Embodied AI FRONTIER
/ Deep Dive / China's Embodied AI Companies Haven't Learned How to Spend Money
Deep Dive 📡 Embodied AI Frontier · Deep Dive

China's Embodied AI Companies Haven't Learned How to Spend Money

WHY READ

Why it matters: the first data-driven look at capital efficiency in China's embodied intelligence boom, built entirely on public filings and disclosures.

By Embodied AI Frontier

Abstract

China's embodied intelligence sector raised more than RMB 46 billion in disclosed funding in the first half of 2026, and over RMB 96 billion in the twelve months to June. Money is arriving faster than any previous technology cycle, yet "how to spend it well" has become the industry's least understood skill. In the same period, one company reached global leadership in humanoid shipments with annual R&D of less than RMB 100 million and turned a profit, while another, branded the "first humanoid-robot stock," lost more than RMB 4.2 billion over four years. The world's most famous robotics company has never recorded a profitable year in three decades, and an American star startup went from a $1.2 billion market cap to $3.7 million in the bank. Based on public prospectuses, listed-company filings, HKEX listing rules, and public reporting, this article dissects the two extremes of spending in China's embodied AI industry and the three cost structures behind them, and asks a question with no standard answer: once money far beyond a company's operating capacity enters its account, who decides how fast it is spent, and in what direction?


Article structure
  1. 1. The Flood: RMB 100 Billion With No Playbook
  2. 2. The Two Extremes: Frugality and Excess
  3. 3. The Uncountable Ledger: Talent, Compute, and Data
  4. 4. Where the Money Goes: Rules, Governance, and Valuation Gaps
  5. 5. The Watershed: The Spending Test of Late 2026

In 2025, one company sold more humanoid robots than anyone else in the world: revenue of RMB 1.71 billion, net profit excluding non-recurring items of RMB 591 million, and gross margin above 60%. A year earlier, its entire annual R&D budget had been about RMB 70 million. That is roughly the same order of magnitude as the total compensation package of a single top research scientist at an American AI giant.

Another company, branded "the first humanoid-robot stock," generated RMB 2.00 billion in revenue in 2025 and lost RMB 790 million. Its R&D spending from 2022 to 2024 totaled RMB 1.40 billion, more than nine times that of the first company. Over four years, it lost more than RMB 4.2 billion.

Same industry. Same time window. Two ways of spending money, two outcomes. After a funding frenzy unmatched in the history of Chinese tech, China's embodied intelligence sector is confronting, for the first time, the backlash of capital: the industry has no consensus on what good spending looks like, but the results of bad spending are already written into the financial statements.

1. The Flood: RMB 100 Billion With No Playbook

1.1 The scale and rhythm of fundraising

Start with the numbers. According to IT Juzi, a Chinese startup data provider, the domestic embodied intelligence sector recorded 288 financing events involving 226 companies in the first half of 2026, with disclosed funding exceeding RMB 46 billion (about $6.5 billion). Extend the window to the twelve months from July 2025 through June 2026, and the count grows to 503 rounds totaling more than RMB 96 billion (about $13.5 billion). On average, more than one round per day.

Mega-rounds have become routine. More than 25 transactions of RMB 1 billion or above closed in the first half of 2026 alone. In the first quarter, disclosed events exceeded 40 and the number of companies valued at RMB 10 billion or more expanded past ten: Zibianliang Robotics (自变量机器人) raised RMB 1 billion in an A++ round in January; LimX Dynamics (逐际动力), one of China's better-known humanoid startups, closed a $200 million Series B in February; AI2 Robotics (智平方) raised RMB 1 billion across its Series B in the same month; and Galbot (银河通用) closed a RMB 2.5 billion round in March.

1.2 Records falling in sequence

Records keep being broken. TARS (它石智航), a Shanghai startup founded in February 2025 by alumni of Huawei's autonomous-driving and DJI's computer-vision teams, completed a $120 million angel round in March 2025, at the time the largest angel round the sector had seen that year. In July 2025 it added another $122 million angel-plus round. On April 16, 2026, TARS announced a Pre-A round of more than $455 million, co-led by Hillhouse, HongShan (formerly Sequoia Capital China), and Meituan's strategic investment arm, setting new records for both the largest single round and the largest Pre-A round in the sector's history. In under fourteen months, TARS raised close to $700 million.

Galaxea (星海图), founded in September 2023 by a Tsinghua-linked team including former Waymo engineers, completed a RMB 1 billion Series B on February 11, 2026, pushing its cumulative funding to nearly RMB 3 billion at a valuation above RMB 10 billion. A month and a half later, on April 2, it closed a further RMB 2 billion B+ round, lifting cumulative funding to nearly RMB 5 billion and its valuation above RMB 20 billion. The valuation roughly doubled in under two months, one of the fastest rises in the sector in 2026.

Galbot, founded in 2023 by Peking University researcher Wang He, had accumulated about $800 million by the end of 2025, already the most of any Chinese embodied intelligence company. In March 2026 it added another RMB 2.5 billion, with the National Artificial Intelligence Industry Investment Fund among the investors, the first time the state's flagship fund system has backed an embodied intelligence company. Galbot remains the sector's most heavily capitalized private company.

1.3 The richest-per-employee sector in Chinese internet history

The result of this injection is a phenomenon unseen in Chinese internet history: embodied intelligence may be the sector with the most "capital per employee" in Chinese history. A company with barely a hundred employees can hold hundreds of millions of yuan in cash. At the end of September 2025, Unitree (宇树科技) held RMB 1.80 billion in cash and cash equivalents, and more than RMB 2.5 billion including wealth-management products and deposits, with a headcount in the low hundreds.

But money arriving this fast creates its own problem. The industry is crowded with scientists and engineers who excel at reconstructing world models in three-dimensional space, yet few are veterans of capital operations, and fewer still have experience spending money at this scale. When lab thinking collides with a flood of capital, putting money to work becomes a harder test than raising it. The investors who prided themselves on financial discipline have no benchmark either. Spending, in other words, has become the industry's least familiar lesson.

2. The Two Extremes: Frugality and Excess

2.1 Unitree: a year of R&D cheaper than one AI star's pay

Unitree's prospectus, filed for a listing on the STAR Market, China's NASDAQ-style board for tech firms, is the extreme example of frugality.

From 2022 through 2024, Unitree's R&D expenses were RMB 30 million, RMB 50 million, and RMB 70 million respectively, totaling RMB 150 million over three years, never exceeding RMB 100 million in any single year. In the first nine months of 2025, R&D was RMB 90.2 million, bringing the R&D ratio down to 7.73% of revenue. For comparison, UBTECH (优必选)'s R&D expenses over the same three years totaled RMB 1.40 billion, more than nine times Unitree's. Chinese media ran the arithmetic: Unitree's full-year 2024 R&D of about RMB 70 million sits in the same range as the annual compensation package of a top researcher at OpenAI, whose total pay for elite research staff can reach the RMB 70 million level. A robot company shipping the most humanoids in the world spent roughly what one AI superstar earns.

Marketing is equally restrained. In the first nine months of 2025, Unitree's selling expenses were about RMB 76 million, or 6.51% of revenue; its entire advertising and promotion budget, including sponsorship of the CCTV Spring Festival Gala, was RMB 22.6 million. UBTECH's selling-expense ratio exceeded 40% in 2024.

Frugality, however, does not mean not spending. It means spending where it counts. In 2025 Unitree generated revenue of RMB 1.71 billion, up 335% year on year, with net profit excluding non-recurring items of RMB 591 million and more than 5,500 humanoid units shipped, the most globally. In August 2026, Unitree priced its IPO at RMB 150.80 per share, raising RMB 6.10 billion in gross proceeds at 219 times earnings, one of the most expensive robotics listings ever. The four main uses of proceeds: RMB 2.02 billion for embodied foundation-model research, RMB 1.11 billion for robot body development, RMB 445 million for new product development, and RMB 624 million for manufacturing facilities, about RMB 4.2 billion in total. The company that had never spent more than RMB 100 million a year on R&D is now committing most of its IPO money to the "brain" it openly admits it lacks: the prospectus states that its general-purpose embodied foundation model has not yet been applied at scale in products, and its low R&D ratio drew questions from the Shanghai Stock Exchange.

2.2 UBTECH: the first humanoid-robot stock, and RMB 4.2 billion in losses

UBTECH is the mirror image. On December 29, 2023, it listed on the Hong Kong Stock Exchange as the "first humanoid-robot stock," with a first-day market capitalization of about HK$38 billion. Before listing, it had raised about RMB 5.6 billion over a decade from Tencent, Qiming Venture Partners, CDH Investments, and a roster of state funds.

UBTECH is the archetype of high spending: R&D expenses of RMB 429 million, RMB 517 million, and RMB 428 million in 2020, 2021, and 2022, averaging 56.5% of revenue; RMB 507 million, or about 25.3% of revenue, in 2025; and nearly RMB 1.9 billion of cumulative R&D over four years. High spending did not produce proportionate returns. Net losses from 2022 through 2025 were RMB 987 million, RMB 1.27 billion, RMB 1.16 billion, and RMB 790 million, more than RMB 4.2 billion in total, narrowing each year since 2024 but still far from breakeven. At listing, its much-publicized Walker humanoid line carried an average price of RMB 5.99 million per unit and had sold ten units in the two and a half years to mid-2023.

The contrast within one industry is stark: the company spending one-ninth as much on R&D turned profitable first; the company spending nine times more is still losing money. The point is not that spending less is right. It is that in an industry where the technology has yet to converge, the structure and efficiency of spending matter more than the total. Unitree's low R&D ratio reflects a deliberate focus on the "body plus cerebellum" path (hardware plus motion control), vertical integration of core components, and cost engineering. UBTECH pursues full-stack self-development across hardware and software, a much bigger footprint and a much heavier ledger. Which route wins is still an open question, but capital has begun to vote: Unitree's IPO at 219 times earnings was oversubscribed by roughly RMB 1.9 billion.

2.3 Galaxea: turning prudent spending into a public strategy

On the frugal side, one company has turned the question of how to spend into a public strategy statement. When Galaxea closed its RMB 1 billion Series B on February 11, 2026, CEO Luo Tianqi explained why the company is deliberately cautious: embodied intelligence is fundamentally an AI business that follows scaling laws, and with data still scarce, R&D investment will inevitably rise exponentially, so the company must keep cash reserves as a buffer for the coming explosion in data collection, model training, and compute demand. "This industry is a marathon, not a sprint," he said. On April 2, after the RMB 2 billion B+ round lifted Galaxea's cumulative funding to nearly RMB 5 billion, Luo added: "In 2026, we want spending to move from efficiency to results. In the past six months, our R&D expenses have been several times what the company spent in its entire history. The starting gun for the scaling race has gone off."

Galaxea has raised five Series A rounds in 2025 and two billion-yuan rounds in two months in 2026, yet keeps a cash buffer at the top of its operating priorities. It has not disclosed burn figures, but its public statements treat prudent spending itself as a competitive asset. A founder openly explaining his spending logic is rare in China's private markets (venture capital and pre-IPO funding), and it signals that anxiety about capital efficiency is surfacing across the sector.

🔒
Unlock the full deep dive
Sign in to read the complete analysis — members get full access to every deep dive.
Sign in / Sign up
✓ Verified 2026-08-16
Recommended
Embodied AI FRONTIER