Skild AI's brains are on hundreds of factory robots, and the buying decision moves up a layer

Divide the revenue by the fleet and the business gets legible: somewhere between about $33,000 and $100,000 per robot for the period, depending on whether "hundreds" means three or nine ^[18]. That is not seat pricing. It reads like equipment or engineering work, which fits deployments that still look hand-fitted to each site: Nvidia's Houston factory, several unnamed wiring and construction firms, and a trial at LaGuardia ^[7].

The OEM line in the source has the longest reach. Three established machine vendors are putting Skild's model inside products they sell under their own brands ^[8], and Skild makes no hardware at all ^[1]. So a plant that orders a familiar arm can end up operating a model from a company three years old ^[2] that it never evaluated, on commercial terms the source does not disclose. Whether that is a per-unit royalty, a site licence or a subscription decides who keeps the margin once the mechanical half of the package is the cheaper half, and none of it is public.

Horizontality is the whole pitch. One model is meant to drive quadrupeds, humanoids, tabletop arms and mobile manipulators without being rebuilt per body ^[9], on the argument that no single task and no single machine produces enough data to train on, so you pool everything and learn one model that transcends the form factor ^[11]. Co-founder Abhinav Gupta's motto is "any task, any hardware, one brain" ^[10]. The buyer's upside is that competence learned on one body arrives on another. The exposure runs the same way: a regression in the one brain does not stay inside one product line.

Then the arithmetic on the money. The valuation is roughly 467 times that trailing half-year revenue, or about 233 times if you simply double it ^[15]. Everything raised to date is about 67 times the same revenue ^[16]. Notably, several of the January investors are industrial suppliers rather than pure financiers ^[3], which is a distribution signal as much as a funding one.

On timing, the sources do not agree. The publisher's framing is that the GPT moment for robots has already arrived, just not in the home ^[14]. Gupta says what is visible now is "the start of a GPT moment which will happen over the next one, one and a half year" ^[17]. Jensen Huang's line at CES, eight months before the piece, was that it was "nearly here" ^[12]. Those are three different purchase decisions. What has actually crossed over is narrower than any of them: robot competence is now something a factory buys from a software vendor, and the machine it arrives in is chosen separately.