Liquid cooling in AI data centers has mostly been a GPU story. Supermicro's newly published design for an NVIDIA Vera CPU rack adds a second chapter. The company describes a 48U rack built on NVIDIA's MGX architecture that holds 256 NVIDIA Vera CPUs and is, in Supermicro's words, 100% direct liquid-cooled and fully factory-integrated. The target workloads are agentic-AI sandbox execution and high-performance computing.
For operators, the main point is not the core count. NVIDIA presents the rack as something that sits next to its GPU systems, not in place of them. On its Vera CPU Rack product page, the company says the rack is designed to scale CPU capacity alongside Vera Rubin NVL72. In other words, the CPU racks in an AI hall may need the same coolant loops as the accelerator racks.
What is in the rack
According to Supermicro, each rack contains:
- 256 NVIDIA Vera CPUs, totaling 22,528 cores and 45,056 threads
- Up to 300 TB/s of aggregate memory bandwidth
- Up to 400 TB of memory
- NVIDIA BlueField-4 DPUs and Spectrum-X Ethernet networking
- Support for more than 22,500 concurrent CPU environments
NVIDIA's product page uses matching but slightly looser numbers: "up to 256 Vera CPUs" per liquid-cooled rack and "more than 22,500 concurrent CPU environments." It also lists BlueField-4 DPUs and Spectrum-X Ethernet.
heise online's analysis of the processor, from its GTC coverage, explains the core math. Each Vera chip has 88 Olympus cores based on the ARMv9.2 architecture, so 256 processors give 22,528 cores and 45,056 threads. heise reports that each Vera CPU uses eight SOCAMM2 LPDDR5X memory modules, for about 1.2 TB/s of bandwidth and up to 1.5 TB of capacity per CPU.
Multiplying those per-chip figures out gives totals close to Supermicro's rack numbers. 256 CPUs at about 1.2 TB/s each is roughly 307 TB/s, in line with Supermicro's "up to 300 TB/s." 256 CPUs at 1.5 TB each is 384 TB, a bit under Supermicro's "up to 400 TB." None of the sources explains that gap, which may come down to rounding or how capacity is counted.
Why a CPU rack for AI?
Both vendors pitch the rack for agentic AI. In these systems, AI models do more than generate text: they run code, call tools and run tasks in isolated software environments. That is why both companies cite the number of concurrent CPU environments a rack can host, more than 22,500, as a main metric next to core count.
The performance claims come from the vendors. Supermicro says the rack is up to 1.8x faster than leading x86 processors on key agentic workloads. It also claims twice the memory bandwidth at half the power compared with x86 systems using DDR5. NVIDIA's page gives the same up-to-1.8x figure for compiling generated code, running Python tool chains and analyzing software code, and notes that the figure is based on measured data, baselined against a latest-generation x86 CPU and subject to change. No independent benchmarks appear in the sources reviewed for this article.
How it compares on density
heise adds a caveat. A Vera rack does not have the most cores a rack can hold. heise notes that x86 racks can exceed 22,528 cores: a rack of 160 AMD Epyc 9965 processors, for example, reaches 30,720 cores. The Vera case rests on memory bandwidth per CPU, integration with NVIDIA's networking and DPUs, and NVIDIA's workload claims, not on raw core count.
That matters for anyone planning capacity. Core count, environment count and memory bandwidth are three different measures. Which one matters most depends on the workload a tenant actually runs.
The cooling and facility angle
Supermicro says the rack uses its DLC-2 direct liquid cooling technology. It is delivered through the company's Data Center Building Block Solutions (DCBBS) approach and arrives fully integrated from the factory. The company calls the design 100% direct liquid-cooled and says DLC-2 eliminates chilled-air overhead entirely.
The sources do not give rack power draw, coolant flow rates, supply temperatures or facility water figures for the Vera rack. Operators should not assume any specific per-rack load until vendors publish those specifications. The design direction is still clear. If CPU racks built to sit next to Vera Rubin NVL72 systems are also fully liquid-cooled, a larger share of an AI hall's IT load could need liquid cooling. That would affect how coolant distribution and heat rejection are sized for new halls and retrofits.
Supermicro describes DCBBS as covering site survey, project design, rack integration, solution validation and ongoing support, so part of the integration work moves from the data hall floor to the vendor's process.
Who is building and buying
According to heise, Dell, HPE, Lenovo and Supermicro all plan to sell Vera-based servers, using NVIDIA's modular MGX system. Customers named in heise's coverage include Meta, Alibaba, Cloudflare, Nebius and OCI. The sources do not give order volumes, deployment sites or deployment timelines for any of these customers.
What Happens Next
heise reports that NVIDIA announced the Vera CPU Rack at GTC for availability in the second half of 2026. Supermicro has published its rack design, and NVIDIA's product page lists the rack configuration. None of the sources reviewed here gives a specific ship date, pricing, or named customer deployment for the Supermicro system. According to heise, Dell, HPE and Lenovo also plan Vera servers, but the sources do not schedule their releases, and they include no independent benchmarks of the 1.8x agentic-workload claim.
