State Grid Corp. of China — the largest electric utility in the world by customer count, serving roughly 1.1 billion people — has issued its 2026 Embodied Intelligence Development Plan. Per Humanoids Daily’s read of the internal documents, the Interesting Engineering writeup, and the South China Morning Post recap:
- 6.8 billion yuan ($994.7 million) earmarked for embodied-AI procurement in 2026.
- 8,500 units on order, in three categories.
- Five named domestic suppliers: Unitree, Deep Robotics, AgiBot, UBTECH, and Fourier Intelligence.
This is the single largest procurement order in the history of the embodied-robotics industry. It is also the first contract for these robots that is not a demonstration, a pilot, or a sponsored marketing event. It is a utility writing a multi-year capex line item.
The procurement, in three rows
| Category | Units | Budget | Job |
|---|---|---|---|
| Quadruped “robot dogs” | 5,000 | ¥1.5B | Substation patrol, mountainous terrain inspection |
| Humanoid robots | 500 | ¥2.5B | Live-line work, ultra-high-voltage maintenance |
| Dual-arm wheeled manipulators | 3,000 | ¥1.8B | Equipment operation, fault processing in substations |
A few things worth noting about this table.
The dogs are the volume. 5,000 quadrupeds, by far the largest cohort, going to the most boring and most useful job in the grid: walking the perimeter of substations and trekking up mountains to inspect transmission towers. This is the Deep Robotics Jueying line’s home turf, and by unit count, the dogs eat the contract.
The humanoids are the prestige line item. Only 500 units but ¥2.5 billion — roughly ¥5 million per humanoid, ~$700K each. They are going to ultra-high-voltage live-line work — the bipedal form factor is required because the legacy switchgear, the catwalks, and the safety harness anchor points are designed for a human body. A wheeled robot cannot reach the same hardware. Per State Grid’s projections, each humanoid is expected to save 500,000–800,000 yuan in annual labor costs, with a 2-3 year payback.
The dual-arm wheeled manipulators are the working-class robot. 3,000 units, ¥1.8 billion, ~¥600K each — handling switchgear racking, breaker operation, and fault response inside substations. This is the category that quietly absorbs the most repetitive shop-floor work.
The five suppliers, all pre-IPO
Per Humanoids Daily:
- Unitree — humanoids and quadrupeds, paving for a Hong Kong IPO.
- AgiBot — humanoids, fresh off the Longcheer eight-hour livestream pilot and the SF Express customs deployment.
- UBTECH — Hong Kong-listed; will recognize this revenue first.
- Deep Robotics — quadrupeds, #1 by quadruped sales in Europe per Hannover Messe.
- Fourier Intelligence — rehab humanoids pivoting hard to industrial.
Four of the five are racing toward 2026 IPOs on Shanghai or Hong Kong. The State Grid contract is the kind of multi-year, single-customer revenue line that fits very neatly into an S-1. This is not an accident of timing.
ROI math the buyer is willing to put in writing
State Grid’s published projections, per eWeek and TechRadar:
- 5x increase in inspection efficiency vs. human crews.
- 90% reduction in personnel risk in high-danger environments.
- ¥500K–¥800K annual labor savings per robot.
- 2–3 year payback period.
- 80% target penetration of intelligent agents in high-risk scenarios by 2027.
- 30% target penetration in key regions by end of 2026.
- Full integration with “Digital Twin” grids by 2030.
The 5x and 90% figures are the ones US humanoid pitch decks tend to claim. State Grid is the first major buyer to underwrite them.
The market that just appeared
State Grid is one utility. The math of the rest of the Chinese energy sector is what makes this story matter beyond a single procurement.
- Southern Power Grid is expected to issue a similar, smaller plan within the year.
- Local utilities and provincial energy groups are queuing.
- Total Chinese energy-sector embodied-AI spend in 2026 is projected to exceed ¥10 billion ($1.46 billion) — the first ten-billion-yuan vertical market for embodied AI.
- The longer-tail TAM, per Xianheng International’s analyst note: 100,000 substations across China requiring 1–2 robots each, implying a ¥100 billion ($14B) substation-automation market.
For context: that ¥100B substation TAM alone is roughly equivalent to the entire 2026 capex of a single hyperscaler.
What this looks like next to the US humanoid story
Same week, two very different markets.
In the US, 1X opened its Hayward factory targeting 10,000 home-consumer NEO units in year one. Figure published its 03 ramp curve at 240/month doubling. Both are betting on the consumer or general-purpose market — and both are still substantially in the preorder/pilot phase.
In China, the buyer just placed a 8,500-unit order, with a vendor list, a budget, a Q1/Q3/Q4 phase plan, a ROI model, and a 2030 full-autonomy roadmap. The check is signed.
The strategic difference is the buyer. The US humanoid industry is selling to consumers, factories, and venture capital, in roughly that order. The Chinese humanoid industry is selling to the state utility that runs the electricity grid for 1.1 billion people. The two go-to-market paths produce very different unit economics, very different deployment velocities, and very different IPO narratives.
The dry coda
A wheeled industrial robot has been in factories for fifty years. A bipedal robot doing live-line work on a 1,000-kilovolt transmission tower in Qinghai province has been in a deck slide for two years. State Grid just made it a line item.
The “Deployment Year One” branding on the procurement plan is doing a lot of work, and so far the work is honest. The price-per-unit at ¥5 million for a humanoid is roughly 25x what 1X is asking for a NEO. The deployment is also 25x the consequence — these things are not folding laundry, they are operating equipment energized to a million volts.
The first $1B annual contract in the embodied-AI industry went to a state utility, in China, for energy-grid maintenance. Nobody who drew the Magnificent Seven Optimus deck slides in 2023 had that on the bingo card.