GPU count has stopped being a reliable measure of capacity for agentic AI. As agents plan, call tools, and retry, the work spreads onto the CPU layer and the network fabric that keep every accelerator fed. AMD reports the CPU-to-GPU ratio moving toward 1:1, and since October 2026 NVIDIA, AMD, and Intel have all put agentic CPU demand at the center of their data center plans.
What Agentic AI Changes in the Cluster
Agentic AI describes systems that complete multi-step tasks on their own. An agent plans a sequence of actions, calls tools, checks outputs, and retries when a step fails. Each task, in turn, fans out into many steps, and every step consumes CPU time and network traffic alongside GPU cycles.
As a result, a cluster can hit its GPU target exactly and still stall. For years, buyers reduced the AI capacity conversation to a single number. That number still matters; however, it captures only part of what an agentic system consumes. Chip vendors, network vendors, and financial researchers now reach the same conclusion from different directions. As AI systems shift from answering questions to completing tasks, the bottleneck moves off the GPU.
How Much More Compute Does an Agentic Task Use?
J.P. Morgan Asset Management estimates that a single user could demand 10 to 100 times more compute with the latest agentic tools. Quartz puts the token figure much higher, reporting that AI agents can consume 1,000 times more tokens than a single chatbot query. Ciena describes the same effect in concrete terms: a single inference request of 50 tokens can evolve into a 50,000-token job as the work fans out across agents.
A chatbot query is bounded, because the transaction closes once the model answers. By contrast, agentic work behaves more like a running program. One request fans out into planning steps, tool calls, verification loops, and retries, and every hop costs tokens. We quantified that loop in our analysis of agentic AI compute.
Ciena is explicit that the expansion covers more than GPU compute, since it also means more of the orchestration logic and tooling that run on CPUs. Meanwhile, the J.P. Morgan and Quartz figures do not conflict. J.P. Morgan describes demand per user, while Quartz and Ciena describe what a single complex task can grow into. On either figure, a capacity plan that tracks users but ignores steps per task will undercount demand.
Why the Multiplier Compounds
The multiplier keeps growing while a task runs. Agentic systems retry failed steps and validate outputs against external tools. Often, a system also runs several candidate paths before it commits to a final answer.
Each of those steps adds tokens on top of the original request. Consequently, a forecast built on user count alone understates real demand by a wide margin. The planning unit therefore has to change: count completed tasks per month, and state an explicit assumption about steps per task.
The CPU-to-GPU Ratio Is Moving Now
AMD reports that agentic AI is moving the CPU-to-GPU ratio toward 1:1 and, in some cases, higher on the CPU side, up from the one CPU per four to eight GPUs typical of chatbot-era inference servers. For that reason, an audit that counts accelerators only will miss the added CPU capacity entirely.
That ratio changes how a cluster gets budgeted. Under the older ratio, an inference server spent nearly all of its silicon budget on accelerators, while a small number of CPUs shuttled traffic. In the newer ratio, by comparison, the CPU layer carries real load. AMD frames the fix as architectural, calling for a newly engineered CPU compute layer for agent execution and tool calls, with real-world checks on every autonomous action.
AMD CEO Lisa Su made the point first on the company’s first-quarter 2026 earnings call in May, describing a CPU that had served mostly as a host node in a one-to-four or one-to-eight configuration and was now getting closer to one to one. By the third-quarter call in November, AMD described personal and enterprise agents as “quickly becoming the next killer app for CPUs.”
A Note on the Ratio Itself
The one-to-four and one-to-eight figures describe the classic eight-GPU server with x86 host processors. Rack-scale systems already sit closer to the middle: NVIDIA’s Vera Rubin NVL72, for example, pairs 72 Rubin GPUs with 36 Vera CPUs. The newer step, therefore, happens at the level of the data center rather than the server. Operators now add dedicated CPU racks beside their GPU racks, so the ratio that matters is the one across the whole cluster.
The Industry Is Now Building CPU Racks for Agents
When we first published this analysis in October, AMD was the clearest voice on the ratio. Since then, all three major chipmakers have put agentic CPU demand at the center of their data center plans, and two of them now sell systems built for it.
NVIDIA: A CPU Built for Agents
In May 2026, NVIDIA launched Vera as “the first CPU built for AI agents,” with 88 custom Olympus cores and up to 1.2 TB/s of memory bandwidth. A Vera rack holds up to 256 CPUs and supports more than 22,500 concurrent CPU environments, according to NVIDIA. Jensen Huang framed the reason directly: “AI agents will be the largest users of computing.” NVIDIA reports that the first Vera systems went to leading AI labs, including OpenAI, Anthropic, and SpaceXAI.
AMD: CPU-Only Racks Beside GPU Racks
In October 2026, AMD introduced Helion, an open rack design for agentic AI. Its CPU-only version holds up to 32,768 EPYC cores, while a second version combines EPYC CPU nodes with Instinct GPU nodes. AMD said in November that the first systems ship this quarter. In the same update, AMD reported data center revenue up 97 percent to a record $8.5 billion and said it now expects server revenue to double in 2027.
Intel: Demand Ahead of Supply
Intel reported data center and AI revenue of $7.3 billion in its third quarter of 2026, up 78 percent, and said demand “continues to meaningfully exceed available supply.” As a result, Intel is shifting manufacturing capacity from PC processors toward data center CPUs. CEO Lip-Bu Tan estimated that the server CPU market, largely driven by agentic workloads, could grow from $40 billion to more than $200 billion over five years.
Why Agents Need So Many CPU Environments
The Vera figure of 22,500 concurrent CPU environments points to a practical reason for the shift. Agents do more than generate text: they run code, open browsers, query databases, and call tools, usually inside isolated sandboxes. Each of those sandboxes runs on CPU cores. Consequently, every agent session that executes work needs CPU capacity of its own, on top of the GPU time its model consumes.
What the CPU Actually Does in an Agent Loop
Semiwiki, summarizing an Intel white paper on agentic AI and CPUs, lays out the mechanics. In agentic workflows, CPUs handle orchestration tasks such as control flow management, branching logic, retries, and coordination between agents and external services. This control work runs on the CPU while the GPU waits for its next input. Consequently, a thin CPU layer leaves expensive accelerators idle.
The same article states the consequence plainly. Under-provisioned CPU resources “can introduce latency in orchestration, delay tool execution, and slow verification loops, all of which reduce GPU utilization and increase operational costs.” A cluster with ample GPU headroom can therefore underperform its own hardware specification, and in practice that gap shows up only once an agentic workload runs on it.
One caveat applies to the vendor sources in this article. NVIDIA, AMD, and Intel all sell server CPUs, and Ciena sells networking, so each benefits from the trend it describes. That is why the demand figures above come from outside that supply chain, from J.P. Morgan Asset Management and Quartz. Reported quarterly revenue, however, is harder evidence than a vendor blog, and both AMD and Intel now report it.
The Fabric Has to Behave Differently
The shift also reaches the network fabric. Ciena compares an agentic workflow to an operating system: spawning threads, managing memory, calling tools, engaging through APIs, and iterating through reinforcement loops. Each of those actions is a call from orchestration code on a CPU to a model, a memory store, or an external tool. As a result, every interaction crosses the network, and many cross it repeatedly inside a single user request.
The causal chain runs in a straight line. More CPU-side orchestration means more traffic between CPUs, accelerators, memory, and tools, and that traffic arrives as short, bursty calls. Fabrics built for training expect large, predictable batches, so they handle the bursty pattern badly. Separate CPU racks add another hop to that path. Ciena therefore argues that networks must evolve from passive interconnects into programmable, high-capacity infrastructure. We cover the fabric choice itself in InfiniBand vs Ethernet for AI networking.
Chatbot-Era Cluster Versus Agentic-Era Cluster
The table below sets out what changes between the two, using the vendor and analyst figures above.
| Dimension | Chatbot-era inference | Agentic-era inference |
|---|---|---|
| CPU-to-GPU ratio | 1:4 to 1:8 per server (AMD) | Toward 1:1 across the cluster, sometimes higher on CPU (AMD) |
| CPU form factor | Host processors inside GPU servers | Dedicated CPU racks beside GPU racks (NVIDIA Vera, AMD Helion) |
| Compute per user | Baseline | 10x to 100x with agentic tools (J.P. Morgan) |
| Tokens per request | Bounded, single pass | 1,000x more than a chatbot query (Quartz); 50 to 50,000 tokens (Ciena) |
| Bottleneck location | GPU-centric inference | CPU-heavy orchestration, tool execution, and sandboxes (Semiwiki, Intel white paper) |
| Network profile | Predictable, batch-oriented | Bursty, needing programmable, high-capacity fabric (Ciena) |
Why GPU Count Was Always an Incomplete Metric
The GPU stays central to agentic inference. Even so, GPU count on its own makes a weak proxy for capacity. A cluster sized against a target accelerator number can hit that number precisely and still fail its workload, because the CPU layer, fabric, storage, and orchestration software were never sized against the same target.
Under agentic workloads, the scarce input therefore extends past the chip. It now includes the whole system that keeps the chip fed and connected to the tools an agent needs.
For a business forecasting capacity a year out, the consequence is uncomfortable. A plan to double agentic users has to cover CPU capacity and fabric alongside GPU orders. Under AMD’s ratio figures, CPU demand can grow faster than accelerator demand, and orchestration software may need re-architecting at the same time. Intel’s supply shortfall adds a further reason to plan early: server CPUs are no longer the easy part of the order. A GPU-only forecast, however, never surfaces that schedule. Our review of enterprise GPU strategy for 2026 reaches the same point from the buyer’s side: the workload defines the specification.
Designing Against the Ratio
A cluster built for agentic work treats the CPU layer, fabric, storage, and orchestration stack as first-class design decisions, made before the accelerator count is frozen. Capacity conversations that begin and end with chip count leave those decisions to chance. Axe Compute organizes the design work around four principles, each matched to a problem the ratio shift creates.
Choice: Configured to the Ratio
Agentic workloads need fabric and storage chosen for the job. A provider that sells only what sits in stock forces the customer to adapt, and a generic configuration retrofitted afterward will not hold the ratio. Through Axe Compute, teams select NVIDIA Blackwell, Grace Blackwell, or Vera Rubin against the ratio a workload actually demands. Our comparison of Vera Rubin and Blackwell explains how that choice shifts as workload intensity rises.
Expertise: The CPU Layer Sized on Purpose
The under-provisioning risk is the reason cluster design, fabric, storage, and operations sit with the team that runs the fleet. A CPU layer sized by guesswork becomes the silent bottleneck behind an expensive GPU cluster. Virtualization compounds the loss, because it can skim performance off every GPU hour. For that reason, single-tenant bare metal is a design decision: the performance bought is the performance delivered, including on the orchestration path agentic workloads depend on.
Execution: One Plan for CPU, Network, and GPU
No team can correct a cluster built against the wrong ratio without a second buildout. Axe Build therefore plans the CPU, network, storage, and GPU specification together, working backward from the committed date. With server CPUs now in short supply, the CPU layer belongs in that plan from the first conversation, so procurement, facility readiness, and burn-in are scheduled before anything is signed.
Global Reach: Region in the Specification
Agentic systems call external services constantly, and residency rules govern where that traffic may land. Available capacity rarely lines up with those rules by accident, so region belongs in the specification at design time. With capacity across multiple regions, buyers can pick the region alongside the CPU ratio and the fabric, so agent loops stay inside the boundary the business is allowed to operate in.
Buying Capacity While the Ratio Is Still Moving
The ratio is moving inside live deployments right now, which creates a timing problem for buyers. Committing five years of infrastructure to a chatbot-era design is expensive. Waiting for the ratio to settle costs money too, because the workload is already in production.
Axe Access for Teams Still Measuring
Axe Access provides single-tenant GPU capacity, ready when the workload is, on terms from month to month up to 36 months. Shorter terms let a team measure steps per task on a live workload before it commits to a longer one. In addition, the customer receives a delivery date in writing at reservation, before committing.
Axe Build for Teams That Know Their Ratio
Once a team knows its ratio, Axe Build delivers a cluster built to its specification on a 36 or 60 month term, single-tenant, at a fixed monthly cost. The equipment sits on Axe Compute’s balance sheet, so the customer carries no residual hardware risk if the ratio shifts again. A mid-term GPU upgrade is also available where the agreement provides for it.
Both offerings run under a 99.93% uptime SLA, monitored continuously, with service credits. Pricing terms matter more than usual while the ratio moves. Zero CapEx means no hardware purchase and no facility commitment. Because agent loops move data between services constantly, zero egress fees also carry real weight, and zero hidden fees mean the number you agree to is the number you pay.
What Changes in the Year Ahead
NVIDIA, AMD, and Intel now describe the same shift from three directions, and Ciena describes the network demands that come with it. Any team running multi-step, tool-calling workloads on a cluster with a thin CPU layer therefore risks the same mismatch: accelerators that sit idle while orchestration catches up.
That mismatch is where teams will make the next round of infrastructure decisions. Teams that size CPU ratio, fabric behavior, storage, and orchestration against the workload will keep their clusters busy as tokens per task climb. A chatbot-era default carried forward, by contrast, surfaces the gap mid-deployment, when the hardware is already installed and a workload is already stalling on it.
Start with a measurement. First, instrument one agentic workflow in production. Next, count the steps per completed task and the CPU environments each task opens. Then size the CPU layer against those numbers before the next capacity conversation opens.
Design for the ratio, not the count.
Frequently Asked Questions
What is agentic AI and how is it different from chatbot-style AI?
Agentic AI refers to systems that plan, execute multi-step tasks, call external tools, and iterate toward a goal, rather than returning a single response to a single query. A chatbot query is one bounded exchange, while an agentic task can spawn many sub-steps, tool calls, and verification loops before it finishes.
Why does agentic AI need more compute per user?
Each agentic task can fan out into planning steps, tool calls, and validation loops, all of which add tokens and processing on top of the original request. J.P. Morgan Asset Management estimates that a single user could demand 10 to 100 times more compute with the latest agentic tools.
Why is the CPU-to-GPU ratio changing?
Orchestration, tool execution, and the sandboxed environments agents run code in all rely on CPUs. As agentic workloads add more of this work, AMD reports the ratio moving from roughly one CPU per four to eight GPUs toward 1:1, and in some cases higher on the CPU side.
Are chipmakers building CPU-only systems for AI agents?
Yes. NVIDIA launched Vera in May 2026 as a CPU built for AI agents, with a rack of up to 256 Vera CPUs supporting more than 22,500 concurrent CPU environments. In October 2026, AMD introduced Helion, which includes a CPU-only rack with up to 32,768 EPYC cores.
Does this mean GPUs matter less?
No. GPUs still perform the core model computation and remain the most capital-intensive part of the cluster. The shift means GPU count alone is an incomplete measure of capacity, because an under-provisioned CPU or network layer can leave GPUs idle while they wait on orchestration.
What happens if the network is not upgraded for agentic workloads?
Agentic workflows generate short, bursty, unpredictable traffic between agents, tools, and models, unlike the large, predictable batches typical of training. Ciena argues that networks need to evolve from passive interconnects into programmable, high-capacity infrastructure to keep pace.
How should a business plan capacity for agentic AI?
Capacity planning should size the CPU layer, network fabric, and orchestration software against expected workload intensity, rather than against a target GPU count alone. A practical starting point is to measure the steps per completed task on one live agentic workflow, then size the CPU layer against that number.
Is the CPU-to-GPU ratio shift already happening?
Yes. AMD reported 97 percent data center revenue growth in its third quarter of 2026 and expects server revenue to double in 2027. Intel reported 78 percent growth in data center and AI revenue, with demand exceeding available supply.
About Axe Compute
Axe Compute Inc. (NASDAQ: AGPU) is a neocloud AI infrastructure platform built on a fundamental premise: AI innovation should not be constrained by hardware choice or availability. The company provides enterprises and AI innovators with flexibility across hardware, geography, and deployment models through two core offerings: Axe Compute Access, delivering a wide range of the latest high-performance GPU infrastructure across global locations, and Axe Compute Build, enabling the design, deployment, ownership, and operation of large-scale, dedicated AI infrastructure worldwide. All solutions are supported by enterprise-grade SLAs and operational expertise. Axe Compute is headquartered in Pittsburgh, Pennsylvania. For more information, visit axecompute.com or contact us at info@axecompute.com.
Sources
- J.P. Morgan Asset Management, April 17, 2026, “Is AI running out of compute?”: a single user could demand 10 to 100 times more compute with the latest tools
- Quartz, May 14, 2026: AI agents can consume 1,000 times more tokens than a single chatbot query
- Ciena, March 25, 2026, “Agentic AI: Rewriting the rules of compute and networking”: 50 to 50,000 tokens; programmable, high-capacity networks
- AMD, May 7, 2026, “Agentic AI Changes the CPU/GPU Equation”: CPU-to-GPU ratio moving toward 1:1
- AMD first-quarter 2026 earnings call, May 5, 2026 (third-party transcript): Lisa Su on the CPU-to-GPU ratio
- AMD, November 3, 2026, third-quarter 2026 prepared remarks: data center revenue up 97% to $8.5 billion; server revenue expected to double in 2027; agents as the next killer app for CPUs; first Helion systems shipping this quarter
- AMD, October 12, 2026, “Introducing AMD Helion”: CPU-only rack with up to 32,768 EPYC cores
- NVIDIA, May 31, 2026, “NVIDIA Unveils Vera, the CPU for Agents”: 88 Olympus cores, up to 1.2 TB/s memory bandwidth
- NVIDIA Vera rack product page: up to 256 Vera CPUs and more than 22,500 concurrent CPU environments
- NVIDIA blog, 2026: first Vera CPU systems delivered to leading AI labs
- NVIDIA Vera CPU page: Vera Rubin NVL72 pairs 72 Rubin GPUs with 36 Vera CPUs
- Intel, October 29, 2026, third-quarter 2026 results: data center and AI revenue up 78% to $7.3 billion
- Intel, third-quarter 2026 prepared remarks: demand exceeding available supply; server CPU market could grow from $40 billion to more than $200 billion over five years
- Semiwiki, April 8, 2026, “Agentic AI Demands More Than GPUs” (summarizes an Intel white paper): CPU orchestration and under-provisioning