" when an agent does - Between model calls, the parsing and reasoning "is going to probably happen on a CPU" - That CPU then pulls context "from DRAM local to me, maybe from someone else's DRAM on another CPU, maybe on SSD, or maybe on HDD somewhere else" - A storage rack draws 10 to 40 kilowatts against "hundreds of kilowatts" for a TPU or GPU rack, so Google may put CPUs and hard drives in their own buildings and network them to the accelerators - Serving clusters are "not necessarily cheaper" per megawatt than training clusters, because serving needs "the mix of storage, compute and accelerators" in one place I put his agent answer next to what he said about Google's new inference chip, the TPU 8i. Google split it off because it expects inference to be 30% to 60% of the chip market over the hardware's life, and inference is the part of the build he says can't run on accelerators alone. So CPU, DRAM, flash and disk orders now ride the same curve as Google's inference build-out. That CPU is x86 from $INTC and $AMD or Google's own Arm-based Axion. The context sits on DRAM from $MU, flash from $SNDK and hard drives from $STX and $WDC. Every extra step an agent takes is extra work for all of them, whether or not another TPU gets installed. A chatbot waited seconds on a human. An agent waits milliseconds, and spends them on a CPU and a disk.
Show more