Analysis
At Hot Chips 2026 -- the annual chip-architecture conference held August 23-25 at Stanford -- Nvidia disclosed the internal architecture of its custom "Vera" CPU in detail for the first time, Tom's Hardware reported. The chip packs 88 custom Arm-based "Olympus" cores, uses a technique Nvidia calls spatial multithreading to improve how the cores share work, and pairs with 1.2 TB/s of SOCAMM2 memory bandwidth -- a meaningfully faster memory interface than the LPDDR5X Nvidia used in its prior Grace CPU generation. Nvidia is claiming roughly 1.8x faster performance on agentic AI workloads specifically, a category (multi-step reasoning, tool calls, long context) that stresses CPU-GPU coordination differently than a single large training run does.
Why Nvidia Is Building Its Own CPU
Vera isn't Nvidia's first custom CPU -- the company shipped Grace, its first Arm-based server CPU, starting in 2023, positioning it as the CPU half of the Grace Hopper and Grace Blackwell superchip pairings. What's new with Vera is both the scale of disclosure (Nvidia rarely details this much CPU microarchitecture publicly) and its role in Nvidia's next full rack-scale system, Vera Rubin NVL72 -- the successor to the Blackwell/GB200 platform that currently anchors most large AI-training deployments. Building its own CPU, rather than pairing its GPUs with merchant silicon from Intel or AMD, gives Nvidia tighter control over the interconnect between CPU and GPU and reduces how much of its own server economics depend on a separate chip vendor's roadmap and pricing.
The Competitive Field
- AMD -- its EPYC server CPU line is the most direct merchant-silicon comparison; ServeTheHome's benchmarking found Vera outpacing AMD's EPYC 9655P specifically on Linux kernel compilation, an early third-party data point beyond Nvidia's own marketing claims
- Intel -- Xeon remains the incumbent data-center CPU standard Nvidia's custom silicon push is gradually displacing in AI-server configurations
- Google (Axion), Amazon (Graviton), Microsoft (Cobalt) -- hyperscalers building their own Arm-based custom CPUs for the same reason Nvidia is: reducing dependence on merchant silicon and tuning hardware to their own workloads
- Nvidia Grace -- Vera's direct predecessor, still shipping in current Grace Hopper and Grace Blackwell systems
The pattern across all of these is the same: every major AI-infrastructure player with the scale to justify it is now designing custom CPU silicon rather than buying off the shelf, and Nvidia joining that list on the CPU side -- after already dominating the GPU side -- is a meaningfully different competitive position than Google, Amazon or Microsoft occupy, since Nvidia's CPU investment reinforces a platform it already controls end to end rather than diversifying away from one.
The Numbers in Context
An 88-core count and 1.2 TB/s of memory bandwidth are large by current server-CPU standards, but the more relevant comparison for Pulse's audience isn't the microarchitecture spec sheet -- it's what Vera signals about Nvidia's broader infrastructure strategy, which Pulse has tracked extending well beyond chip design into financing the AI buildout directly. Nvidia is separately assembling a $500 billion financing pool with BlackRock, KKR and Apollo to help customers afford its hardware; a custom CPU that improves agentic-workload performance by a claimed 1.8x is the product-side lever pulling in the same direction as that financing effort -- both are aimed at keeping AI compute economics favorable to Nvidia as competition from custom hyperscaler silicon intensifies.
The Counterweight
The honest limitation: Nvidia's 1.8x agentic-workload figure is the company's own claim, not an independently verified benchmark, and the only third-party data point disclosed so far -- Vera beating one specific AMD EPYC part on Linux kernel compilation -- is a narrow, non-AI workload that doesn't directly validate the agentic-performance claim. Hot Chips disclosures are also architecture reveals, not shipping-product announcements; Vera Rubin NVL72's actual production timeline and real-world availability haven't been detailed with the same specificity as the chip's internals.
Independent agentic-workload benchmarks against Vera, once the chip and Vera Rubin systems actually ship, will be the real test of Nvidia's 1.8x claim -- Hot Chips disclosures routinely undersell or overstate real-world gains once hardware reaches customers running production workloads rather than vendor-selected demos.
