Article Extraction

**Title:** 2年估值破100亿，千寻智能却不敢松口气

**Date:** February 27, 2026 (2026年02月27日 18:27)

**Source:** 36氪 (36kr.com) - Originally from 金角财经 (Golden Horn Finance)

---

## Full Article Content

Humanoid robotics has welcomed its sixth "centibillion-dollar club" member. On February 24, Spirit AI announced completing two consecutive rounds of financing totaling nearly 2 billion yuan, with valuation exceeding 10 billion yuan.

The timing itself signals something important. The capital roster is equally intriguing: Cloud Feng Fund, Sequoia China, TCL Ventures, and 360 Fund entered as new shareholders, while Shunwei Capital, Prosperity7, and Damorning continued adding to their positions. Those familiar with early-stage investing understand this "new and old co-investment" structure means institutions have moved beyond observing to clearly choosing sides.

Before 2026, only three companies occupied humanoid robotics' "centibillion-dollar club": Zhiyuan Robotics (by Bilibili personality Chihui), Unitree Technology and Galaxy Universal (which appeared on the Spring Festival Gala). Entering 2026, Xinghai Tu, Zhipingfang, and Spirit AI successively joined this "table" around Chinese New Year.

**The centibillion valuation in humanoid robotics has shifted from anomaly to sector effect.**

Founder Han Fengtao stated bluntly: "2026's embodied AI resembles 2023's large models—without substantial capital and top-tier model performance, you cannot reach the table."

**In humanoid robotics, financing has transformed from "accelerator" to "ticket." Insufficient funding means direct elimination.**

Spirit AI's 2-billion-yuan financing reflects subtle market sentiment: capital remains willing to invest in the future, but only for first-tier players.

Valuation represents more than a label—it's both entry ticket and pressure gauge. Upon receiving it, the countdown begins.

## The Frenzied Record of Hundred-Billion Valuation in Two Years

Among club members, Spirit AI—established in 2024—ranks among the youngest companies, reaching centibillion valuation in merely 26 months.

This financing curve shows minimal interruption. Founded January 2024, the company completed nearly 200-million-yuan seed and angel rounds within seven months; 2025 saw consecutive 528-million-yuan Pre-A and approximately 600-million-yuan Pre-A+ rounds; opening 2026 brought nearly 2-billion-yuan more. Six rounds across two years, totaling approximately 3.328 billion yuan.

This pace represents "time-racing" rather than speed.

**By comparison, Unitree Technology—founded 2016—required nine years reaching centibillion valuation.** Spirit AI compressed this to two-plus years, not through efficiency gains but environmental differences: **it rode the large-model explosion's window.**

Yet Spirit AI didn't emerge from nowhere. Founder Han Fengtao represents robotics industry's second venture. In 2015, he co-founded Lobot, serving as CTO, delivering dozens of models totaling over 20,000 industrial robots.

By his account, this decade witnessed "industrial robot nationalization rates rising from 3% exceeding 50%, **yet most companies failed earning profits.**"

The core problem wasn't insufficient robots but limited execution capability. Industrial robots depend heavily on customization and scenario fixation; switching tasks requires logic rewrites. Weak generalization, limited scale effects, and repeatedly-compressed margins plagued the industry. Han's earlier attempts at "hardware upgrades plus early deep learning" couldn't overcome then-limited AI capability.

The genuine shift occurred in 2023. When ChatGPT appeared, demonstrating cross-task transfer abilities, he recognized the variable had arrived.

"Reducers, integrated joints, motors, complete machines...but these aren't humanoid robotics' greatest opportunity. The biggest opportunity is the brain. So I searched for whoever could build that brain."

Summer 2023, Han met roughly 100 people, ultimately discovering **"one of Berkeley's returning four"** and Tsinghua's Visual & Humanoid Robotics Lab director Gao Yang. "Gao researched end-to-end autonomous driving since 2017, understanding internet video training, **absolutely technically reliable.**"

This represented strategic direction choice: **betting on "large-model plus embodied" fusion.**

January 2024, Spirit AI launched; July that year, the Moz0 prototype emerged. August brought nearly 200-million-yuan seed and angel investments.

March 2025, Spirit V1 VLA model's preview version launched, enabling humanoid robots completing complex extended-duration tasks like **folding clothes.** Technical breakthroughs triggered immediate financing—528 million Pre-A round followed swiftly.

January 2026, the company further open-sourced Spirit v1.5 model. On RoboChallenge's Table30 benchmark, Spirit v1.5 scored 66.09 points with 50.33% success rate, surpassing Pi0.5—becoming the first Chinese open-source embodied model publicly outperforming Pi0.5. Subsequently, nearly 2-billion-yuan financing materialized.

Reviewing Spirit's rapid financing reveals clear rhythm: **each model performance leap corresponds to valuation increases.** This exemplifies "technology-driven financing curves," yet carries elevated risks—should iteration decelerate, capital patience evaporates rapidly.

## The High-Stakes Data Collection Gamble

Unitree's Spring Gala performance reignited humanoid robots' emotional appeal; Spirit's two-year centibillion valuation quietly raised industry ceilings. **2026's humanoid robotics enthusiasm remains undeniable.**

Yet newly-funded Han Fengtao appears unrelaxed. To him, this enthusiasm resembles dividing lines rather than celebrations: "2026's embodied resembles 2023's large models. Without sufficient capital and top-tier model performance, you lack table access."

Clear subtext: windows narrow rapidly. Combined capital-technology pressure means lagging players face near-impossible comeback odds.

Post-financing, Spirit set aggressive targets: **claim global top-3 embodied brain position, scaling effective training data to 1-million hours.** Previous best open-source models used merely 10,000 hours—a 100-fold expansion ahead.

Unlike language models relying on internet text, robot models require real-world physical interaction data: action execution, environmental feedback, force-control errors, failure corrections repeating. Each grasping mistake and path deviation becomes training samples. **Data scale and structure directly determine generalization capability ceilings.**

The challenge: this path carries extreme costs. Real-world data resists "web-scraping extraction," requiring hardware, space, personnel, and time. Scaling creates exponential overhead.

Therefore, Spirit allocates 80% resources toward data systems—essentially strategic gambling. Han's assessment proved direct: **humanoid robotics haven't exploded from insufficient data rather than algorithm intelligence. With adequate early funding, he'd "all-in data" unhesitatingly.**

Yet critical differences appear in "how to acquire data" rather than "whether needing it."

Spirit avoided "small but refined" paths, rejecting limited lab samples. **They selected industrial-scale approaches prioritizing scale with controlled costs.**

Leveraging Gao's autonomous-driving and internet-video training expertise, rather than frame-by-frame complex world-model construction, the team used massive internet videos for pretraining, exchanging fewer parameters for higher starting points while drastically reducing computational costs. This represents standard large-model approaches: **broad coverage first, then physical alignment.**

Offline data employed parallel pathways: teleoperation, wearable-device capture, real-scenario replay. Goals weren't "perfect samples" but cost reduction. Reportedly, compared with traditional approaches, costs dropped ~90%. Only lowered costs support massive scaling.

Interestingly, teams deliberately avoid "cleaning dirty data." Failed grasps, muddled motions, non-standard operations remain preserved. Their logic: the real world itself proves chaotic; overly-clean data prevents learning complex-scenario handling. Compared against heavily-filtered lab samples, **this noisy, diverse data better enables cross-scenario transfer capability, approaching zero-shot generalization.**

This training logic already manifests at model levels. **Spirit v1 released March 2025 completed extended soft-object manipulation—cloth folding represented continuous action planning and execution sequences rather than single grasping.** Focus shifted from "successfully grasped" toward "sustained correct execution."

**December that year, Moz completed consecutive tasks—desk organizing, trash disposal, chair repositioning, microwave heating—in 2025 robotics competition's eldercare category.** Demonstrations showcased not isolated skills but cross-step decision chains.

From industrial-robotics eras, Han understood something most missed: true commercial value never derives from singular spectacular displays, but from stable repetition across thousands of instances. Only repeatable, transferable, scalable capabilities generate genuine commercial returns.

**"All-in data" represents current-stage most rational choice, simultaneously a high-risk gamble.** Once scaling maximizes without proportional performance gains, capital patience evaporates rapidly. Markets buy technical vision but won't indefinitely fund efficiency problems.

## Big-Factory Partnership Insufficient

Yet Spirit AI, despite cloth-folding abilities, primarily chose factory entry first.

Although Gao proposed 2024's "ten-year global 10% personal robot ownership" "double-ten plan," practically Spirit didn't prioritize home scenarios, instead returning to "factory work."

This reflects pragmatic risk-benefit calculations. Home scenarios offer vast imaginary space yet extreme variables, unclear purchase intent, and obscured cost structures; industrial scenarios provide clear tasks, transparent standards, quantifiable verification. For sub-two-year-old companies, cash flow and validation speed matter more than narrative.

Han's industrial robotics assessment proved unsentimental. After ten-year immersion, he witnessed difficult monetization realities: traditional industrial robots provide singular functions, completing fixed operations, with sluggish penetration increases. **China holds approximately 100-million manufacturing workers but only ~3-million robot inventory, with ~300,000 annual additions. Structural replacement never materialized.**

His reasoning suggested: **should embodied brains replace 30% of this labor, automation market scaling experiences quantum jumps.**

December 2025, CATL's Zhongzhou facility commenced operation, heralded as the world's first humanoid robotics factory-scale new-energy battery production line. Spirit's Moz robots handled battery insertion-extraction inspection.

From company establishment to industrial-scale scenario entry required merely 22 months. Per Han: "CATL contact through final acceptance consumed 11 months. Spirit ranks first truly production-line-functional embodied company, having participated producing 1,000+ batteries thus far."

Versus demonstration videos, "production participation" carries weightier significance. Industrial standards demand stability, yield rates, and rhythm alignment rather than fluid motions. Production-line access indicates system controllability within certain parameters.

**Yet this guarantees nothing.**

Spirit's investors include both Jingdong and 360 Fund alongside Shunwei and Cloud Feng Fund, separately connected with Xiaomi and Alibaba ecosystems. Han frequently stated internally: "Our true future opponents are these giant corporations. Hardware-and-software companies like Huawei, Xiaomi, Ideal prove more capable."

Therefore, "two-year centibillion valuation" marks commencement rather than conclusion. Han's targets prove quantifiable: **before major-company entry, achieve "mid-tier" status—selling minimum 100,000 robots annually.** This represents measurable mid-tier benchmarks corresponding to scale effects, cost amortization, and negotiating power.

**Capital environments also tighten. Despite two additional humanoid companies reaching centibillion valuations this year, sector financing enthusiasm declines overall.** Public data showed third-quarter 2025 humanoid robotics financing dropped 42% year-over-year with 35% fewer events, with early-stage project shares declining from 2024's 68% to 32%. Markets shift from "concept betting" toward "implementation demonstration."

This also marks humanoid robotics transitioning from concept phases toward validation stages.

Following the robotics-laden Spring Gala performance, Spirit's 2-billion-yuan financing epitomizes this structural transformation: **capital remains future-betting, yet no longer purchases imagination—instead wagering on "factory-work" capability.**

When industries progress from performances toward production lines, genuine humanoid robotics testing has merely commenced.