The coming robotics wave

Start with why anyone in this sector wants robots, which is labor. Lubershane points to three trends that are not reversing soon: fewer young people being born in advanced economies, the likelihood of a more restrictive United States immigration environment closing off the route by which the country had escaped that demographic squeeze, and labor productivity that has been flat for 10 to 15 years in manufacturing and declining for a decade in construction. Then Kann presses on how binding that constraint actually is. He says he has heard endless complaints that labor is hard to find, but has never heard anyone say a project failed because the workers could not be found, and Lubershane agrees he has not either. What both see instead is rising wages, high turnover and a general drag on cost. That is a real headwind, but a slow one, and Kann draws the consequence: nothing about it would trigger a wave of deployment unless the machines themselves changed.

They did, on two tracks at once. On capability, generative AI has begun to produce robots trained to handle unstructured tasks rather than the deterministic ones factory arms have done for decades, and Lubershane points to Google’s PaLM-E model and to a Stanford group’s low-cost Mobile ALOHA system as the seminal work. That matters because this sector’s real work happens outdoors, where nothing can be programmed move by move. On cost, everything inside a robot has gotten cheaper: the chips to run the models, laser ranging sensors (lidar) that he puts at roughly 90% cheaper over 15 years thanks largely to autonomous vehicle investment, electric actuators for the joints, and batteries. Kann draws out the economics that follow, and Lubershane agrees with them. With capital cost that low, a robot priced at or near what the human-labor alternative would have cost pays back remarkably fast, but only if utilization is high. That leaves three open questions rather than one: whether the pricing holds, how much the machine actually gets used, and whether it is reliable enough to trust. Electrification compounds the effect, since an electric drivetrain costs less to run and needs less maintenance than an engine or hydraulics, which supports the high utilization the payback math depends on.

Kann then supplies the counterweight, and Lubershane partly concedes it. Substituting capital for labor is exactly the wrong trade when interest rates are high, because the whole cost now has to be financed, and rising electricity prices cut against a technology that is itself a form of electrification. Lubershane agrees cost of capital is the bigger of the two, and says the market’s answer is the robotics-as-a-service model, where the producer keeps the asset and sells the service, absorbing the financing and raising utilization by spreading one machine across several customers. Kann is careful to call both headwinds marginal and possibly transient rather than blocking: he has heard nobody say they would have automated a task but for the price of power.

Which leaves the question of where this works first, and the sorting principle is how long the tail of strange situations is. A robot bolted to a factory floor sits at one end, a self-driving car in San Francisco at the other, and field construction in the middle. That is why Lubershane calls a utility-scale solar project the perfect frontier: the work is nearly identical across thousands of acres, yet slope, ground moisture and the mounting hardware vary enough that it cannot be scripted, so it needs the new AI without needing a machine that can handle anything at all. He hedges that as the frontier “today,” not a permanent ranking. The tail is still real, and Kann’s example is a one-month pilot in the southwest that met 110-degree heat, high winds, sleet and ice inside that single month. On top of the technology sits a commercial hypothesis: a company building a vertical robot has to get units into the field early, learn the exceptions, and come down the cost and performance curve fast enough that a follower arriving two or three years later cannot catch up. Lubershane prefers that bet to the alternatives. General-purpose humanoids are, by his estimate, at least one and perhaps two orders of magnitude harder, even though individual humanoid companies have raised hundreds of millions and in some cases billions. He holds the emerging robotic foundation models at arm’s length too, doubting both their long-term moat and whether sophisticated robot builders will hand that layer to a third party.