Nvidia has unveiled a few more details about its now-shipping Vera CPU, announced earlier this year, which is clearly differentiated in a crowded market, with more efficiency and better performance, especially for agentic AI workloads. But the real surprise is that the CPU is gaining traction in other areas besides agents, due to its excellent memory bandwidth, power efficiency and single-core performance. Let’s take a deeper look. (Note that Nvidia, like many AI semiconductor companies, is a client of my firm, Cambrian-AI Research.)
Vera CPU Boasts 2X Faster Core Performance to Keep GPUs Busy
As most everyone now realizes, CPUs are back in fashion thanks to the rise of agentic AI, which needs fast and efficient CPU cores and memory. In agentic workloads, CPUs plan and execute a series of sequential steps to complete a task in a loop which can execute hundreds of times. Multiple cores don’t help as much as they do for traditional cloud workloads. What matters are fast cores for agentic speed and latency, and efficient memory that can keep up with the cores.
The graph below shows the execution of a 10-step agentic loop, and you can see how the expensive GPUs are waiting for the agent (grey bars) to finish each step. Faster CPUs mean faster results, and less latency improves overall data center efficiency. And that is where the rubber meets the road as inference becomes a profit center for clouds and enterprises.
Anticipating the impact that slower CPUs have in the agentic work, Nvidia decided to re-engineer its CPU line (Grace) to have fast single-threaded core performance which communicate efficiently without crossing chiplet boundaries which incur a performance hit. The results are appealing as shown below. More benchmarks are needed to fully evaluate this platform, and Nvidia promises they will be forthcoming. Much of the advantage that Vera brings to market is the LPDDR5 memory, which was originally designed for mobile phones and delivers
The Vera CPU Business Case
Nvidia could ship four to five million Vera CPUs in the 2nd half of 2026, compared with 2.5 million Grace CPUs shipped to date. BofA estimates that could amount to $20B in revenue in 2026, and perhaps as much as half of that could be stand-alone CPUs. Public reporting from SemiAnalysis cited by Reuters and other outlets indicates Nvidia is quoting Vera at “well north of $20,000” per processor prior to volume discounts. A fully configured rack with 256 Vera chips will cost “around $10 million,” implying an average effective price in the $30–40k range once you include memory and system costs. Nvidia estimates the Total Addressable Market (TAM) for Rubin to be a $200B opportunity.
Vera CPU Application Examples Include HPC, not just AI Agents
Nvidia went on to sharing a few customer stories around Vera. The first is not a surprise: Perplexity, in whom Nvidia has invested, is seeing 1.5-1.9X performance for Vera compared to x86 CPUs in agentic AI.
But there are some surprises here beyond agentic AI workloads which could allow Nvidia to carve more deeply into the Intel and AMD x86 CPU business. The NYSE is seeing a six-fold improvement in P99 latency, demonstrating the consistency that can be achieved with the Vera CPU monolithic architecture. “P99 latency” refers to the 99th percentile of response times in a system. In practical terms, it answers the question: How slow are the slowest 1% of requests?
And in scientific simulation, the faster per-core performance and memory bandwidth are allowing Los Alamos National Labs to realize three to seven times better performance in their Veritas supercomputer, which uses Vera CPUs and Rubin GPUs.
Seeing the opportunity Vera promises, Nvidia has been able to build an ecosystem of customers, Supercomputing Centers, cloud providers, and system OEMs that could enable a significant upside to Nvidia’s existing GPU-centric data center business.
Vera CPU’s Strategic Impact on Nvidia
Vera finally allows Nvidia to claim a significant technology lead in every segment of the data center compute business, with top-shelf differentiated platforms including rack-scale systems, CPUs, GPUs, LPUs and networking of all scales (up, out, and across). Consequently, the $200B server CPU TAM is now open to Nvidia to address, and they have a strong hand to play.
Disclosures: This article expresses the opinions of the author and is not to be taken as advice to purchase from or invest in the companies mentioned. My firm, Cambrian-AI Research, is fortunate to have many semiconductor firms as our clients, including Baya Systems BrainChip, Cadence, Cerebras Systems, D-Matrix, Flex, Groq, IBM, Infleqtion, Intel, Micron, NVIDIA, Qualcomm, SImA.ai, Synopsys, Taalas, Tenstorrent, Ventana Microsystems, and scores of investors. I have no investment positions in any of the companies mentioned in this article. For more information, please visit our website at https://cambrian-AI.com.
