Agentic AI workloads are refusing to follow a predictable path, and that changes the hardware equation. Telemetry from 163,594 agentic sessions reveals a level of variation that can make traditional CPU fleet planning impractical, while NVIDIA Vera targets the challenge with up to 1.5x the per-core performance of the latest AMD Venice CPUs.
The message is clear: AI systems that act through changing sequences need processors built for workload diversity, not just familiar averages. At the same time, MLCommons has released MLPerf Client v2.0, bringing new AI categories into benchmarking and adding another way to measure how computing hardware handles modern AI tasks.
Agentic Workloads Break the Old Fleet Model
Agentic AI does not produce one fixed workload pattern. Each session can move through a different trajectory, creating a difficult target for teams planning CPU fleets for AI factories. The data from more than 163,000 sessions puts a number on that challenge: more than 97% show unique trajectory profiles.
Eduardo Alvarez and Praveen Menon describe the pattern in direct terms: “Real-world telemetry from over 163,000 agentic sessions shows highly variable and unpredictable workload trajectories, with more than 97% of sessions exhibiting unique profiles, which makes traditional multi-design CPU fleet strategies impractical for AI factories.”
That finding turns workload unpredictability into a hardware design problem. A fleet built around multiple CPU designs must account for many different agentic behaviors, yet the session data shows that shared patterns are rare. When nearly every session brings a unique profile, choosing processors for one expected path leaves the fleet facing many others.
The result is a strong push toward a CPU that can serve a broad range of agentic workloads. NVIDIA Vera, a CPU designed for agentic AI workloads, enters this picture as a response to the diversity shown in the telemetry. Its role is not tied to one narrow session pattern; the central claim focuses on per-core agentic workload performance across this changing environment.
NVIDIA Vera Targets Per-Core Agentic Performance
NVIDIA says the Vera CPU achieves up to 1.5x the per-core agentic workload performance of the latest AMD Venice CPUs. That figure gives the hardware story a clear point of comparison: performance per core becomes a key measure when agentic sessions vary so widely.
“The NVIDIA Vera CPU achieves up to 1.5x the per-core agentic workload performance of the latest AMD Venice CPUs,” Alvarez and Menon wrote. The claim connects Vera directly to the fleet challenge, presenting a processor designed around agentic AI rather than a general performance figure detached from the workload.
Per-core performance matters in this context because the workload profiles do not settle into one repeatable shape. The 163,594-session dataset shows that agentic activity can follow unpredictable trajectories, and the 1.5x figure positions Vera as a way to address that variation through CPU performance.
The comparison also sharpens the competitive picture for AI hardware. AMD is part of the benchmark point through its latest Venice CPUs, while NVIDIA is positioning Vera against that hardware for agentic workloads. The central question is no longer only how much computing a fleet contains, but how well its CPUs handle the many paths that agentic sessions can take.
MLPerf Client v2.0 Expands the Measurement Picture
Hardware claims need benchmarks that reflect the work AI systems perform, and MLCommons has now released MLPerf Client v2.0. MLCommons describes itself as “an open engineering consortium dedicated to improving machine learning performance and transparency,” and its new benchmark includes new AI categories.
The release adds a second major development to the agentic CPU story. NVIDIA Vera addresses a hardware challenge revealed by real-world session telemetry, while MLPerf Client v2.0 expands the measurement framework used for AI performance. Together, the developments point toward a computing landscape where workload-specific results carry more weight.
MLCommons developed MLPerf Client v2.0 with collaboration from top PC OEMs. The benchmark release also involves technology companies including AMD, Intel, Microsoft, NVIDIA, and Qualcomm Technologies, Inc. That group brings multiple hardware and software perspectives into a benchmark designed to improve machine learning performance and transparency.
New AI categories give the benchmark a broader role as AI workloads continue to take different forms. In the agentic case, the session data shows why a single simple workload cannot represent the full challenge: 97% of sessions display unique trajectory profiles, so measurements must connect performance to the work AI systems actually perform.
A Faster Path for Agentic AI Fleets
The combination of telemetry, CPU design, and benchmarking creates a sharper roadmap for AI factories. First, real-world data exposes the workload problem. Next, NVIDIA Vera offers a per-core performance claim aimed at that problem. MLPerf Client v2.0 then adds new AI categories that can help frame performance discussions across the broader computing market.
Those pieces do not erase the unpredictability of agentic sessions, but they make the challenge easier to describe. The numbers are hard to miss: 163,594 sessions, more than 97% with unique profiles, and up to 1.5x the per-core performance claimed for NVIDIA Vera against the latest AMD Venice CPUs.
As agentic AI workloads keep demanding hardware that can handle changing trajectories, CPU strategy will move closer to the center of AI infrastructure planning. NVIDIA Vera and MLPerf Client v2.0 mark two connected developments: one aims at the processor, and the other strengthens the way AI performance gets measured.
The next stage will belong to systems that match flexible hardware with measurements built for flexible workloads. Agentic AI has shown that predictable fleet assumptions no longer fit every session, and the race to build, test, and scale around that reality is already underway.
Based on
