AI Agents Need CPUs, Not Just GPUs: The Coming Compute Crunch Nobody Is Talking About

The market has obsessed over graphics processors for two years. Nvidia’s market capitalization tripled. Startups such as CoreWeave and Nebius raised billions to build “neoclouds” stuffed with GPUs for model training and inference. But the infrastructure bottleneck for the next phase of artificial intelligence may not be graphics chips at all.

It may be the humble central processing unit.

Daytona, a San Francisco startup that provides instant virtual machines for AI agents, closed a $60 million Series B on Tuesday at a $1 billion valuation. FirstMark led the round. The cap table reads like a directory of the people actually building agentic products: Stripe’s Patrick Collison, Shopify’s Tobi Lütke, Perplexity’s Aravind Srinivas, Quora’s Adam D’Angelo, Zoom’s Eric Yuan. Strategic backers include Qualcomm Ventures, Samsung Next and Okta Ventures. About $50 million came from existing investors, bringing total funding to roughly $91 million.

The funding is notable. The reasoning behind it is more so.

“The inference clouds is where the agent thinks, and our cloud is where the agent works,” said Ivan Burazin, Daytona’s cofounder and chief executive. He argues that every autonomous agent — one that writes code, browses the web, operates business software — needs a computer to act on. A virtual machine, or sandbox, isolated from the user’s laptop and from sensitive data. Existing hyperscalers can spin those up in 30 seconds to several minutes. Daytona does it in 60 milliseconds by keeping pre-configured machine snapshots on physical drives and attaching compute and memory on demand.

That speed difference is not a convenience. It is a prerequisite for the scale agent builders now anticipate.

Meta’s Muse, a personal agent launched in September, provisions “a computer underneath for every single chat,” Burazin said. His back-of-the-envelope math: serving 100 million users would require $2.8 billion worth of sandboxes. Snorkel, an AI data company, has already provisioned more than one million sandboxes on Daytona’s platform. Customers include Anthropic, OpenAI, Nvidia, Cursor, Cognition, Meta and Cohere. Daytona’s revenue run rate grew 25-fold between January and September, averaging 50.9 percent month-over-month.

Hyperscalers could theoretically match this architecture. Burazin’s pitch to skeptical investors was that Amazon Web Services, Microsoft Azure and Google Cloud Platform optimized their entire stacks for maximum utilization of a few core services — built for human users, not millions of ephemeral agent sessions. Rebuilding that infrastructure would amount to creating a new company inside the old one.

Chipmakers are already seeing the shift. On AMD’s first-quarter earnings call in May, Chief Executive Lisa Su said agentic AI was increasing the need for server CPUs. The company raised its forecast for server CPU market growth from roughly 18 percent annually to more than 35 percent, projecting a market exceeding $120 billion by 2030. Arm estimates that agentic AI data centers could require 120 million CPU cores per gigawatt — four times the density of traditional data centers. Intel reportedly told some Chinese customers earlier this year that lead times for certain server CPUs had stretched to six months.

“We internally sort of believe that the number of dollars spent on CPUs will probably be higher than the dollars spent on GPUs in the foreseeable future,” Burazin said. “Which seems sort of counterintuitive today.”

He forecasts a CPU crunch. “The amount of CPU machines that we will need if we want to keep growing at this trajectory actually doesn’t exist right now.”

Daytona plans to spend most of the new capital on go-to-market sales staff and additional engineers. The company was founded in 2023 by Burazin, Goran Draganić and Vedran Jukić to automate software development environments for large enterprises. When the founders realized agents would need the same infrastructure, they gave their existing customers to competitors, rebuilt the product from scratch with roughly three months of runway left, and relaunched.