Vertiv SmartIT bundles power and direct-to-chip liquid cooling in a factory-validated AI rack

What Vertiv announced

Vertiv has begun selling SmartIT, a pre-engineered rack package aimed at high-density AI deployments. Announced on 26 October 2026 and described by the company as available immediately, a single order covers the rack enclosure, DC power conversion and distribution, direct-to-chip liquid cooling, in-rack coolant distribution and monitoring, all assembled and tested before the unit leaves the factory. The details come from Vertiv's press release, distributed through PR Newswire and summarized in a dev.to write-up.

The intended buyers are enterprises, research institutions and colocation operators, in purpose-built data halls as well as ordinary commercial buildings.

Vertiv positions the product around integration. Scott Armul, the company's chief product and technology officer, frames the point as bringing power, cooling and IT hardware into a single architecture that is validated up front, rather than having separate vendors each sign off on their own gear while the seams between them get sorted out on site.

What factory validation changes

The practical effect is that the work performed in the data hall shrinks. According to Vertiv, on-site commissioning reduces to hooking the package up to the building: the incoming power feed, the coolant loop and heat rejection, floor loading and network uplinks all remain the buyer's responsibility.

Direct-to-chip liquid cooling sharpens that boundary. The rack expects a supply temperature, a flow rate, a pressure budget and a fluid specification, plus a route for heat to leave the building, and none of that ships with the order. What does ship with the rack: leak detection that can shut the flow, automated shunt-trip protection, monitoring over Redfish, SNMP and Modbus, and a one-year limited warranty on Vertiv's standard terms.

Vertiv also makes a notable claim here: existing enterprise rooms and Class A commercial buildings can host AI-class rack density without a wholesale rebuild. The claim covers the enclosure and the cooling it carries; the electrical capacity of an office block designed for desks, frequently the binding constraint, is left unaddressed.

Named NVIDIA systems and the numbers left open

The compatibility list is specific. Vertiv names four NVIDIA platforms, GB300 NVL72, Vera Rubin NVL72, DGX Rubin NVL8 and HGX Rubin NVL8, built on NVIDIA's MGX and OCP's ORV3 reference architectures. The company says it has already deployed the platform with NVIDIA in support of DGX GB300, and showed a unit alongside an HGX Rubin NVL8 demonstration system at NVIDIA GTC Berlin in October.

The deployment range runs from a single rack up to a full pod, per Vertiv's product page. What the announcement does not provide is a measured power envelope, a cooling capacity figure, per-rack power and heat numbers, or any independent cost comparison. For facility planners those figures are usually the starting point: a buyer can confirm which accelerator generations fit the enclosure, but not what density to promise before a site survey or a feed upgrade.

The rest of the row

SmartIT does not cover the whole installation on its own. Vertiv sells adjacent pieces, including SmartRun for row-level integration, rack PDUs for distribution and metering, and Unify for monitoring and controls across power and thermal infrastructure. A buyer who stops at the rack package inherits a monitoring boundary: the enclosure reports on itself, and the rest of the row sits outside that coverage unless ordered separately. The dev.to write-up suggests naming that boundary explicitly in the purchase order.

Why it matters

Buying accelerators has become routine; making the room around them work is the hard part. Rack-scale packages like SmartIT move integration, testing and the associated finger-pointing from the data hall floor into the factory, which matters most for first-time AI builders and organizations retrofitting existing buildings rather than constructing new ones. The model also formalizes a split that phased deployments need: the platform team can order compute and rack as one item while the facilities team works the power feed, coolant loop and floor loading in parallel, with no redesign between the single-rack and pod stages. The gap to watch is capacity. Until Vertiv publishes per-rack power and heat figures, and third-party cost comparisons appear, the announcement works best as a scoping document, a checklist of what to ask before signing anything.