NVIDIA says its Vera CPU ends a choice AI data centers face with x86 hardware: a fast processor or a dense, efficient one. In a blog post from November 5, NVIDIA puts Vera at 1.6 times as many agents per watt as the fastest x86 CPU in a 100-megawatt data center comparison, and at twice the agent speed of the most energy-efficient x86 CPU.
NVIDIA’s pitch is that agents spend much of their time on the CPU, not the GPU
The argument starts from how an agent works. It moves between GPU reasoning and CPU execution, so every search, database query, code compile and test run lands on a processor, and NVIDIA expects power, not rack space, to be the limit. Its answer is one CPU that serves both ends of the x86 range. NVIDIA’s own Vera product page adds a memory claim, pairing LPDDR5X with detachable SOCAMM modules for five times the bandwidth per watt of memory power and three times the bandwidth per core of leading x86 CPUs with DDR5.
Partner results start at 1.5 times faster, and each one has a different baseline
NVIDIA lists results from early adopters. Perplexity ran agentic coding tasks 1.5 times faster than on traditional CPUs, DeepInfra reached up to 2.2 times the speed of the best x86 processor on its agent workload, and Modal reached up to twice the sandbox throughput per core. Daytona brought 1,000 agent sandboxes online in 10.8 seconds, against 88.4 seconds with x86. DuckLabs measured about 1.5 times the median TPC-H SF1000 result of an Intel Xeon 6 baseline, and ClickHouse’s standard binary scaled near-linearly across 176 cores and topped the ClickBench hardware leaderboard.
Only one comparison in the post names the chip it beats, so the figures cannot be lined up
That is the limit of the numbers. The post refers to “the fastest x86 CPU” and “the most energy-efficient x86 CPU” without naming either, and appears to give no core count, clock speed or power figure for the Vera part in the headline comparison, and does not describe how the 100-megawatt comparison was built. The product page gives up to 1.8 times on agentic sandbox work against “leading x86 CPUs” and marks its relative figures as subject to change, which is a different headline from the blog’s 2 times against the efficient part.
A named test shows what is missing. Intel’s agent-throughput post, which we covered in October, replayed one Terminal-Bench trace at 24-way concurrency for 60 minutes on Xeon 6, 5th Gen EPYC and Arm v9.2A instances, and put Xeon 6 at up to 1.66 times the EPYC one. That is a claim with a baseline a reader can identify. NVIDIA’s Vera claims have none yet, and neither vendor’s numbers say anything about the other’s chip.