NVIDIA says its Vera CPU is the fastest way to run AI agents, and the most efficient. In a new benchmark report, the company claims one Vera chip beats the fastest x86 processor on agents per watt and doubles the agent speed of the most efficient x86 chip. Early partners report the same pattern in production-style tests, and Perplexity already plans to deploy Vera in a coming production system.
What NVIDIA measured
The comparison uses a 100-megawatt data center, where power limits how many agents can run. Against that constraint, NVIDIA says Vera delivers 1.6 times more agents per watt than the fastest x86 CPU, and twice the agent speed of the most energy-efficient x86 chip. One chip covers both jobs instead of mixing fast and dense x86 parts.
Agents are unusually hard on CPUs. They search for information, query databases, write and test code and take action, moving between GPU reasoning and CPU execution at every step. As companies put more agents to work, those in-between steps decide how fast a job finishes, and a data center that wastes power on them can run fewer agents overall.
What the partners found
Partners tested Vera on their own workloads and reported faster results across the board. Perplexity ran agentic coding tasks 1.5 times faster than on traditional CPUs. DeepInfra measured up to 2.2 times the speed of the best x86 chip on agent orchestration. Daytona brought 1,000 agent sandboxes online in 10.8 seconds, against 88.4 seconds on x86 hardware.
The database tests agreed. ClickHouse ran its software untuned and topped its benchmark leaderboard, scaling near-linearly across 176 cores. DuckDB ran about 1.5 times faster than an Intel Xeon 6 baseline, and Modal saw twice the sandbox throughput per core. Vera is the host CPU in NVIDIA’s Vera Rubin systems, the platform behind this year’s biggest AI infrastructure deals.
What NVIDIA did not say matters too. The post gives no price, no ship date and no general-availability window for Vera, and the partner results are early tests rather than independent benchmarks. Perplexity’s plan to deploy Vera in a production system is the strongest signal here, because a customer committing real workloads is worth more than any vendor chart.
Our read
The numbers are impressive, and they are all NVIDIA’s own. The 100-megawatt comparison, the partner tests and the leaderboard results come from the company selling the chip, so independent benchmarks will matter more than this launch post. The pattern is still hard to ignore: five different partners, on five different workloads, all reported the same direction of travel.
Vera shows how NVIDIA’s ambitions have widened. The company built its empire on GPUs, then moved into networking and software; a CPU beating x86 at agent work would let it sell the whole AI factory, not just the accelerators. Intel and AMD will answer in the same currency: agents per watt, not just raw speed.
Why It Matters
Agent workloads are becoming the measure for data centers. Models get the headlines, but agents do the searching, coding and database work, and they run all day. A CPU serving more agents per watt cuts the power bill for every deployment, which is why NVIDIA framed Vera around a 100-megawatt building, not a single score.
The timing also fits a market hungry for NVIDIA hardware. Chinese firms have ordered billions of dollars’ worth of export-compliant NVIDIA cards this year, and demand for AI capacity keeps outrunning supply. If Vera ships as claimed, it gives NVIDIA a second product line to sell into the same buildout, one aimed at the CPU-shaped gap its GPUs never filled.
Background
NVIDIA unveiled Vera in March as its first purpose-built data center CPU, built on its custom Olympus core, and the chip now anchors its Vera Rubin systems. Microsoft recently walked away from expanding a Texas data center deal, a reminder that even in a boom, buyers are choosing capacity carefully.
Source: NVIDIA’s Vera announcement and Unite.AI’s report
0 Comments