YTD performance tells you everything you need to know about CPUs becoming the new AI bottleneck: • $INTC +110% • $ARM +93% • $AMKR +87% • $AMD +50% As AI shifts toward inference and agents, the bottleneck moves to the CPU layer that schedules work, manages memory and keeps multi-step systems running at scale.
This chart (from NVIDIA’s 2026 Dynamo post) shows cumulative KV-cache reads vastly outpacing writes (read/write ≈ 11.7x) for agentic inference workflows—illustrating the heavy memory/cache pressure and reuse patterns that force systems to offload and manage KV state across GPU→CPU→NVMe tiers. It directly supports the tweet’s point that as inference and agentic workloads grow, the host CPU (which schedules work and manages host memory/KV cache) becomes a critical bottleneck.
Source: NVIDIA Developer Blog
Research Brief
What our analysis found
A viral tweet claiming that year-to-date stock performance of CPU-adjacent companies proves CPUs are becoming the new AI bottleneck has sparked debate across the investment and technology communities. The core thesis — that the shift from AI training to inference and agentic workloads is dramatically elevating the importance of CPUs — is well-supported by industry data. TrendForce expects CPU-to-GPU ratios in agentic AI deployments to shift from the traditional 1:4–1:8 toward 1:1–1:2, while Arm estimates a fourfold increase in CPU cores demanded per gigawatt in the AI agent era, rising from 30 million to 120 million cores. Morgan Stanley projects the server CPU total addressable market could exceed $100 billion, with $32.5–$60 billion in incremental CPU TAM by 2030.
The operational evidence is equally compelling. CPU-side orchestration and tool processing now account for 50% to 90% of total workload latency in agentic systems, and a quiet supply crisis has emerged — server CPU prices rose approximately 30% in Q4 2025, with AMD reporting delivery lead times stretching beyond ten weeks and some models facing delays of up to six months. Intel and AMD have confirmed that high-core-count server processors are effectively sold out. As AI models scale beyond 1 million tokens, Key-Value caches can reach approximately 200 GB, far exceeding typical GPU VRAM and making CPU memory capacity and bandwidth critical infrastructure.
However, the tweet's specific stock performance figures are significantly at odds with verified 2024 data. Intel's stock was actually down 58.7% for full-year 2024, not up 110%, and Amkor Technology declined 20.8% rather than gaining 87%. AMD's 2024 performance was also negative by roughly 11.3% as of late December 2024, despite record revenue. The underlying thesis about CPU importance in agentic AI carries substantial weight, but the stock data used to illustrate it appears to be inaccurate or drawn from a different, unspecified time period.
Fact Check
Evidence from both sides
Supporting Evidence
CPU-to-GPU ratio shift in agentic AI
TrendForce expects CPU-to-GPU ratios in agentic AI deployments to move from the traditional 1:4–1:8 toward 1:1–1:2, reflecting dramatically increased CPU demand per AI server. Arm estimates CPU core demand per gigawatt will quadruple from 30 million to 120 million in the agent era.
CPU orchestration dominates agentic workflow latency
CPU-side orchestration and tool processing account for 50% to 90% of total workload latency in agentic systems, confirming that CPU capacity has become a material bottleneck in multi-step AI pipelines.
Server CPU supply crisis and price surges
Intel and AMD have confirmed high-core-count server processors are effectively sold out. Server CPU prices rose approximately 30% in Q4 2025, and AMD reported delivery lead times exceeding ten weeks, with some models delayed up to six months — consistent with a genuine supply bottleneck.
Massive projected server CPU market growth
Morgan Stanley estimates $32.5–$60 billion of incremental CPU total addressable market by 2030, with the overall server CPU TAM exceeding $100 billion, up from $26 billion in 2025.
GPU idling caused by CPU constraints
In agentic workflows, insufficient CPU resources can leave GPUs sitting idle while waiting for preprocessing, tool execution, or verification steps to complete, making CPUs a critical factor in overall system throughput and cost efficiency.
Industry leadership acknowledgment
AMD CEO Lisa Su noted in March 2026 that agentic workloads are pushing computation back onto traditional CPU tasks, and Intel's CFO admitted server CPU supply is absolutely constrained.
Memory offloading to CPUs
As AI models scale beyond 1 million tokens, KV caches can reach approximately 200 GB — far exceeding typical GPU VRAM — making CPU memory capacity and bandwidth essential for inference at scale.
Contradicting Evidence
The tweet's stock performance figures are largely inaccurate for 2024
Intel ($INTC) was down 58.7% for full-year 2024, not up 110% as claimed. Amkor Technology ($AMKR) declined 20.8%, contradicting the claimed +87%. AMD ($AMD) was down roughly 11.3% by late December 2024, not up 50%. The ARM figure of +93% also appears inconsistent with available data showing limited upside for the year. The tweet may reference a different or unspecified time period.
GPUs remain dominant for training and high-throughput inference
GPUs are still superior for large-scale parallel computation, particularly AI model training and high-throughput batch inference workloads, meaning the CPU bottleneck thesis applies primarily to agentic and orchestration-heavy use cases rather than all of AI.
CPUs outperform GPUs only in specific inference scenarios
For low-latency single-request inference or smaller models, CPUs can sometimes outperform GPUs by avoiding memory transfer overhead, but when a model fits entirely in VRAM, GPUs remain faster — limiting the universality of the CPU bottleneck argument.
Hybrid system-level optimization complicates the narrative
The optimal AI infrastructure solution increasingly involves balanced CPU-GPU systems where performance is defined at the system level rather than by any single component, suggesting the bottleneck framing oversimplifies a more nuanced architectural evolution.
Stock performance is a weak proxy for technical bottlenecks
Year-to-date stock movements reflect a complex mix of earnings, guidance, market sentiment, and macroeconomic factors — not a direct measure of whether CPUs have become AI's primary constraint. Using inaccurate stock data further weakens this line of reasoning.
Report an Issue
Found something wrong with this article? Let us know and we'll look into it.