It Wasn't Smart AI That Took the Jobs. It Was a Clumsy Robot That Keeps Calling in Sick.

Looking six months out, into the back half of 2026, I feel fairly confident about one specific thing: the noise surrounding the numbers is going to get louder before the numbers themselves get any clearer. The evidence is right there in how wildly institutional forecasts for 2026 China humanoid-robot shipments already diverge. UBS says 30,000 units and draws a hard line that "real mass production starts in 2027." The China Center for Information Industry Development says 38,000. Morgan Stanley says 50,000. Meanwhile Xiong Rong, chief scientist at the Zhejiang Humanoid Robot Innovation Center, throws out a figure of 100,000 to 200,000. The low and high ends of that range differ by more than sixfold. When forecasts for the same year, same country, and same product category diverge sixfold, that tells you this market doesn't yet have a demand function anyone can trust. My own read is that when year-end numbers land, whatever they are, they'll probably fall into that ambiguous 40,000-to-60,000 band where both the bulls and the bears will claim vindication.

There's a second signal I expect to be far more reliably observable over the same stretch: the type of deployment. Break down every success story that's surfaced so far and one pattern jumps out immediately. BMW's Figure 02 in Spartanburg did sheet-metal insertion and weld prep. Figure 03 is moving into logistics sequencing. In Chinese factories, robots are loading and unloading. Every single one of these tasks is "moving something heavy, from a fixed spot, repeatedly." These are jobs that call for strength and stamina, not dexterity. I expect more than 80 percent of newly announced deployments in the second half of 2026 to stay squarely within that category. If cases outside that category start appearing — precision assembly or handling of irregular objects, showing up simultaneously across multiple automakers or electronics manufacturers — that should be read as the first real sign that Brooks's skepticism is starting to crack. If the second half of the year keeps producing nothing but loading and logistics sequencing, that means the trough of disappointment is arriving right on schedule.

The medium term, 2027 through 2028, is where I think the real contest gets decided, and I'd narrow it down to exactly two variables: cost and battery. IDTechEx projects the average price of a humanoid robot will fall by as much as 68 percent from roughly $35,000 in 2025 to a $17,000-to-$37,000 range by 2030. UBTech has already cut the production cost of its Walker series by 25 percent versus 2024 and has set 300,000 yuan as its internal commercial-viability threshold. I honestly think the cost side is largely a matter of time. Once component standardization and volume kick in, prices fall almost by definition, and the fact that the International Federation of Robotics counted commercial humanoid platforms rising from three in 2024 to twelve by 2026 is itself evidence that the supply chain is already forming underneath this industry.

Battery life is the harder problem. Leading Chinese models currently run only two to three hours per charge. That matters enormously because the entire economic case for robotic substitution rests on round-the-clock operation. A machine that needs recharging every three hours isn't a 24-hour machine — it's a shift worker, and a shift worker running at 30 percent of human efficiency at that. I think whether the industry achieves a meaningful breakthrough in battery chemistry or hot-swap systems between 2027 and 2028 is the single largest fork in the road determining whether this technology stays confined to niche use cases or scales into mass deployment. Cost falls with economies of scale almost by default. Energy density does not fall the same way, because chemistry doesn't double in performance every year the way semiconductors do.

Whatever happens in employment over that same window will be far less dramatic and far more brutal, in my view. Plenty of people read the news that Goldman Sachs' AI-driven job-loss tracker "slowed" from 16,000 to 11,000 net jobs a month as a reassuring signal. I read it the opposite way. Look at what's actually been cushioning that number: data-center construction employment, which has added 212,000 jobs since 2022, of which only 69,700 are permanent positions. The rest disappear the moment the building is finished. In other words, jobs erased by AI are being temporarily papered over by jobs building the physical infrastructure AI runs on, and buildings, eventually, get finished. I expect that once the data-center construction boom peaks sometime in 2027, the curve that currently looks like it's "slowing" will steepen again. When that happens, people will describe it as things suddenly getting worse, when really, the painkiller just wore off.

Looking further out, across the 2028-to-2031 window, I'd assign probabilities across three distinct scenarios. Start with the bullish one. The fact that Morgan Stanley more than tripled its China shipment forecast within half a year, and is now talking about a $5 trillion ecosystem by 2050, signals that the entire consensus is being repriced in real time. If robot productivity climbs from UBTech's current 30-to-40 percent toward its own targeted 80 percent by early 2027, if IDTechEx's projected 68 percent price decline materializes on schedule, and if BMW's pilot-to-standard-deployment path gets copied across other automakers and logistics operators, the current twelve-platform commercial ecosystem tracked by the International Federation of Robotics moves quickly toward standardization. For context, that platform count was three in 2024 and five in early 2025 — a fourfold jump in two years. If this path plays out, we'd be looking at something like a "smartphone moment" for robotics before 2030. I'd put roughly 25 percent odds on this outcome.

The base case is far duller, and precisely because of that, far more plausible. Goldman Sachs' own baseline projects over 250,000 unit shipments by 2030 and a $38 billion market by 2035. $38 billion sounds like a lot until you realize that's smaller than a single semiconductor company's annual revenue. In other words, even Goldman Sachs is framing this as a specialty-equipment market, not an industrial revolution. UBS's assessment that "real mass production starts in 2027" points in the same direction.

Under this path, humanoid robots settle into a narrow niche — logistics sorting, loading, repetitive assembly assistance — and get quietly better within that lane. The employment shock shows up not as a dramatic collapse but as slow erosion. That's essentially the picture the World Economic Forum has painted: 92 million jobs eliminated by 2030, 170 million created, for a net gain of 78 million. The aggregate is positive. I'd assign this scenario 50 percent odds.

But I think that "net 78 million" figure is the single most dangerous statistic in this entire debate. The World Economic Forum's own report states plainly that the people losing jobs and the people benefiting from newly created jobs are, overwhelmingly, different people. "Net gain" is a concept that only works on a balance sheet. The worker inserting sheet metal in Spartanburg does not become a prompt engineer next month. A 23-year-old loading trucks in a Guangzhou factory would need two to three years of training — and income to survive that training — to become a robot maintenance technician, and that income was supposed to come from the very job that just disappeared. That's exactly why the WEF separately flagged roughly 120 million people as facing serious displacement risk within the next three to five years. Aggregate optimism and individual-level pessimism aren't contradictory; they're two faces of the same coin. I expect the core political fault line of the next five years to form right along that gap.

The bearish scenario gets 25 percent from me. If Rodney Brooks's predicted "trough of disappointment following massive hype" arrives as forecast, here's roughly how the sequence plays out. Two or three humanoid robotics startups miss their promised shipment targets during 2027 earnings season. Batteries are still stuck at three hours. Efficiency still hasn't cleared 40 percent of human output. The 300,000-yuan commercial breakeven point is still a long way off. Kalshi's 21 percent probability for mass-market sales in 2026 gets rediscovered and re-cited everywhere. Morgan Stanley's own caveat — that most demos "should be viewed with caution" — starts getting quoted again, but this time as a warning that came true. Speculative capital, like the 30 billion yuan that piled in during the first quarter of 2026, starts getting pulled back out, a shakeout begins, and even UBS's conservative 30,000-unit forecast ends up missed.

And here's the part of this scenario that matters most, in my view: even here, the jobs don't come back. Capital exiting robotics doesn't automatically flow into entry-level hiring, because hiring was never the default state it's assumed to be. The reason Figure AI's gray line stayed flat for four years was never really about robots at all — it's because not hiring had already become a management virtue in its own right, one that survives regardless of whether the robots underperform or overperform. That's the uncomfortable thread that connects all three scenarios: the bull case, the base case, and the bear case each produce a different robot market, but none of them produce a return to the old hiring habits. Whichever version of 2028 arrives, the muscle for skipping entry-level hiring will already be fully built.

So let me close with one actionable takeaway. At the individual level, I think the old advice — "learn skills AI can't replace" — has become close to useless. If the real front line of substitution isn't dexterity or intelligence but cost structure, then your defense isn't measured by how capable you are; it's measured by how expensive it would be to cut you. There's a hint buried in something Sabrina White said in that GMAC survey: employers are still investing in people who can exercise judgment, define problems, and help organizations navigate change. What robots can't do isn't the loading part — it's the part right before it, deciding what should be loaded and why. And at the societal level, I believe that if we leave the first rung of the ladder entirely to market forces, that rung will disappear. Companies never hired entry-level workers out of charity; it was always an investment in future skilled labor. The moment a company realizes it can coast for three years on robots and senior staff without making that investment, nobody gets hired. And three years from now, when everyone suddenly notices there are no skilled workers left to hire, the first rung that would have produced them will already be gone. I think that's the real bill coming due over the next five years.