One number from Shanghai and one from Wellington describe the same invoice. Unitree’s first filing as a listed company shows quarterly R&D up 249.6 percent — 127 million yuan, 19.1 percent of revenue — booked to win embodied intelligence before profit is kept, Reuters reported on 28 October. The same morning, Stats NZ’s September employment data showed New Zealand’s job growth skipping the under-30s almost entirely. Two economies, one arrangement: automation’s next phase is prepaid — in yuan by robot makers, in first jobs by everyone else.
🔍 THE BOTTOM LINE
The AI economy now pays for its robotics future the way it pays for compute: up front, off today’s margins, against returns that arrive later or never. Unitree’s 127-million-yuan R&D quarter is that bill priced in money; the zero-growth door for 22-to-25s in the most exposed US occupations — against 9.6 percent for their peers — is the same bill priced in first jobs. Neither appears as a loss. Both compound.
The robot makers are pre-paying for intelligence they don’t have yet
Unitree’s Q3 filing is the first hard look at what embodied-AI competitiveness costs. Revenue grew 70 percent year-on-year to 665 million yuan; net profit still fell 17.4 percent against forecasts of roughly double. The nine-month core is softer still: profit excluding one-offs fell 18.8 percent, operating cash flow dropped 54.7 percent while R&D rose 191 percent. Booking nearly a third of quarterly revenue into one month’s R&D isn’t a product cycle — it’s an option premium on a category that may not price rationally for years.
Unitree’s public repos show the posture on the ground: unitree_ros was pushed the day the earnings landed (1,619 stars, 457 forks), with unitree_sdk2 at 1,425 stars pushed a fortnight prior — development plumbing still shipping, not winding down. XPeng’s solid-state Iron-robot batteries are rolling off the line in small batches before the robot itself reaches production. The bill comes due against hardware margins, in an industry where Physical Intelligence — closest pure-play on the model layer — is reportedly raising US$1 billion at a $20 billion valuation, pre-revenue.
Who is NOT paying attention: the margin-priced incumbents
The counterpoint is every employer pricing automation as a line-item discount. NZ hiring surveys read as recovery — ANZ’s read of NZIER’s QSBO puts employment intentions at +23, the cycle’s strongest — but who got the jobs is the tell: 35-to-39s took the largest gain, 65-plus grew 4.3 percent, and interest.co.nz’s read of the same release is blunt that under-30 employment is declining sharply. Firms are hiring back the 2019 workforce composition — experienced, credentialed, immediately productive — and treating the entry rung as a deferrable expense. The wiki tracker that reached Hacker News’ front page shows the same shape in US data: exposed firms keep the 40-year-olds and quietly don’t hire the 22-year-olds, because a deferred first job costs nothing on any board dashboard.
The door-narrowing economy has a 12–24 month clock
Nimwiki’s public tracker, updated 26 October with US Census data through September, states the quietest verdict in the debate: not yet across the economy — but check the young workers. Economy-wide, the 135 most AI-exposed occupations grew 2.8 percent against 3.7 percent for the rest — inside the noise. But 22-to-25s in the most exposed quarter grew 0.1 percent against 9.6 percent elsewhere, the July-September gap the widest on record at 17.9 points. This morning’s read of the tracker carries the detail. A mechanism this quiet produces no crisis moment — it produces a cohort arriving at 30 without the job where the learning used to happen, and firms that in 2027-28 discover mid-level pipelines have a lead time no hiring surge compresses. The tracker names its falsification dates: October’s Challenger report and Census microdata, both due early November.
Where New Zealand actually sits in this
New Zealand is the pre-paid side without the robots: this morning’s data-verdict article mapped growth concentrated in care, government, construction and teaching — the categories automation vendors market least and NZIER’s youth-jobs paper called the refuge doors. New Zealand has no Unitree, no XPeng battery line, no frontier-model layer to pre-pay into — only the demand side of the same trade, waiting for the intelligence layer to become someone else’s product. The exposure arrives in first jobs rather than yuan: the entry-narrowing mechanism applies to young New Zealanders whether or not anyone here builds a humanoid. For a Kiwi firm, the pre-pay era rewards booking capability spend now — training pipelines, first-job intakes — while competitors treat the rung as deferred maintenance. For a Kiwi 22-year-old it’s blunter: the occupations that grew beyond the noise — software development, project management, oversight roles — all moved up the judgement stack, the shift yesterday’s review-economy essays pointed at. Entry hasn’t closed; it has moved.
❓ FAQ
Is AI taking everybody’s jobs? The broad data says no: the most AI-exposed US occupations grew 2.8 percent against 3.7 percent for the rest (year to September 2026), inside the survey’s noise. The clear effect is on who gets hired, not headcount.
Why are young workers hit first if there’s no layoffs wave? The mechanism is hiring, not firing: exposed firms keep staff and stop hiring the bottom rung, so 22-to-25s in the most exposed jobs saw zero growth while peers elsewhere grew 9.6 percent. A missing first job never appears as a layoff.
What does Unitree’s profit fall have to do with NZ jobs? Two sides of one arrangement: Unitree pre-pays for embodied intelligence out of hardware margins because falling behind costs more; exposed-occupation employers pre-save on entry wages because hiring juniors feels optional. Both defer the cost.
Should NZ care about China’s humanoid economics at all? As a price signal: Unitree stays profitable booking 19 percent of revenue into R&D while XPeng tools battery lines before robots ship — the stack NZ buys rather than builds is being priced elsewhere, and the component-first ramp shows where the bottleneck sits.
What’s the single thing to watch next? The Household Labour Force Survey on 4 November (whether NZ’s under-30 slide mirrors the US split), plus the Challenger and Census US releases that week, which the nimwiki tracker names as its falsification dates.
🔍 THE BOTTOM LINE
Two stories on one morning agree in opposite currencies: China’s robot makers booking the cost of embodied intelligence ahead of any product that could pay for it, New Zealand’s labour market booking the cost of cognitive automation ahead of the workers it will need to hire back. Up-front spend against deferred returns is heavy industry’s oldest financing trick — and the cohort carrying the deferred half is too new to the labour market to have a lobby. The 12-24-month question isn’t whether the bill comes due; it’s who has paid it, and who discovers in 2028 that a mid-level pipeline can’t be bought back at any price.
📰 Sources
- Reuters (via SRN News) — Unitree Q3 earnings
- Robotico.market — Unitree filing analysis
- nimwiki — AI and jobs tracker (Census CPS data)
- CnEVPost — XPeng Iron solid-state battery production
- GitHub API — Unitree public repositories
- Stats NZ — Employment indicators, September 2026
- ANZ Economics — NZIER QSBO Q3 2026 commentary
