Supermicro announced its Nvidia Vera CPU rack on 22 October 2026, adding a fully liquid-cooled system with 256 processors to its data-centre range. The company says the rack can supply the separate CPU environments in which AI agents run tools and carry out actions, either on its own or alongside its Vera Rubin NVL72 GPU system, which it says is shipping.
Key points
- Supermicro rates the 256-CPU rack for up to 22,500 concurrent agent sandboxes.
- Published specifications include up to 400TB of memory and a 208kW operating load.
- OpenShell runs on the CPUs, while Sentry monitors agent activity from BlueField-4 DPUs.
- The rack is available to quote, but Supermicro has not given a shipping date.
22,500 sandboxes on 256 Vera CPUs
Supermicro says the rack contains 22,528 CPU cores and 45,056 threads, and rates it for up to 22,500 agent sandboxes running at once. In the workload the company describes, GPUs generate text, code and queries, while CPU-based environments execute actions, assess results and run tools. An agent that branches into further tasks can therefore need more than one place to do its work, even when a GPU handles generation.
A sandbox is an isolated computing environment in which an agent can run code and use tools, Yahoo Finance reported. Think of separate workbenches: giving each job its own space helps keep the jobs apart, but the number of benches alone says little about how demanding each job can be. StorageReview reported that Supermicro does not state the memory allocation, workload or software configuration behind its sandbox rating, and described the figure as a sizing ceiling for lightweight sandboxes.
Running a software build could involve code executing in one of those separate environments, with other agent tasks running alongside it. The rack’s CPU capacity would make that arrangement possible at scale, while the published sandbox rating would not say how a particular build performs.
StorageReview also lists large-scale software builds, database operations, data analytics and cluster scheduling among Supermicro’s intended standalone uses. Nvidia’s Arm-based Vera processor enters work historically served by Intel Xeon and AMD EPYC, according to Yahoo Finance. AI Affairs previously reported Modal’s plan to use Vera after tests against Intel and AMD.
A 208kW rack beside Vera Rubin
Supermicro’s published configuration puts the CPUs in 32 compute trays, with eight processors in each, StorageReview reported. The 48U rack accommodates up to 400TB of SOCAMM2 LPDDR5X memory, with 300TB/s of aggregate memory bandwidth, and specifies up to 1.8TB/s of NVLink-C2C bandwidth per CPU. Each tray also carries four Nvidia BlueField-4 data processing units, or DPUs, for networking and security, while Spectrum-X Ethernet provides up to 400Gb/s of north-south rack connectivity.
StorageReview puts the rack’s operating power at 208kW. Its listed power options are three AC shelves arranged with redundancy, or an 800 VDC feed through a sidecar power rack. The sidecar can use battery backup units for protection during a power loss. For companies planning a deployment, those specifications put the electrical feed, cooling and network design alongside processor count as decisions about the system.
Supermicro says the rack uses Nvidia’s third-generation MGX modular design and its own DLC-2 direct liquid-cooling stack, sharing an infrastructure approach with Vera Rubin NVL72. It also offers a design pairing the CPU and GPU racks under a defined Spectrum-X network layout. The CPU rack can be deployed without the GPU system, according to the company.
Other system vendors are building around Vera. Dell announced its PowerEdge M9882 at GTC Berlin earlier in the same week, making Supermicro’s the second announced 256-CPU Vera rack, StorageReview reported. AI Affairs has also reported HPE’s addition of a 256-CPU Vera rack and agent controls to its Nvidia AI factory range.
OpenShell and Sentry divide agent controls
Supermicro says Nvidia OpenShell runs on the Vera CPUs and applies policies that limit an agent’s access while it is operating. Nvidia Sentry runs on the BlueField-4 DPUs, separate from the host CPUs, to monitor activity and enforce policy. The company says Sentry can quarantine an agent that crosses its permitted boundaries in milliseconds, and that placing monitoring on the DPUs avoids consuming host CPU resources for that work.
The arrangement gives the rack two places to act on an agent’s permissions: where its work runs, and on separate hardware watching that work. AI Affairs previously covered Nvidia’s OpenShell architecture for governing agents on enterprise systems. Supermicro’s announcement describes the software’s roles in this rack.
Supermicro’s Vera range also includes air-cooled 1U and 2U systems, the company says. The liquid-cooled rack is available to quote, but Supermicro has not given a shipping date.
