Solar units for deploying AI off the power grid

Startup Liquid Energy presents NanoDC, a solar module designed for AI inference. The first deployments are announced for November 2026.

Solar panels that unfold like an accordion to power a computing module: that is the format presented by Liquid Energy for its NanoDC. The whole unit is factory-built, transported to the site, and designed to operate without connection to the electrical grid. The startup announces a startup time of thirty minutes after delivery, with solar production, battery storage, and Internet connection via Starlink.

The Canadian company, part of Y Combinator’s Fall 2026 batch, is targeting inference, meaning the execution of already trained models to produce responses. Its offering focuses on this computing capability: the commercial website says it sells inference to businesses. The module is the infrastructure meant to deliver this service. Generating electricity in the same place as the computing

Each unit brings together the computing equipment and its energy supply. The panels provide electricity, while the batteries store part of the output. Additional capacity is obtained by adding modules, each arriving with its own cooling and connectivity equipment.

According to the company, this built-in power supply is meant to avoid grid-connection delays and certain electrical equipment, including the transformers and turbine generators mentioned in its announcement.

Cooling is advertised as air-based, with a system the company describes as suited to high compute density. The launch message states that the unit does not need a water supply. It also presents Starlink as the means of connection, to avoid depending on a terrestrial link that would have to be built on site.

In a presentation published by LvlUp Ventures, drawn from its funding application, the company says it uses lithium iron phosphate batteries, or LFP, and claims continuous operation. This description remains a company statement: the documents reviewed provide neither storage capacity per module nor an operating record allowing an assessment of service continuity during periods without sunlight. Four weeks, thirty minutes, or eight weeks

The stated timelines do not all start from the same point. In its launch message, the company speaks of a four-week deployment and a thirty-minute startup once the module has arrived. The exact scope of the four weeks is not detailed; the thirty minutes apply to a unit that has already been delivered.

The commercial website uses another benchmark: less than eight weeks between contract signing and delivery of the first token, in other words the actual start of the inference service. This timeline includes a commercial relationship and a commitment to provide compute. It should not be confused with the thirty minutes announced for installing a unit that has already been delivered.

The promise of deployment “anywhere on the planet” remains, for its part, poorly documented. The public pages do not specify the solar conditions assumed, the acceptable climate constraints, or the sites on which the timelines were measured. The absence of a grid connection is the mechanism presented to simplify installation, without any published demonstration of deployment valid in all environments. Selling compute from a modular fleet

The commercial proposition consists of providing compute remotely, with results transmitted over the Internet. It therefore does not necessarily require each customer to host a unit on their own land. The location of the modules and the provision of the service can be separated.

The company does not publish, in the pages reviewed, any inference pricing or performance per module.

The company’s profile at Y Combinator mentions twenty units. It does not specify how many are already in operation, or what computing equipment they carry. The launch message announces the first deployments for November 2026, followed by shipments of thousands of units starting in 2027.