TrendForce predicts that autonomous AI agents will drive a four-fold increase in data center CPU core requirements, forcing a fundamental rethink of server infrastructure balance. The research firm estimates core demand could jump from 30 million to 120 million cores per gigawatt of capacity.
The Resurgence of the CPU in the AI Era
For years, the prevailing wisdom held that the GPU had rendered the CPU a mere supporting actor in AI infrastructure. Traditional deployments settled on a ratio of roughly one CPU for every four to eight GPUs, sized for the chatbot era of large language model inference. TrendForce now says that model is obsolete. Agentic AI workloads — where autonomous agents plan tasks, schedule subtasks, call external tools and evaluate results — are intensely CPU-dependent. GPUs execute the model weights; the CPU orchestrates everything else.
The technical reality is that the CPU’s fundamental operation — the fetch-decode-execute cycle driven by the ALU, registers and control unit — has remained almost unchanged since the stored-program computer. What has changed is the packaging: modern server chips are multi-core microprocessors, often multithreaded, with vast cache hierarchies. That architecture, it turns out, is precisely what agentic workflows demand. Python execution, database queries, web crawling and inter-agent coordination run on integer logic and branch prediction, not tensor cores.
Why Agentic Workflows Require CPU Coordination
TrendForce highlights a striking figure: CPU-based tool processing can account for up to 90.6% of total latency in certain agentic workloads. The bottleneck is not the model inference but the scaffolding around it. An agent that must search a database, write code, call an API, parse the result and decide the next step spends the vast majority of its time in conventional compute. Dynamic energy consumption on the CPU can also represent a substantial portion of total system power under these conditions.
Arm estimates cited by TrendForce suggest the CPU-to-GPU ratio in agentic infrastructure will shift toward one CPU for every one to two GPUs. That is a radical inversion of the recent past. It implies not just more CPUs, but different ones: chips optimized for memory bandwidth, thread count and serialization latency rather than pure vector throughput.
TrendForce Forecast on Core Requirements
The four-fold core projection — 30 million to 120 million cores per gigawatt — is the headline number, but the detail that matters is the composition of that demand. Futurum’s channel checks find agentic deployments specifying ratios from 1:2 up to 100:1 CPUs per GPU as sandboxes and tool runners multiply. Their AI Supply Chain Market Model puts the server CPU market at roughly $246 billion by 2030, roughly six times the $40 billion market Intel CEO Lip-Bu Tan cites for today. Standalone AI CPUs are forecast to grow from $23.7 billion in 2026 to $164.7 billion in that window.
Intel’s Substrate Shortage and Supply Constraints
The supply chain, however, was sized for the old ratio. Intel reported Q3 2026 revenue of $16.5 billion, with Data Center and AI revenue up nearly 80% year over year to $7.3 billion. Yet the company warned that an industry-wide shortage of advanced organic substrates for EMIB packaging is limiting its ability to scale output. CEO Lip-Bu Tan noted that even as Intel 3 wafer output ramps, the substrate constraint bites. CFO Dave Zinsner said Intel has secured 40% more substrate capacity, most arriving in the second half of 2027, and projected the industry will remain supply constrained throughout 2027 and into 2028. Futurum’s market map traces more than 70 companies across design, manufacturing and end markets; the binding constraint has moved below the wafer, into the package.
Nvidia’s Entry Into the Server CPU Market
Nvidia’s Vera CPU arrives into this squeeze. Built on the Armv9.2-A architecture with 88 custom Olympus cores and Nvidia’s “spatial multithreading,” Vera is explicitly marketed for agentic AI workflows. Benchmarks show chart-topping performance in Python — the backbone of the AI software stack — and commanding memory bandwidth at both socket and system levels thanks to high-speed C2C interconnects. In contentious agentic environments where agents query large databases, that bandwidth is critical. Vera also demonstrates competent general-purpose server performance, often matching or beating Intel’s Granite Rapids-AP and AMD’s EPYC Turin in backend, data-rich operations, though x86 core density still wins on frontend workloads.
Implications for Data Center Infrastructure Balance
The shift rewrites the bill of materials for a data center. AMD’s Ryzen 9 9950X3D, launched in March 2025 with 128MB of L3 cache and a 5.7GHz turbo clock, illustrates how cache-heavy architectures benefit latency-sensitive workloads, though that part targets the client socket. In the server room, the rise of custom Arm and RISC-V designs — SiFive and others — will influence who owns the socket once foundry capacity expands. For now, market share through 2027 will be determined by who controls the slowest layer: advanced organic substrate supply. As ever, the bottleneck is not the silicon. It is the laminate.
Sources
- Central processing unit
- Intel's Substrate Shortage Moves the CPU Supply Chain Bottleneck Into the Package
- Benchmarking Nvidia’s Vera CPU for the data center — standalone Vera tested against Granite Rapids-AP and EPYC Turin
- AMD Ryzen 9 9900X3D vs Ryzen 9 9950X3D CPU Review
- Agentic AI boom expected to reshape CPU market
- What speeds up in a Game Boy Color game when you use double-speed mode? | GB Rogue Development Story #84 ROM Distribution｜河豚だろう | AIゲーム開発
Photo:
Dmitry Grigoriev (CC BY 3.0)
