TL;DR · 30-second read
The Short Version
The newest AI systems don’t just answer questions. They act as “agents” that carry out tasks, such as writing and testing software. Each task needs its own walled-off workspace running on ordinary computer processors.
Server maker Supermicro just launched a single cabinet holding 256 of NVIDIA’s new processors. It can host about 22,500 of those workspaces at once, and it is cooled by liquid piped through the cabinet instead of by fans.
Why it matters: data centers building for AI may now need that plumbing across more of the building, not only in the rooms full of AI chips.
Super Micro Computer (Nasdaq: SMCI) on October 23, 2026 announced the NVIDIA Vera CPU rack, a fully liquid-cooled rack holding 256 NVIDIA Vera CPUs, which adds up to 22,528 cores and 45,056 threads. Supermicro says one rack supports up to 22,500 simultaneous agent sandbox environments. The rack pairs the CPUs with NVIDIA BlueField-4 DPUs, uses NVIDIA’s third-generation MGX rack design and Supermicro’s DLC-2 direct liquid cooling stack, and is available now for quotation.
The rack can run on its own or alongside Supermicro’s Vera Rubin NVL72 GPU racks, which the company says are now shipping. The launch comes three days after an 8-K filed October 20, in which Supermicro estimated fiscal first-quarter revenue of $18.5 billion to $19.5 billion, ahead of its $14.5 billion to $15.5 billion guidance, and attributed the beat primarily to increased demand for AI platforms.
Executive Summary
Supermicro has added a CPU-only rack to its Data Center Building Block Solutions (DCBBS) portfolio. The rack is built for the part of agentic AI that GPUs do not handle: running the actions an AI agent decides to take, such as executing code, calling tools, checking results and orchestrating multi-step workflows. Each of these runs in a sandbox, an isolated software environment that keeps an agent’s actions contained. Supermicro rates one rack at up to 22,500 concurrent sandboxes, roughly one per CPU core.
The more consequential detail for facility planners is how the rack is built. It is fully liquid-cooled, it shares NVIDIA’s third-generation MGX rack architecture with Vera Rubin NVL72, and Supermicro offers a combined layout in which both rack types sit in rows served by redundant coolant distribution units and rear-door heat exchangers. In that design the CPU tier of an AI cluster draws on the same liquid-cooling and power infrastructure as the GPU tier. It does not sit in a conventional air-cooled hall next door.
Supermicro still sells air-cooled 1U and 2U Vera servers, so liquid is an option rather than a requirement. But the rack-scale product, and the reference design that pairs it with NVL72, assumes that the CPUs serving AI agents live on the liquid loop.
Why Agents Need a CPU Tier of Their Own
Supermicro’s case for the product rests on a description of how agentic AI differs from chatbot-style inference. A conventional inference request is a straight line: a prompt goes in and generated text comes out. Agentic and reinforcement learning workloads branch and loop. A model proposes an action, something executes it, the result is evaluated, and the model tries again. In the company’s framing, GPUs still do the heavy lifting of generating tokens, code and queries. CPU-based sandboxes execute the actions, run the tools and orchestrate the workflow. Reinforcement learning here means training a model by letting it attempt tasks and scoring the outcomes.
The arithmetic in the announcement is consistent with that picture. 256 CPUs delivering 22,528 cores works out to 88 cores per Vera CPU, and 45,056 threads means two threads per core. The 22,500-sandbox ceiling is therefore close to one sandbox per physical core. That is a plausible density for lightweight, isolated execution environments. It is also an “up to” figure whose real value depends on how much memory, storage and compute each agent’s sandbox consumes, and Supermicro has not defined that.
NVIDIA’s Chris Marriott made the same argument in the announcement: supporting thousands of concurrent agent workflows requires “a new class” of high-density CPU infrastructure. The claim is directional rather than quantified. Neither company has published a recommended ratio of CPU racks to GPU racks, and that ratio is the number that would tell an operator how much extra floor space agentic workloads actually add.
The Liquid Loop Now Extends Past the GPU Hall
This is where the announcement matters for people who build and run data centers. Liquid cooling has so far been justified by GPU density. Racks of AI accelerators produce too much heat for air alone, so operators have plumbed coolant to those rows. Supermicro’s Vera CPU rack is described as 100% liquid-cooled through its DLC-2 direct liquid cooling stack, in which cold plates sit directly on the chips. It is built on the same third-generation MGX architecture as Vera Rubin NVL72. Supermicro presents that shared architecture as a way to standardize deployment.
The optional combined design spells out the facility consequence. Vera Rubin NVL72 racks and Vera CPU racks are laid out together with hot-aisle containment, which walls off the hot exhaust side of the racks so heat does not mix back into the room. The rows are served by redundant in-row coolant distribution units (CDUs), the pump-and-heat-exchanger units that move heat from rack coolant into the building’s water system, and by rear-door heat exchangers, radiators mounted on the back of a rack that capture residual heat. In that layout, every CPU rack added for agent sandboxes is another load on the CDUs, the facility water loop and the containment scheme. It is not a cheap air-cooled addition at the edge of the hall.
The affected groups are colocation providers, enterprise data center teams and the engineers sizing cooling plants for AI clusters. Their liquid-cooling capacity plans have been built around GPU rack counts. If buyers adopt the rack-scale CPU product rather than the air-cooled 1U and 2U alternatives, CDU capacity and piping need to cover the CPU rows as well. The size of that addition cannot yet be calculated, because Supermicro has not disclosed the rack’s power draw or heat load. What can be said is that the reference design treats the CPU tier as liquid-cooled floor space by default.
800 VDC and Batteries in Place of a UPS
The power design follows the same logic of standardizing around the GPU rack. Customers can feed the compute rack with conventional AC power, or with an 800-volt DC sidecar power rack, a dedicated rack beside the compute racks that converts and distributes power. The sidecar option allows battery backup units to replace a conventional uninterruptible power supply (UPS), the room-scale battery system that keeps equipment running through a utility outage.
Supermicro says this approach frees data center floor space, can improve end-to-end power conversion efficiency, and lowers both capital and operating costs. These are reasonable design goals, since fewer conversion steps generally mean less energy lost as heat. The announcement does not quantify any of them, however. Operators evaluating the option will want efficiency figures, battery ride-through times and an account of how a sidecar-based design interacts with their existing electrical distribution and redundancy standards.
Integration Is the Product, and the Margin
Security is offered as a hardware feature. NVIDIA’s OpenShell Reference Architecture runs on the Vera CPUs and applies least-privilege, policy-based controls to agents at runtime. NVIDIA Sentry runs on the BlueField-4 DPUs, network cards with their own processors, as an out-of-band monitor. It operates separately from the host system, enforces policy, provides attested telemetry and, according to the announcement, can quarantine an agent within milliseconds if it oversteps its limits. Moving that monitoring off the host CPUs is meant to preserve compute for the sandboxes themselves. Independent validation of these claims will matter to buyers putting autonomous agents into production.
Supermicro’s broader pitch is end-to-end delivery. One team covers site survey, design, rack- and cluster-level integration and testing, on-site deployment and ongoing support. That positioning fits the company’s financial profile. In the October 20 business update filed with the SEC, Supermicro estimated first-quarter fiscal 2027 revenue of $18.5 billion to $19.5 billion and gross margins of 12.0% to 12.5%, both above guidance. It attributed the revenue beat primarily to increased demand for AI platforms and the margin improvement primarily to customer and product mix. The figures are preliminary and unaudited. They are also thin margins on very large volumes, which helps explain why the company emphasizes integrated racks, cooling and services over standalone servers. The Vera CPU rack is quote-only today and contributed nothing to that quarter. Investors will hear more on the November 3 earnings call.
Background
Super Micro Computer, based in San Jose, California, designs and manufactures servers, storage and networking systems in the US, Taiwan and the Netherlands. It has become one of the largest suppliers of AI server racks built around NVIDIA platforms, and it increasingly sells complete data center packages, which it calls DCBBS, covering power, liquid cooling and deployment services. In an 8-K filed October 15, the company reported that Mark Long had been appointed chief financial officer, succeeding David Weigand.
NVIDIA’s Vera Rubin platform pairs its Vera CPU with its Rubin GPU. Vera Rubin NVL72 is the rack-scale GPU system that Supermicro says it is now shipping. The new Vera CPU rack extends the platform into CPU-heavy workloads created by AI agents, using the same third-generation MGX modular rack design so that both rack types can be deployed and cooled as one system. Source: Supermicro introduceert het NVIDIA Vera CPU-rack: een volledig vloeistofgekoelde oplossing op rackschaal met 256 CPU’s, als aanvulling op het NVIDIA Vera Rubin-platform voor agentische AI op grote schaal. Supermicro’s announcement of its liquid-cooled, 256-CPU rack for agentic AI sandboxes. Primary sources: Super Micro Computer Form 8-K filed October 20, 2026; Supermicro Provides First Quarter of Fiscal Year 2027 Preliminary Business Update (Exhibit 99.1); Super Micro Computer Form 8-K filed October 15, 2026 (Chief Financial Officer Transition).Sources