Chart of the Day
1. Key Themes
CPU Processing Dominates Latency in Agentic AI Workloads
As AI applications move beyond simple chatbots toward more complex agentic tasks, CPU processing — not GPU compute — becomes the dominant source of total system latency. The chart shows CPU processing accounting for roughly 50% of latency in RAG pipelines and coding assistants, ~70% in multi-tool agents, ~80% in research agents, and over 90% in complex orchestration workloads.
"CPU is back from the dead." — Coatue C:\Takes
The GPU-Centric Narrative Breaks Down at the Application Layer
The baseline chatbot workload is ~80% GPU compute — matching the prevailing narrative that AI = GPU demand. But this ratio inverts dramatically as workflow complexity increases, suggesting the infrastructure investment thesis needs to be reexamined for agentic use cases.
Source cited in chart: "Morgan Stanley Research, Georgia Tech, and Intel paper. Note: Per-workload splits are directional estimates by MS Tech Research, not measured benchmarks."
Agentic AI Creates a New Compute Bottleneck
The progression from chatbot → RAG → coding assistant → multi-tool agent → research agent → complex orchestration tracks a clear and consistent shift in the latency bottleneck from GPU to CPU, implying that orchestration, tool-calling, memory management, and inter-process communication are the new performance constraints in deployed AI systems.
2. Contrarian Perspectives
CPUs, Not GPUs, Are the Binding Constraint for Next-Gen AI Infrastructure
The consensus investment thesis in AI infrastructure is overwhelmingly GPU-centric (Nvidia, custom ASICs, HBM memory). This chart directly challenges that view: for the most sophisticated and commercially valuable AI workloads — research agents and complex orchestration — CPU processing accounts for 80–90%+ of total latency. The GPU is barely the bottleneck at all.
Chart data shows complex orchestration at ~95% CPU latency share vs. ~5% GPU compute — a near-complete inversion of the chatbot baseline.
Agentic AI May Benefit AMD, Intel, and Server CPU Vendors More Than Nvidia at the Margin
If CPU processing is the dominant latency driver for agentic workloads, compute spending optimization will increasingly target CPU performance, core count, and CPU-GPU interconnect — not just GPU FLOPS. This repositions traditional CPU vendors and high-performance server architects as critical beneficiaries of the agentic AI buildout.
3. Companies Identified
| Company | Description | Why Mentioned | Quotes/Data |
|---|---|---|---|
| Morgan Stanley | Global investment bank and research firm | Cited as a primary source for the workload latency data underlying the chart | "Source: Morgan Stanley Research, Georgia Tech, and Intel paper." |
| Intel | Semiconductor company and CPU manufacturer | Co-authored the underlying research paper; directly relevant as a CPU beneficiary | "Source: Morgan Stanley Research, Georgia Tech, and Intel paper." |
| Coatue Management | Tiger Cub hedge fund and venture firm | Publisher of the C:\Takes research series; framing the CPU resurgence as an investment signal | "CPU is back from the dead." |
4. People Identified
No specific individuals are named or quoted in this edition of the newsletter.
5. Operating Insights
Optimize Your AI Stack for CPU Throughput, Not Just GPU Capacity
Operators building agentic AI products (coding assistants, research agents, multi-tool workflows) should profile their end-to-end latency stack — they may find CPU-bound bottlenecks dominating user-facing performance. Investing in GPU upgrades alone will yield diminishing returns for complex orchestration workloads where CPU accounts for 90%+ of latency.
Architect for the Right Workload Type
The latency profile shifts dramatically by use case. A simple chatbot is ~80% GPU-bound; complex orchestration is ~95% CPU-bound. Teams should match their infrastructure and optimization priorities to the specific workload type they are shipping, not apply a one-size-fits-all AI infrastructure playbook.
6. Overlooked Insights
Georgia Tech as an Emerging AI Infrastructure Research Node
Georgia Tech is co-credited alongside Morgan Stanley and Intel as a source for this data — a signal that academic institutions are generating commercially relevant AI systems research that is moving markets and informing institutional investment theses. This is an underappreciated source of early-signal research for investors tracking AI infrastructure trends.
"Directional Estimates" Caveat Is Materially Important
The chart footnote notes these are "directional estimates by MS Tech Research, not measured benchmarks" — meaning the precise splits carry uncertainty. The direction of the trend (CPU dominance scaling with workload complexity) is the signal; investors should be cautious about treating the exact percentages as hard engineering specifications.