Yuanming Hu

AI · AI 3D generation
胡渊鸣ChinaAge ~32

Founder and CEO of Meshy AI · Author of the Taichi programming language

Tsinghua’s Yao Class and an MIT graphics PhD; CEO one month after graduation; after two tool products failed following a $50 million Series A, he pivoted Meshy to outcome-based pricing and grew ARR from about $1 million to $100 million in under two years

One-line read: Scientist turned entrepreneur: the result came not from starting at Yao Class and MIT, but from pivoting after two failures in the same company—from selling creative tools to selling the 3D assets they produce.

Started after graduationFirst-time founderScaled company$100M+ · 2026-09VC-backedChina to globalFrom Mainland ChinaBorn 1994Top-school trainedAcademic backgroundMIT EECS George M. Sprowls Dissertation Award (2022)SIGGRAPH Outstanding Doctoral Dissertation Honorable Mention (2022)NOI gold medal

Key numbers

Each figure includes its definition, date, and evidence level.

1994Born · age ~32
~27First venture · 2021
Media report$100M+↗annual recurring revenue · 2026-09
Self-reported$1.5B↗post-money Series B valuation · 2026-07
Self-reportednearly $400M↗Series B funding · 2026-07
Media report15M+↗registered users · 2026-09
Media report$50M↗Taichi Graphics Series A · 2022-02

Key stages

What he was doing at each age

~25

Wrote Taichi

~26

Gave up a faculty career

~28

Two products hit trouble

~29

Switched to selling outcomes

~32

$100 million ARR

Geographic moves

The city behind each change in direction

  1. Yangzhou, Jiangsu, China
    ~17 high school student
  2. Yangzhou
  3. Beijing
    ~19 undergraduate
  4. Boston, United States
    ~23 MIT EECS graduate student~25 Wrote Taichi~26 Gave up a faculty career
  5. China
    ~27 Founder and CEO

Selected work

Generates 3D models from text and images for direct use in game engines and 3D printing; charges by credit subscription and API

An open-source, high-performance programming language for graphics computing that Yuanming Hu led during his PhD; it does not generate revenue directly, but brought him reputation, a co-founder, and an early team

Taichi Graphics / cloud rendering

Taichi Graphics' early cloud-native 3D creation platform and subscription cloud rendering both failed to find a working path; cloud rendering was shut down in 2024

How the business works

market
AI · 3D content generation; transactions center on game assets and 3D-printing models
users and payers
Indie game developers, 3D-printing enthusiasts, and 3D artists and designers self-serve subscriptions; teams and enterprises buy Studio / Enterprise; platforms such as Bambu Lab's MakerWorld embed the capability for their users through the API. Revenue shares by segment have not been disclosed
pricing
credit subscriptions (October 2026 pricing page): Free includes 100 credits per month, with assets under CC BY 4.0; Pro $20/month for 1,000 credits; Premium $40/month for 3,000 credits; Ultra $100/month for 8,000 credits; team and enterprise plans, extra credit packs, and an API are also available
acquisition and distribution
The company says more than half of its growth comes from organic channels and LTV/CAC is above 4; observable channels include many SEO landing pages, Blender / Unity / slicer-software plugins, an affiliate program, and plugin access through ChatGPT
main costs
GPU inference (cited on the pricing page as a reason for no refunds), foundation-model training and R&D staff, paid acquisition, and enterprise sales
validation level
scale validated$100 million ARR and 15M+ registered users come from company and investor disclosures and are unaudited; gross margin, revenue mix, and profit have not been disclosed, so profitability cannot be determined
main risks
General-purpose models make 3D modeling a built-in feature (GPT-6 Astra can already operate Blender); open-source models such as Hunyuan 3D and Seed3D drive prices down; individual subscription churn; GPU costs
Growth flywheel · analysis

Credit subscriptions tie revenue to GPU usage → proprietary models lower the cost per generation → cheaper, better models attract more organic traffic and platform integrations → generation volume continues to grow. This is an analytical judgment; the company has not disclosed unit economics.

Meshy does not sell modeling software; it sells a 3D model that can go straight into a game engine or be printed. The pivot was prompted by one user's comment: "I wouldn't pay for this software, but if you took the 3D models inside it out and sold them, five bucks each, I might buy one."Huxiu So it charges by credits: each generation consumes one unit, making the billing unit and the delivered result the same thing.

The timing also mattered. At the end of 2022, ChatGPT and text-to-image products had accustomed the public to "generating content with one sentence". 3D generation had made academic progress, but there was not yet a product leader; at the same time, consumer 3D printers brought in many new users who could not model but needed models. Three years earlier, diffusion models were not mature enough, so this could only have been built as a tool—Taichi Kwu was an example. Three years later, Tripo, Rodin, and open-source models from major companies had all entered the market, and the first-mover traffic and brand window would have been gone.

Scale has been validated: ARR grew from about $1 million at the end of 2024 to $15 million in November 2025, $30 million in March 2026, more than $40 million in April, and over $100 million in September.NetEase But the company has not disclosed gross margin or profit. Whether profit remains depends on whether inference costs fall faster than general-purpose and open-source models drive prices down.

Timeline

  1. Won only a bronze medal at his first NOI attempt as a tenth-grader at Yangzhou High School

    About 17 years oldFailureYangzhou, Jiangsu, China · high school student

    He recalled being "unable to think about anything but the result". That year he won first prize at NOIP, securing admission to Tsinghua, and had already won an APIO silver medal in middle school.

    Source ↗
  2. Won an NOI gold medal as an eleventh-grader, secured admission, and entered Tsinghua's Yao Class

    About 19 years oldStarting outYangzhou → Beijing · undergraduate

    Interned at Microsoft Research Asia; his undergraduate thesis, Exposure, was published at SIGGRAPH 2018.

    Source ↗
  3. Went to MIT for graduate study, briefly considered dropping out, then changed groups and developed ChainQueen and the original idea for Taichi

    About 23 years oldTurning pointBeijing → Boston, United States · MIT EECS graduate student

    His personal website lists a master's degree from 2017-09 to 2018-12 and a PhD from 2019-01 to 2021-03; in an interview with Jiemian, he wrote, "I went to MIT for a PhD in 2017", while media reports have also said he completed the PhD in "three and a half years". This site follows the personal website.

    Source ↗
  4. Taichi paper published at SIGGRAPH Asia; wrote "Frozen in 99 Lines of Code" at the end of the year

    About 25 years oldMilestoneBoston, United States · MIT PhD student

    After Taichi was open-sourced, it accumulated nearly 30,000 GitHub stars. His future CTO, Ye Kuang, and the early team all learned about him through Taichi on GitHub.

    Source ↗
  5. Chose entrepreneurship after using a "coin toss" test to decide between academia, a major tech company, and a startup

    About 26 years oldTurning pointBoston, United States · MIT PhD student

    He first ruled out joining a major tech company; during the coin toss, he realized he hoped heads (entrepreneurship) would come up. He began looking for a co-founder at the end of 2020.

    Source ↗
  6. Founded Taichi Graphics with Ye Kuang one month after defending his PhD

    About 27 years oldStarting outBoston, United States → China · Founder and CEO

    Ye Kuang became CTO. Hu had never held a full-time job and described himself as going "straight from intern to company CEO". Sequoia Seed Fund was the sole angel investor.

    Source ↗
  7. Taichi Graphics raised a $50 million Series A

    About 28 years oldMilestoneFounder and CEO

    Source Code, GGV, BAI Capital, and Sequoia China jointly led the round. It was a bet on him and the market, before there was a validated commercial transaction.

    Source ↗
  8. Neither the foundational development tools nor the cloud-native 3D platform "Taichi Kwu" found a working path

    About 28 years oldFailureFounder and CEO

    The foundational tools were better suited to NVIDIA, and building them himself risked turning into low-margin outsourcing; Taichi Kwu had to compete for users with established local software companies that had been around for two or three decades. He described himself as depressed from March 2022 to March 2023.

    Source ↗
  9. Pivoted to AI 3D and launched the first tool, Meshy Texturer

    About 29 years oldTurning pointFounder and CEO

    The pivot was based on one user's feedback: they would not pay for the software, but would pay for the 3D models inside it. Meshy Texturer generated textures only for existing models; it was not yet a full text/image-to-3D product.

    Source ↗
  10. Released Meshy-1, formally launching Text to 3D and Image to 3D

    About 29 years oldMilestoneFounder and CEO

    An official website post by Yuanming Hu. The new Text to 3D launched on Discord first that day, with the web version following a few weeks later. Early models often generated grotesque "four faces", and the first real users were instead horror-game developers.

    Source ↗
  11. Shut down subscription cloud rendering and bet everything on Meshy

    About 30 years oldFailureFounder and CEO

    The reasons were direct competition with established software, rising GPU costs, and difficult-to-solve global access latency. He later recalled considering returning the money to investors at the hardest point.

    Source ↗
  12. Meshy ARR reached $15 million, with about 6 million registered users

    About 31 years oldMilestoneFounder and CEO

    ARR was about $1 million at the end of 2024; the company later disclosed that full-year 2025 revenue grew 14-fold year over year. This was the first clear commercial result.

    Source ↗
  13. Completed a nearly $400 million Series B at a $1.5 billion post-money valuation

    About 32 years oldMilestoneFounder and CEO

    IDG, Matrix Partners, and Lightspeed China Partners led the round, while Source Code, BAI, Sequoia, and Granite Asia followed on above their pro-rata share—the existing investors who had watched the first two products fail continued to double down. Funding and valuation are not revenue.

    Source ↗
  14. Meshy ARR surpassed $100 million

    About 32 years oldMilestoneFounder and CEO

    It took less than 24 months to grow from about $1 million to $100 million; registered users exceeded 15 million. That same month, Xin Tong, formerly of Microsoft Research Asia, joined as chief scientist. The data comes from investors and media and is unaudited.

    Source ↗

From games to graphics

Yuanming Hu’s parents were both university professors, and his father taught computer science. He encountered 3D Max at five or six, programmed in Visual Basic in third grade, and made five games with RPG Maker XP and Ruby in middle school. He later attributed his choice of 3D to this experience: as a child he wondered, “If a computer can only do addition, subtraction, multiplication, and division, how does it draw such dazzling images?”

At Yangzhou High School he followed the informatics-competition track: he won bronze at the NOI on his first attempt in tenth grade, won gold the next year, and was admitted to Tsinghua’s Yao Class. After arriving at MIT in 2017, he clashed with his advisor over research direction and briefly considered dropping out. After changing groups, he instead had his most productive academic year; the original idea for the Taichi programming language came from this period.

Becoming a CEO immediately after graduation

After Taichi was open-sourced and accumulated nearly 30,000 GitHub stars, it gave him what a startup needed most: partners and a team. At the end of 2020 he approached Ye Kuang, then a software engineer at Google Japan; the two had only written code together on GitHub, and “it took only three minutes to decide to work together.” He defended his PhD in March 2021, incorporated the company on April 6, and became its CEO.

He never held a full-time job; his understanding of business came from a minor in financial mathematics during his PhD and YouTube videos. In February 2022 the company raised $50 million in Series A, but after the money arrived, neither the foundational tools nor the cloud-native 3D platform found a working path.

The pivot after two failures

From March 2022 to March 2023 was the low point by his own account, and outsiders began describing him as a “high start, low finish.” The turning point came from a user who said they would not pay for rendering software, but if the 3D models inside it were sold, “at five bucks each, I might buy one.”

That sentence, together with the generative-AI wave brought by ChatGPT, pushed the team toward 3D generation. In March 2023 they launched the texturing tool Meshy Texturer; in October they released Meshy-1, which could generate 3D models directly from text and images. In 2024, after the last old product, a subscription cloud-rendering service, was shut down, the company placed its entire bet on Meshy.

From $1 million to $100 million ARR

Meshy’s ARR was about $1 million at the end of 2024, $15 million in November 2025, and exceeded $100 million in September 2026—taking less than 24 months. In the July 2026 Series B, Source Code, BAI, and Sequoia, which had invested in Taichi Graphics in 2022, all followed on above their pro-rata share. According to Huxiu, BAI Capital’s Tianfan Wang explained it this way: the greatest risk in AI investing is not getting the direction wrong, but getting the direction right and standing on the wrong floor. They had originally backed the underlying programming language; the company eventually found traction in vertical AI applications.

Meshy is moving in two directions next: first, making 3D generation one part of an Agent workflow (ChatGPT plugins, MCP); second, generating worlds that can be played in real time (the Mora 1 research demo). In September 2026, Xin Tong, formerly of Microsoft Research Asia, joined as chief scientist.

Open verification items

Unknown numbers stay blank instead of pretending to be reliable.

  • No public information is currently available about his birthplace; he attended Yangzhou High School in Jiangsu. The lead that "a 2021 The Paper report described Yangzhou Evening News as his hometown media" could not be traced to the original article.
  • There are two accounts of when ARR reached $30 million: a BAI Capital post says February 2026, while a GDC report says March 2026.
  • July 2026 ARR: Huxiu reported $60 million, while the financing press release only said "approximately 12-fold year-over-year growth".
  • The city where Taichi Graphics was incorporated has not been verified.
  • The revenue mix (shares from individual subscriptions, API, and enterprise), gross margin, profitability, and team size have not been disclosed.
In their words
When the coin was spinning in the air, the answer was already crystal clear in my mind—I desperately hoped it would land heads up.
—— Yuanming Hu · Dec 2023 ↗

More quotes

For a company to move at full speed, its core roles must be full-time.
Dec 2023#commitmentSource ↗

Explaining why he did not pursue academia and entrepreneurship at the same time.

As the top executive of a startup, I believe Why is more important than How, so we should spend more time thinking about Why.
Dec 2023#strategySource ↗

Reflecting on the failure of two products in 2022.

Compared with effort, I think choosing a good direction for a startup is more important, because the returns differ enormously between directions.
Dec 2023#choosing a directionSource ↗
For a startup, if you wait until the timing is fully mature before taking the plunge, it may already be too late.
Dec 2023#timingSource ↗

Discussing the productization of 3D generative AI.

My read

Yuanming Hu is most easily written off as a genius who coasted from one win to the next. In reality, during the three years after his $50 million Series A, two products at the same company failed to find traction, and he at one point considered returning the money to investors. What actually mattered was not Yao Class or MIT, but changing the business model after failure—from selling tools to selling outcomes.

The first lesson worth copying is to change the unit of payment to the result when users say, “I will not pay for a tool, but I will pay for an outcome.” Meshy charges by credits; every payment gives the user a usable model. Second, start with a narrow group willing to accept defects: early models generated grotesque “four faces,” which happened to be exactly what horror-game developers wanted. Third, use open-source work as a channel for hiring and fundraising—his CTO and early team learned about Taichi through GitHub.

In business terms, people pay because 3D models have high labor costs and strong demand in games and 3D printing; a usable model in a minute is worth $20 a month to an individual or small team. Profit is most likely to be eaten by two things: GPU inference costs, and price competition from general-purpose and open-source 3D models. When 3D generation becomes a built-in feature of ChatGPT, Meshy will need to prove that its advantages in quality, workflow, and print success are worth paying extra for.

The caution is that this path’s room for experimentation was purchased with an elite academic reputation: most founders do not get a second chance after their first failure, much less continued overfunding from existing investors after two failures. Also, the $100 million ARR and $1.5 billion valuation both come from company and investor disclosures, are unaudited, and gross margin and profit have not been made public.

Next founder