Overview
Reuters' November 1, 2026 long-form report — written by two Beijing-based correspondents after interviews with founders and executives of five Chinese robotics companies plus industry analysts — anchored its narrative on ITjuzi data: in the first half of 2026, Chinese embodied intelligence companies completed 322 financing rounds worth 93.5 billion yuan (~13 billion USD), five times the same period a year earlier. Spread across the 181 days of H1 2026, that is over 500 million yuan per day, roughly two rounds per day at nearly 300 million yuan each.
Reuters' one-line thesis: Chinese robot makers have mastered movement; now the race is for the brain.
Breaking down the 93.5 billion yuan
The headline number is solid, but the accounting method matters:
- A post-Repost Chinese analysis unpacked it: the 93.5 billion figure includes estimates for undisclosed amounts and merges multiple rounds from the same company. Counting only disclosed amounts yields 46.239 billion yuan as of June 4.
- ITjuzi's own split: the top 20 funded companies took 55 billion yuan — 9 humanoid full-body makers 31.5 billion, 7 "embodied brain" developers 14.5 billion, 3 components/chip companies 5.3 billion. Full bodies: 33.7%; brains: 15.5%.
- A domestic newspaper retrospective on October 20 cited the same 322 rounds / 93.5 billion but reported 137% growth (2.37x), not five times. The base-period definition could not be verified against ITjuzi's original table, so both figures are presented side by side without adjudicating.
- The 93.5 bn figure is H1 2026 data, cited retrospectively in November reporting.
- The breakdown and Top-20 split were not verified against ITjuzi's original database (secondhand analysis).
- The 137% vs. 5x discrepancy is unexplained; both are listed without a verdict.
- Company valuations, salaries, and GEN-1.5 success rates are self-reported or Reuters-relayed.
- "US firms lead in some advanced robot AI" is a source's judgment, not an established fact.
| Basis | Amount | Notes | |---|---|---| | ITjuzi via Reuters | 93.5 bn yuan | H1 2026, 322 rounds, 5x YoY | | Disclosed amounts only | 46.239 bn yuan | Through June 4, different cutoff | | Top 20 companies | 55 bn yuan | Bodies 31.5 bn, brains 14.5 bn, components/chips 5.3 bn | | Domestic retrospective YoY | +137% | Same 322 rounds / 93.5 bn, inconsistent with 5x |
Conclusion: capital is drifting toward software, but the share staying with the body is still larger — a slower claim than "capital is pivoting to brains," and closer to the data.
The brain bottleneck: robots must understand rooms on their own
Dermot McGrath (founder, ZenGen Labs): competition has crushed hardware prices, so profits won't come from hardware — they'll come from the brain. Lingqi Intelligence co-founder Gao Yang: the brain is the weakest link in the full robot stack. Galaxy General's Yang Qian: the goal is moving robots from code-driven to common-sense-driven.
The difficulty has concrete shape: chatbots learn from internet text; robots must learn how heavy a cup is, which way a soft cable bends, how much grip force crushes an object. No existing corpus — it must be accumulated through physical interaction.
Cross-Pacific comparison per Reuters: US firms Physical Intelligence, Skild AI, and Generalist build few-shot skill-learning software (Generalist claims GEN-1.5 completes ~60% of simple tasks after one 3–12 second demonstration); China's Galaxy General published research granting manipulation skills from a single human video. Galaxy General, Robbyant-adjacent peers, and Lingqi Intelligence each valued around $3 billion. UBTech posted a chief scientist role with compensation (incl. equity/benefits) capped at 124 million yuan; Unitree reportedly has ~30 researchers on behavior-prediction models. All valuations, salaries, and success rates are company-disclosed or Reuters-relayed, without third-party verification.
Four data-collection routes — the expensive part is humans
| Organization | Method | Disclosed scale | |---|---|---| | Lingqi Intelligence | Hires ~1,000 people to record daily and industrial activities | ~1,000 staff | | ACE Robotics | Wearable sensors capturing store/hotel staff motion and grip force | Not disclosed | | Ant Group Robbyant | Teleoperation on real robots | 160k hours incl. 50k hours human-guided; 20 configurations across 17 brands | | DexForce | Teaching in computer-generated environments | Not disclosed |
Ant's 20 configurations across 17 brands point at a practical goal: one model that doesn't need retraining per machine. Reuters notes Chinese firms are offsetting chip constraints with more efficient models, synthetic data, and simulation, with Huawei and domestic chipmakers filling gaps — background context, not a conclusion drawn here.
The 80/80 test: a measurable threshold
Unitree founder Wang Xingxing (October, after reporting optimistic results testing GPT-6 on robots) defined the "ChatGPT moment for embodied AI": when robots complete ~80% of tasks by voice or text instruction in 80% of unfamiliar scenarios.
Its value: falsifiability. Backflips, dancing, and nine-day walks are impressive but incomparable. 80/80 reduces progress to two countable quantities — but three questions remain unanswered publicly: how unfamiliar scenarios are sampled, who defines the task set, and whether failures count as perception or execution.
From 322 rounds to 80% of unfamiliar scenarios, money alone can't bridge the gap. Whether H1 capital bought real capability depends on whether anyone publishes task sets, failure rates, and human-intervention counts. Until then, "five times" measures capital's appetite — not necessarily robots' competence.