Agents Are the New Users of CPUs
1. Key Themes
Theme 1: Agents Are Replacing Humans as the Primary Users of Computing Infrastructure
The fundamental shift Coatue identifies is that AI agents — not humans — are now the primary drivers of CPU workloads. Tasks previously requiring human cognition and manual execution are now run autonomously in continuous loops.
"Before AI, a human would think about what to do and then take action on their computer screen. Today the GPU plans the action by writing tokens, and the CPU executes it by writing code."
Theme 2: The GPU-CPU Loop Is the New Unit of Compute Demand
The article introduces a structural model of agentic computing: GPUs handle reasoning ("token time") while CPUs handle execution ("tool time"). Every agent action is a round-trip between the two, and those round trips multiply at machine speed.
"The GPU decides the next move, the CPU carries it out, and the cycle repeats. Every cycle is a GPU-to-CPU round trip."
The attached diagram formalizes this as an alternating clock: GPU phases (Plan → Pick Source → Spot Trend → Check → Deliver) interleave with CPU phases (Search Tools → Query Database → Build Chart → Write and Export File).
Theme 3: Agentic Workflows Drive Massive, Compounding CPU Demand
Because agents operate faster, longer, and at greater depth than humans, the volume of CPU calls is orders of magnitude higher than in pre-AI workflows — even for routine tasks like data analysis.
"That is massive CPU demand, because agents can move faster, longer, and deeper than a human possibly could."
"The workflow a lot of people can relate to is pulling data, analyzing it, and displaying a chart. Every single one of those steps can now be done by an agent using a computer driven by a CPU."
Theme 4: The Agent Is the Orchestrating Layer Between GPU and CPU
The diagram in the article positions the "Agent" as a distinct architectural layer — the orchestrating loop — sitting between the GPU (token factory) and CPU+Memory (tool executors). This has implications for where software and infrastructure value accrues.
From the diagram: GPUs are labeled "The token factory," Agents are "The orchestrating loop," and CPUs+Memory are "The tool executors" — suggesting three distinct, investable infrastructure layers.
2. Contrarian Perspectives
The CPU Renaissance: The Infrastructure Story Isn't Just About GPUs
The consensus AI infrastructure narrative is almost entirely GPU-centric (Nvidia, H100s, etc.). Coatue's take is that the agentic shift creates a secondary and underappreciated surge in CPU demand. While GPUs do the thinking, every tool call, database query, file retrieval, and API call runs on CPUs — and agents generate far more of these calls than humans ever did.
"Every single one of those steps can now be done by an agent using a computer driven by a CPU. That is massive CPU demand, because agents can move faster, longer, and deeper than a human possibly could."
This implies that traditional compute infrastructure — cloud CPU instances, data infrastructure, APIs — may be significantly undervalued relative to GPU-focused investments.
Human-Paced Software Was a Bottleneck, Not a Feature
Conventional software was designed around human interaction speeds — click, read, decide. Coatue implies that this was always a performance ceiling, not an intentional design. Agents dissolve that ceiling entirely.
"Before, you had to click, read, and plan a task yourself. Now an agent runs that same sequence in a loop."
This suggests that legacy SaaS products optimized for human-paced workflows will face structural obsolescence pressure, not merely feature competition.
3. Companies Identified
| Company | Description | Why Mentioned | Quotes |
|---|---|---|---|
| Coatue Management | Prominent tech-focused hedge fund and venture firm | Publisher/author of the analysis; providing the investment thesis | "Coatue opinion and analysis as of June 2026." |
Note: No external portfolio companies or case study companies are named in this edition. The content is a conceptual/thematic piece.
4. People Identified
| Person | Description | Why Mentioned | Quotes |
|---|---|---|---|
| Frank (last name not provided) | Coatue analyst or partner | Co-presenter of the C:\Takes video series elaborating on this thesis | "For more on this C:\Take, watch Frank and Nick" |
| Nick (last name not provided) | Coatue analyst or partner | Co-presenter alongside Frank | "For more on this C:\Take, watch Frank and Nick" |
5. Operating Insights
1. Redesign Workflows Assuming Agent Execution, Not Human Execution
Any workflow that currently requires a human to pull data, analyze it, and produce output is a candidate for full agent automation — not just partial automation. Operators should audit their processes with the question: "Could an agent run this entire loop without a human in it?"
"The workflow a lot of people can relate to is pulling data, analyzing it, and displaying a chart. Every single one of those steps can now be done by an agent using a computer driven by a CPU."
2. Infrastructure Costs Will Shift — Plan for CPU/Tool-Layer Spend, Not Just GPU Spend
As companies deploy agents at scale, their compute cost structure will include not just LLM inference (GPU) costs but a growing tail of CPU-driven tool calls — API queries, database reads, file I/O. Finance and engineering leaders should model this "tool time" cost separately.
"Every cycle is a GPU-to-CPU round trip."
6. Overlooked Insights
1. Memory as a Distinct, Investable Infrastructure Layer
The diagram labels the execution side as "CPUs + Memory" — not just CPUs. Memory (likely referring to vector stores, retrieval systems, or context caches) is called out as a co-equal component of the tool execution layer. This is easy to overlook but suggests that memory infrastructure (e.g., vector databases, KV caches) is a critical and potentially undervalued part of the agentic stack.
From the diagram: The right panel is explicitly labeled "CPUs + Memory — The tool executors," distinguishing memory as an architectural primitive alongside compute.
2. The "Token Time vs. Tool Time" Framework as a New Mental Model for Agent Efficiency
The diagram introduces a temporal framing — "token time" (GPU) vs. "tool time" (CPU) — that implies agents have two distinct performance bottlenecks. This framework is briefly presented but has significant implications: optimizing agent performance requires balancing both clocks, not just making the LLM faster. Startups building agent orchestration or observability tools could use this as a core design and pricing framework.
From the diagram title: "An agent's clock alternates between token time and tool time."