# Josh Gao

> Josh Gao is a Canadian entrepreneur and co-founder and CEO of Mecka AI, known for building data and deployment infrastructure for robotics and physical AI.

- Canonical URL: https://iq.wiki/wiki/josh-gao
- Categories: People in crypto
- Tags: Founder
- Created: 2026-10-21T15:51:50.403Z
- Last updated: 2026-10-21T15:51:50.403Z
- Source: IQ.wiki — the world's largest blockchain and crypto encyclopedia (https://iq.wiki)

---

**Josh Gao** is a Canadian entrepreneur and the co-founder and chief executive officer of Mecka AI, a company that collects and analyzes human-movement data for robotics training. He co-founded the company in 2024; it launched publicly from stealth on August 12, 2025.[[1]](#cite-id-sqnu2807ym)[[2]](#cite-id-avmmi6kdv3)[[3]](#cite-id-cdosn8pv1s)

## Early Life and Education

Gao studied Honours Business Administration at Ivey Business School at Western University and earned a bachelor’s degree in philosophy from Western University. He was a fellow of the Cansbridge Fellowship from December 2018 to April 2021, including a period in San Francisco.[[4]](#cite-id-doazb23zpd)[[5]](#cite-id-j83e6xjh4o)

Before Mecka, Gao co-founded software for ghost kitchens with Mogen Cheng; the startup was acquired in 2023. He also built Binko Chat, an artificial-intelligence translation chat app, as a side project to help him communicate with his China-born parents. In April 2023, Gao said the app supported more than 18 languages in addition to Chinese and had been featured on CBC.[[2]](#cite-id-avmmi6kdv3)[[4]](#cite-id-doazb23zpd)

## Career

#### Cansbridge Fellowship (2018–2021)

Josh Gao was a Fellow at The Cansbridge Fellowship from December 2018 to April 2021. During this period, his fellowship was associated with the San Francisco Bay Area, which served as a foundation for his later work with [cryptocurrencies](https://iq.wiki/wiki/cryptocurrency).

#### Co-Founder and CEO of Mecka AI (2025–Present)

Gao co-founded [Mecka](https://iq.wiki/wiki/meckaai) with Mogen Cheng, [Jason Chong](https://iq.wiki/wiki/jason-chong), and Duy Nguyen in 2025. None of the four founders had prior professional experience in robotics. Gao said the group studied robotics research and visited AI laboratories before identifying the availability of physical-world data as a challenge for developing general-purpose robots. The founding team also included engineers Kevin Hu and Anthony Tat, as well as Eric Lok.

#### Founding and Financing

In August 2025, [Mecka](https://iq.wiki/wiki/meckaai) announced an $8 million seed round led by NEO investors Suzanne Xie, Ali Partovi, and Connor Ling. Other participants included SV Angel, A-Star Capital, Cade, Timeless, and Offline VC.

In June 2026, Gao announced $60 million in financing intended to support the company's data infrastructure, expansion into additional sectors, and robot deployment. A post he shared described the financing as a $25 million Series A and a $35 million follow-on round. Reporting in September 2026 described a $60 million round led by Framework Ventures, with participation from Menlo Ventures, SV Angel, and Kindred Ventures.

In October 2026, Gao announced a separate $60 million Series B led by Sequoia Capital. He stated that Mecka's total funding had exceeded $120 million and identified NVIDIA, Samsung Next, Microsoft M12, and Qualcomm as strategic investors. He also acknowledged Sequoia's Anas Biad and other investors and advisers.

On August 17, in a post whose year was not specified, Gao stated that [Mecka](https://iq.wiki/wiki/meckaai) had appeared on a Business Insider list covering robotics companies.

#### Data Collection and Hardware Development

​[Mecka](https://iq.wiki/wiki/meckaai) collects recordings of people performing everyday and workplace tasks, processes and labels the footage, and supplies datasets for robotics training. Its services include custom datasets for business clients and an application through which participants record assigned activities.

The company initially focused on human demonstrations, particularly egocentric video recorded from a person's point of view. Gao described this data as a complement to teleoperation and synthetic data and discussed whether information acquired from human demonstrations could be transferred to robots at scale.

In August 2025, Gao said [Mecka](https://iq.wiki/wiki/meckaai) and its partners had assembled several datasets containing tens of thousands of hours of recordings. The company separately reported collecting more than 30,000 hours of video showing activities such as cooking, cleaning, using a deep fryer, and handling household objects. The home-robotics company 1X Technologies was an early partner. According to Gao, 1X used household-motion data from Mecka to generate its World Model earlier in 2025.

Mecka introduced a points program in 2024 to encourage video contributions. The program later shifted toward daily tasks, including activities such as making a bed. Although users speculated that the points might lead to a cryptocurrency token, the company did not launch an official token. Gao said the program was intended to attract participants and noted that associations with [cryptocurrency](https://iq.wiki/wiki/cryptocurrency) could create difficulties when working with corporate clients.

For some tasks, contributors recorded footage using smartphones. For corporate projects, [Mecka](https://iq.wiki/wiki/meckaai) developed recording rigs and shipped them to selected operators. Gao later described work on tactile gloves, sensorized caps equipped with cameras, and a hub connecting accessories and external cameras.

He said existing recording equipment did not reliably measure the force applied during ordinary activities, such as opening a chair or twisting a bottle cap. Wrist-camera footage could help estimate physical contact but did not provide a direct measurement of force.

Gao described data collection as an operational process involving hardware development, product revisions, distribution to thousands of homes and businesses, and infrastructure for uploading large volumes of recordings. The company also encountered fraudulent submissions, including cases in which contributors presented videos taken from the internet as their own recordings. [Mecka](https://iq.wiki/wiki/meckaai) responded with screening procedures, observation at collection sites, incentives for authentic contributions, and changes to data processing.

Gao stated that the cost of collecting robotics data included contributor compensation, hardware, operational expenses, and quality verification.

#### Research and Evaluation

In March 2026, Gao described Mecka's contribution to EgoVerse as providing data and tools for collecting aligned human demonstrations and evaluating robot performance. He also referred to work involving action-trajectory data, post-processing, quality checks, and platform tools.

In September 2026, Gao described [Mecka's](https://iq.wiki/wiki/meckaai) WetLabs Benchmark, which evaluated three AI models controlling six-degree-of-freedom robot arms across nine laboratory tasks. Each model was given 20 attempts per task. He reported success rates of 25.6% for Astra and 23.3% for Opus 5.5.

Of 192 completions declared by the models, human evaluators classified 79 as partial completions and 10 as failures. Gao discussed difficulties involving delicate object handling and instances in which models reported completing tasks that had not been finished. He also addressed the role of human evaluation in assessing task completion.

He reported mean estimated API costs per attempt of $8.64 for Astra and $2.39 for Opus 5.5. According to his figures, Opus achieved a success rate close to Astra's at approximately 3.6 times lower cost.

#### Deployment and Commercial Applications

Gao has described [Mecka's](https://iq.wiki/wiki/meckaai) planned activities as extending beyond dataset production to include a platform for gathering sector-specific data, post-training models, and supporting enterprise adoption of robotics. He compared this approach with a robotics systems integrator that continues to improve a robot's capabilities after deployment.

He has also discussed the use of specialized robotics models for tasks requiring consistent performance within defined operating conditions.

Regarding commercial applications, Gao cited industrial cleaning and hotel service or delivery robots as existing uses, particularly in China and Southeast Asia. He expected applications involving constrained manipulation, including pick-and-place operations and packaging, to expand.

Gao said Mecka had worked on laundry folding for more than a year and a half and expected the task to be solved in the near term. He described residential environments as more difficult for robot deployment because of safety concerns, including the possibility of a robot falling on a child. He also stated that a continuous training process in which deployed robots generate failure data and corrections remained some way off. [\[1\]](#cite-id-sqnu2807ym)  [\[2\]](#cite-id-avmmi6kdv3)  [\[3\]](#cite-id-cdosn8pv1s)  [\[4\]](#cite-id-doazb23zpd)  [\[5\]](#cite-id-j83e6xjh4o)  [\[6\]](#cite-id-6jjsgqys51)  [\[7\]](#cite-id-lbd1ii5h22) 

## Interviews

#### Robotics Data and Physical AI #01

On October 20, 2026, Josh Gao, founder and CEO of [Mecka AI](https://iq.wiki/wiki/meckaai), discussed robotics data collection, model training, and deployment in an interview hosted by Sonya Huang on Sequoia Capital's YouTube channel.

[YOUTUBE@VID](0idrTKCDhuE)

Gao explained that robotics development differs from large language model training because physical interaction data must be collected through dedicated processes rather than obtained primarily from publicly available internet content. [Mecka AI](https://iq.wiki/wiki/meckaai) collects, processes, and licenses datasets for robotics model training, including first-person recordings of human activities and sensor measurements. The company also develops equipment such as camera-equipped headwear and tactile gloves to capture visual information and physical interactions.

According to Gao, robotics data production involves hardware development, contributor recruitment, data transmission, processing, and quality control. He noted that robotics datasets remain smaller than those available in autonomous driving, where data has been accumulated through vehicles operating in real-world environments over extended periods. He also observed that robotic manipulation presents additional training requirements because it involves numerous degrees of freedom and physical forces.

Gao described a training approach that combines different data sources. He outlined a potential mix of approximately one million hours of human data, tens of thousands of hours of robot demonstrations, and smaller quantities of teleoperation and task-specific data. He also identified simulation and tactile sensing as additional inputs for training systems to perform physical tasks.

Regarding deployment, Gao argued that robotics capabilities would develop incrementally as systems become suitable for specific commercial applications. He identified industrial operations, packaging, cleaning, and hotel services as examples of tasks that could be addressed before robots achieve broader capabilities in less predictable environments. He also anticipated that general-purpose robotics models would coexist with specialized models designed for particular tasks and reliability requirements.

Gao further discussed the potential expansion of robotic deployment and [Mecka AI's](https://iq.wiki/wiki/meckaai) possible role in supporting implementation within business environments. He suggested that the company could extend its activities beyond data provision to include data collection and model adaptation for specific operational settings. In discussing the longer-term outlook, Gao predicted that the global number of robots could exceed the human population by multiple orders of magnitude within a decade and cited asteroid mining as an example of a possible future application. [\[8\]](#cite-id-fDkhARXYnnhPlwux) 

#### Robotics Training Data and Research #02

In an interview on the Training Data podcast, Gao estimated that robotics training datasets comprised millions of hours across egocentric video, UMI data, and teleoperation. He outlined a possible distribution of approximately one million hours of human egocentric video for basic physical understanding and behavior cloning, tens of thousands of hours of UMI data for intermediate manipulation, and a smaller quantity of task-specific teleoperation data to address remaining performance gaps.

He said tactile data, simulation, and world models had not yet been incorporated into this combination at scale but expected them to become part of future training processes. He also discussed methods for using internet video to predict physical events and translate them into action trajectories.

Gao has described the availability of training data as a constraint on robotics development. He argued that manipulation requires coordinating numerous degrees of freedom and that existing datasets remain insufficient for many tasks. He compared robotics data collection with the more than decade-long effort to gather data for autonomous driving, which involves fewer control dimensions.

He also observed that teleoperation facilities focused on repetitive warehouse activities could eventually reach a point at which collecting more similar examples no longer improved model performance.

According to Gao, human demonstrations can provide robots with information about physical interactions and basic behavior before large robot fleets become widespread. However, he has stated that egocentric video alone cannot supply all the information required for manipulation. He expects force measurements to play a larger role as robotics models combine visual and physical observations with outputs such as actions, speech, and control signals.

Gao has also discussed the possible use of smell as an input for robotics systems, while describing olfactory sensing as an early-stage area. Drawing a comparison with AlphaGo, he suggested that robotics could eventually combine human demonstrations with robot self-play. He noted that existing simulators do not yet reproduce the full complexity of the physical world. [\[1\]](#cite-id-sqnu2807ym)