Chipmakers pitch agent CPUs, and buyers need workload tests

Chipmakers have found a fresh label for server CPUs: agentic AI. Arm, Nvidia and AWS now tie standard datacenter processors to AI agents, while Intel and AMD make the same pitch around x86 racks.

The label gives buyers a poor guide. AI agents use CPUs, but agents do not create one clean CPU workload. They call models, pass data to tools, run business software, wait on networks, write files, parse text, query databases and trigger code that companies have run for years.

A team that buys an “agent CPU” without testing its own agent stack risks paying for the wrong strength. One deployment may need memory bandwidth. Another may need cache. A third may care about single-thread latency because the GPU beside it costs too much to leave idle.

Nvidia CEO Jensen Huang calls Vera a “CPU for agents.” Arm calls its datacenter chip the “AGI CPU.” AWS filled its Graviton messaging with agentic AI language. Intel and AMD have brought agent workloads into their server CPU sales pitch.

Buyers should treat those names as marketing. Engineers should measure the code path that their agents execute.

Chipmakers sell a workload story

Nvidia built Vera around fast CPU work beside GPUs. Huang told the audience at GTC Taiwan that agents will use CPUs with little patience because GPUs cost too much to sit idle. That argument fits Nvidia’s business. Nvidia sells systems that need the CPU and GPU to exchange data with less delay.

Vera follows the same idea Nvidia used with Grace. The CPU sits close to the accelerator, moves data across the system and feeds the expensive part of the box. Nvidia’s Grace CPU pitch centers on bandwidth, energy use and work beside accelerated computing.

Arm took a different route with its AGI CPU message. Arm’s design points to Neoverse V3 cores, high memory bandwidth and a stripped-down feature set. Arm removed features that many agent runs may not need, including simultaneous multithreading and heavy vector hardware, according to the article.

AWS pushed Graviton 5 with a similar frame. Amazon described a 192-core Arm server CPU for cloud workloads and agentic AI. That fits AWS’s long-running Graviton strategy: give cloud customers more cores and better power use under Amazon’s own control.

Intel and AMD answer with scale. Intel showed reference racks with tens of thousands of x86 cores. AMD argued that large agent fleets need throughput, so buyers should compare how many tasks a rack completes inside a power budget.

Agents do many kinds of CPU work

An AI agent connects a model to tools. That sounds tidy in a slide deck. In production, the CPU handles messy work around the model.

A customer support agent may query a database, classify a ticket, search a knowledge base and update a customer record. A coding agent may run tests, read files, invoke compilers and handle source control. A security agent may inspect logs, decompress data and call several services before it hands context to a model.

Those examples stress different CPU traits. Database-heavy agents may care about cache and memory latency. Build agents may care about frequency and compiler throughput. Log agents may benefit from compression engines and memory bandwidth. Large fleets may prize core count and power use.

That spread explains why AMD and Intel sell many Epyc and Xeon models. Buyers tune for clocks, cores, cache, power and platform features because one CPU cannot fit every server job.

Agent software does not erase that trade-off. It adds orchestration around known work.

Latency and throughput point to different buys

Nvidia wants buyers to look at latency. The company sells GPU systems, and GPU time costs money. If a CPU waits too long to feed a GPU, the whole system wastes cash.

That case makes sense for certain agent stacks. A real-time assistant that calls a model, asks a tool for data and returns an answer to a user may need low delay across each step. In that setup, memory bandwidth, interconnect speed and single-thread performance can matter more than raw core count.

AMD wants buyers to look at concurrency. That case also makes sense. A business that runs thousands of agents in the background may value completed jobs per rack, not the fastest answer from one agent.

AMD has said its 256-core Venice Epyc chips would produce more throughput per 100 kW rack than Vera, according to the article. That claim pushes the debate toward density: how much agent work can a datacenter finish inside a fixed power limit?

Both metrics matter. The workload decides which one matters more.

Benchmarks need scrutiny

Phoronix tested an Nvidia Vera CPU sample with a subset of its suite that Nvidia saw as representative, according to the article. Vera beat AMD’s 128-core Epyc 9575F by 10% on a geometric mean score and beat Intel’s 128-core Xeon 6980P by 55%.

Those numbers give Nvidia a strong headline. They do not settle the CPU choice for agent deployments.

A geometric mean can hide outliers. One benchmark may reward memory bandwidth. Another may reward clocks. Another may reward cache size. A buyer needs the per-test spread, not the blended score.

Engineering teams should build a benchmark from their own agent traces. They should include tool calls, database waits, serialization, vector search, file work and model handoff. Synthetic tests help only when they match those steps.

The compliance angle for AI infrastructure buyers

AI agents can touch personal data, account records and internal documents. That makes the CPU choice part of a broader governance problem, even though privacy laws do not regulate CPUs by brand.

Under the GDPR, companies that process personal data must limit access, protect data and prove a lawful basis for processing. Under the California Consumer Privacy Act, companies must honor consumer privacy rights and disclose how they handle covered data.

Agent systems can make those duties harder. An agent may call tools across several services, copy personal data into prompts and write outputs into logs. A faster CPU can increase scale, but scale also increases the number of records a faulty agent can expose.

Regulators can impose large penalties for poor controls. GDPR fines can reach 4% of annual global revenue or 20 million euros, whichever amount runs higher. CCPA enforcement can bring civil penalties for violations, and private plaintiffs can sue after certain data breaches.

Infrastructure teams should connect performance tests with compliance tests. They should log which tools agents call, restrict data access by role, redact personal data from prompts and set retention limits for traces. They should also measure whether CPU changes alter logging volume, data movement or access patterns.

Marketing labels cannot replace workload evidence

Chipmakers will keep attaching AI language to CPUs because buyers have AI budgets. That does not make the silicon fake. Vera, Graviton, Neoverse, Xeon and Epyc all solve valid server problems.

The problem sits in the label. “Agentic CPU” suggests one workload with one best chip. Production agents behave like software glue across many systems, so the CPU must match the tools the agent uses.

A buyer should ask four plain questions before signing a purchase order:

• Does the agent wait on GPUs, databases, networks or local compute?
• Does the agent fleet need lower latency or more completed jobs per rack?
• Does the benchmark include real tool calls and real data sizes?
• Does the system keep privacy logs, access controls and retention rules intact under load?

Vendors can help answer those questions with hardware counters, reproducible tests and clear power data. Buyers should demand that evidence before they accept an AI-branded CPU pitch.

The useful conclusion comes from workload math. Agents need CPUs because all server software needs CPUs. The right CPU depends on the agent’s toolchain, data path and compliance burden.