NVIDIA is turning up the heat on traditional data center CPUs. On Tuesday, the company rolled out full specs and new benchmarks for its Vera processor, pitching it as a purpose-built engine for agentic AI and orchestration workloads. Vera will ship both as a standalone server CPU and inside dense, liquid-cooled racks, and Nvidia says early hardware is already in the hands of labs including OpenAI, Anthropic and SpaceX.
In technical notes on NVIDIA's developer blog, the company details an 88-core custom Olympus design with NVIDIA Spatial Multithreading and a second-generation LPDDR5X memory subsystem that delivers up to 1.2 TB/s of bandwidth and as much as 1.5 TB of capacity per socket. NVIDIA is positioning that combo as roughly a 50% uplift in sandbox performance for agentic workloads compared with rival x86 platforms, while also improving efficiency under heavy concurrency. The documentation also describes a monolithic compute die and attachable SOCAMM memory modules meant to make LPDDR5X practical and serviceable in the data center.
Vera Rubin and system options
NVIDIA plans to ship Vera in a range of system footprints, from single- and dual-socket air-cooled servers to tightly integrated rack-scale platforms that pair the CPU with Rubin GPUs. As outlined by Tom's Hardware, the Vera Rubin NVL72 design connects CPUs and GPUs over high-bandwidth NVLink chip-to-chip interconnects, and Nvidia is building dense, liquid-cooled racks that pack large numbers of Vera processors to support tens of thousands of isolated agent sandboxes. The message to operators is clear: think in terms of AI factories at the rack level, not just individual pizza-box servers.
Pricing, power and market math
Wall Street is already reworking its spreadsheets around Nvidia’s CPU ambitions. According to CNBC, Wolfe Research is modeling roughly 1.3 million Vera units shipping this year at an average selling price near $5,000 per chip, while Bernstein pegs the mature server CPU market at about $37 billion in 2025, numbers that help explain Nvidia’s view that the broader CPU opportunity could eventually reach into the hundreds of billions. Vendor guidance and early partner data show Vera parts drawing in the low-hundreds to mid-hundreds of watts, depending on configuration, roughly a 250–450W envelope per chip.
Who's already buying
The sales cycle for Vera appears to be moving faster than a standard science experiment. Data Center Knowledge reports that Nvidia says Vera Rubin and Vera CPU systems are already in production, with labs and cloud providers including OpenAI, Anthropic and SpaceX among the first wave of testers. NVIDIA and its OEM partners say commercial systems from major vendors are slated to roll out more broadly later this year.
What this means for Intel and AMD
Vera’s architecture, with its focus on per-core speed, tight latency and much higher memory bandwidth per socket, is squarely aimed at new classes of agentic AI workloads that are expected to increase CPU demand rather than simply shifting it around. That puts fresh competitive pressure on the x86 incumbents, Intel and AMD. Analysts and hardware watchers such as Tom's Hardware flag the strategic implications but note that eventual market share will depend heavily on supply, software portability and how quickly hyperscalers and cloud providers are willing to shift more of their stacks to Arm-based CPUs instead of x86 designs.
From here, the usual bottlenecks will decide how fast Vera can move from elite labs into more ordinary racks: high-density LPDDR5X SOCAMM module production, liquid-cooling buildouts and board-level component supply. NVIDIA says Vera is already in full production and is scheduled to be available from partners in the second half of 2026. The next several quarters will reveal whether Vera graduates into a mainstream host CPU for AI factories or stays a high-end tool reserved for frontier labs and the biggest hyperscalers.
