Mecka AI raises $60M (Sequoia, Nvidia) to film the real world — and already posts a $100M run-rate

Mecka AI raises $60M (Sequoia, Nvidia) to film the real world — and already posts a $100M run-rate

🔎 Physical data becomes the oil of robotics

LLMs had their internet: trillions of words, free, already digitized, ready to be scraped. Robots, on the other hand, have nothing. A humanoid doesn't learn to brew coffee by reading Wikipedia — it needs hours of real-world movements, with the force applied, the finger pressure, the exact timing. Mecka AI has decided to turn this void into a business. And the market has just put a price on it.

On October 7, 2026, the startup announces a $60M Series B led by Sequoia Capital, with Nvidia, M12 (Microsoft's corporate fund), Qualcomm Ventures and Samsung joining as new investors, according to WebProNews. Around the table: Tony Xu (CEO of DoorDash), Frank Slootman (ex-ServiceNow, ex-Snowflake) and Milan Kovac (ex-Tesla Optimus) as business angels. When a cast like this joins a round led by Sequoia, we're no longer looking at a bet. We're looking at a signal.

The model itself can be summed up in one sentence: pay people to film their everyday tasks — coffee, cycling, laundry — using body sensors and smartphones. From that comes EgoVerse, a dataset of 1,362 hours of human demonstrations. And the company already claims a $100M run-rate since June 2026. My conviction after crunching the numbers: the robot war won't be won on motors. It will be won on data.

Key takeaways

• The round: $60M Series B announced on October 7, 2026, led by Sequoia Capital; Nvidia, M12, Qualcomm Ventures, and Samsung join the cap table.
• The valuation: close to $500M according to a September 2026 TechCrunch report (not officially confirmed); ~$128M raised in total since the $8M seed (August 2025).
• The business: claimed run-rate of $100M as of June 2026, $300M projected by end of 2026 (figures provided by the company).
• The product: egocentric data (movement, force, pressure, timing) captured from humans equipped with body-worn sensors and smartphones.
• The dataset: EgoVerse — 1,362 hours, 80,000 episodes, ~2,000 tasks, more than 2,000 demonstrators.
• The thesis: physical data, not hardware, is the new battleground in robotics.

Recommended Tools

This tour can't be understood in isolation: it fits into a "data for robots" ecosystem in full ferment. Here are the players to know (October 2026):

My advice: if you're just starting out, begin with LeRobot and Isaac Sim — free and sufficient for prototyping. Proprietary data like EgoVerse is negotiated at the enterprise level, and budgets run into hundreds of thousands of dollars per contract.

Who Is Mecka AI, and Why Sequoia and Nvidia Came to the Table

Mecka AI is a company founded in 2024, based in New York and Toronto, that sells roboticists what they cannot produce themselves at scale: real-world data, captured from humans going about their everyday lives (TechFundingNews). Its approach is described as "egocentric" — the camera is on the person, not on the robot — as opposed to teleoperation, where an operator pilots a machine in a lab (TechJuice). The stated ambition: to become the Scale AI — or the Mercor — of robotics.

The funding history tells a story of rapid ramp-up:

In total: roughly $128M raised in a little over a year. Framework Ventures, Kindred Ventures, and Neo return in every round (CryptoBriefing). The valuation being floated — close to $500M according to TechCrunch (September 2026) — is, incidentally, conservative for a company claiming a $100M run-rate. Note the journalistic nuance: the exact post-money has never been officially confirmed.

The most telling part is the cast of angels. Tony Xu runs DoorDash, i.e., physical-world logistics at planetary scale. Milan Kovac led Optimus at Tesla, i.e., the industry's most advanced humanoid data pipelines. Frank Slootman grew ServiceNow and then Snowflake, i.e., selling infrastructure into the enterprise. This cap table is not a prestige guest list: it's a map of Mecka's future customers.

And the timing is no accident. Humanoids have been leaving the labs and entering factories since 2025-2026; every manufacturer hits the same wall: there is no massive physical corpus to train their control policies. Mecka arrives exactly at the moment when demand is exploding and supply is nearly nonexistent.

The "human camera" model: paying people to live in front of sensors

Mecka pays volunteers to film their everyday gestures — making coffee, riding a bike, folding laundry — equipped with body sensors and smartphones. The capture isn't limited to images: the sensors record movement, force, pressure, and timing, four dimensions completely absent from web text and image data (The AI Insider).

The technical feat comes down to one number: sub-centimeter hand pose tracking, achieved outside the lab, with consumer-grade hardware. Internally, a video lab dedicated to motion tracking and 3D reconstruction cleans and structures the raw stream (CryptoBriefing). In other words: the data is dirty on the way in, clean on the way out. That's precisely where the value hides — anyone can film, almost no one can industrialize the cleaning.

Why prefer this approach to teleoperation? Because teleoperation produces lab data: one operator, one robot, a controlled environment, an exorbitant cost per useful hour. The egocentric approach captures the true distribution of human gestures, in real kitchens, with thousands of different hands, body types, and movement patterns. The human is the cheapest robot on the market: it rents by the task, not by the engineer-hour.

This logic isn't new — Figure AI is already turning human videos into training data — but Mecka is industrializing it with a network of paid contributors and dedicated collection infrastructure. It's the Mercor playbook applied to the physical world: pay people to produce rare expertise, then resell it in structured form. The difference between an intuition and a business is the pipeline.

The flywheel is classic but formidable: more contributors, more tasks covered; more tasks, better customer models; better models, more enterprise contracts; more contracts, more money to pay contributors. Each turn of the merry-go-round raises the cost of entry for competitors.

EgoVerse: what the dataset actually contains

EgoVerse today amounts to 1,362 hours of human demonstrations, 80,000 episodes, roughly 2,000 distinct tasks, and more than 2,000 demonstrators (WebProNews, October 2026). Here's the snapshot:

Put these numbers in perspective. Open X-Embodiment, Google DeepMind's big aggregate (2023), federates more than a million trajectories from 22 robot types — but mostly from teleoperation and simulation, with heterogeneous modalities from one lab to the next. EgoVerse is a hundred times smaller in raw volume, but far denser in information: force, contact, pressure, absolute 3D positioning. In physical data, fidelity beats volume.

The number of demonstrators matters as much as the hours. Two thousand different bodies, two thousand ways of holding a cup: that's what keeps a model from learning "the" way of doing things instead of "a" way that's robust to variation. It's the killer argument against lab datasets, captured from a handful of trained operators.

The most promising point remains the human-to-robot transfer study conducted on EgoVerse (The AI Insider, October 2026): demonstrating that a model trained on human gestures actually transfers to a robotic arm would change the economics of the entire industry. That's exactly the hypothesis Sequoia's money is coming to fund.

My take: 1,362 hours is ridiculous if you judge it against LLM corpora. It's enormous if you judge it against what exists in instrumented physical data. The right comparison isn't "tokens versus hours," it's "instrumented hours versus instrumented hours" — and on that playing field, Mecka is ahead.

$100M run-rate: what the numbers say — and don't say

Mecka claims a $100M run-rate as early as June 2026 and projects $300M by the end of the year. In other words: the most recent known month, annualized, would be worth $100M. If the first figure is solid, the second remains a commercial target, not an established fact. The two deserve different treatment.

To gauge the scale, look at what run-rate has become in AI: the new scoreboard. Anthropic first overtook OpenAI with $30 billion in run-rate (our analysis), before estimates climbed to $47 billion and put it in the lead of the race (Anthropic hits $47 billion in revenue run-rate). Mecka is two orders of magnitude smaller — but for a company founded in 2024 that sells physical-world data, this monetization speed is remarkable.

Three caveats are in order, though. One, run-rate is not cashed revenue: it's signed contracts, sometimes spread over years. Two, the Series A was disbursed in tranches — November 2025 then June 2026 — a structure that suggests negotiated milestones, and therefore growth to be confirmed quarter by quarter. Three, the projected $300M comes from the company itself; no public audit validates it to date.

My blunt take: even at half these figures, Mecka is monetizing faster than any humanoid manufacturer. The reason is structural — it sells the shovels and picks of the robotics gold rush. Tool sellers get paid before the gold prospectors, and the margins of a data platform, once the pipeline is built, have nothing to do with those of a hardware maker.

Why physical data — not hardware — is the new battleground

Because hardware is becoming commoditized, and data is not. Chinese humanoids now sell for under $20,000 (Unitree G1, since 2024), and every quarter brings a new chassis with an impressive spec sheet. When the robot becomes a commodity, what differentiates is no longer the machine: it's the model that drives it. And that model is hungry for data that nobody has yet.

That's where Nvidia's entry takes on its full meaning. Nvidia doesn't just sell GPUs: it sells the complete AI infrastructure — simulation with Isaac, training, deployment. Investing in Mecka means securing the data upstream that makes its downstream compute indispensable. The company has already demonstrated its method: Nvidia snapped up the soul of Groq for $20 billion (our analysis), locking in a strategic infrastructure building block along the way. For robotics data, this time it prefers taking an equity stake over an acquisition — faster, cheaper, just as effective for locking down the ecosystem.

The parallel with software is striking. Meta injected $14.3 billion for 49% of Scale AI in June 2025, valuing the data company at around $29 billion. Mercor, the other star of expertise data, was valued at $10 billion in early 2026. Training data is already worth its weight in gold on the software side; Mecka wants to play exactly that role on the robotics side (TechJuice). If the thesis holds, today's $500 million valuation will look like a bargain in three years.

My conviction: whoever controls the upstream of physical data sets the terms for the entire downstream robotics economy. Robot manufacturers will negotiate, labs will subscribe, and future robotics foundation models will train on what Mecka — and a handful of rivals — have captured. It's a bottleneck position, the most profitable one there is in tech.

The Vulnerabilities: Competition, Data Quality, and Sensor Dependence

The Mecka model is appealing, but it rests on three assumptions that can crack. Reviewing them means honestly assessing the risk rather than being swept along by the storytelling.

First vulnerability: competition comes from four directions at once. Teleoperation from the big labs, synthetic simulation data (Nvidia Isaac at the forefront), mining of existing videos — Figure's approach — and above all, manufacturers collecting in-house. Tesla feeds Optimus with its own pipelines, and Milan Kovac, Mecka's angel investor, knows this better than anyone: he comes from there. If every humanoid manufacturer builds its own private data farm, the third-party market shrinks to the players who can't afford to do it themselves.

Second vulnerability: quality and compliance. Consumer-grade sensors produce noise; self-reported tasks produce bias; and filming inside homes raises consent and privacy questions that the GDPR and its American equivalents will not easily forgive. The in-house video lab and sub-centimeter tracking partially address the technical problem — not the legal one, which could slow the European expansion of the contributor network.

Third vulnerability: the pivot to deployment. The raise also funds a robotic deployment infrastructure (The AI Insider). Going from data seller to deployment partner means changing trades: selling data APIs and operating robot fleets have nothing in common, neither in skills nor in sales cycles. The ambition is commendable — it turns Mecka into a complete platform — but the execution remains to be proven, and companies that spread themselves thin across two trades often pay a heavy price for it.

What this means for developers and businesses

Three concrete consequences, depending on where you sit in the value chain.

If you're a researcher or developer: keep an eye on EgoVerse opening up. A public slice of those 1,362 instrumented hours would be worth more than petabytes of raw YouTube video, which comes without sensors. In the meantime, the Hugging Face LeRobot + Isaac Sim duo covers most prototyping needs for free — start there before dreaming of enterprise contracts.

If you're thinking about on-device models: robots won't be offloading their control policies to the cloud, if only for latency and privacy reasons. The embedded constraint favors compact, efficient models — the same logic as for the best LLMs to run locally. Open, self-hostable models like Kimi K2.6 (Self-host) or GLM-5 (Reasoning) offer a glimpse of what will run tomorrow inside a humanoid's torso. And on the video understanding side, multimodal giants like Gemini 3.1 Pro or GPT-5.5 already digest hours of footage: the gap between "understanding a video" and "piloting a robot" is closing as datasets like EgoVerse bridge perception and action.

If you're a business: data partnerships get negotiated early, not once the dataset is saturated with your competitors' needs. Enterprise contracts signed today set the terms for the years that follow. And if you're launching your own data collection project or simply want to publish your robotics experiments to get noticed, a well-built showcase site is enough to get started — Hostinger does the job for a few dollars a month (prices observed October 2026, check hostinger.com).

❌ Common Mistakes

Three pitfalls come up again and again in the commentary and superficial analyses of this funding round. Avoiding them already means analyzing better than 90% of the tech press.

Mistake 1: confusing run-rate with collected revenue

The run-rate annualizes the most recent month: it is neither realized revenue nor available cash. A company can display $100M in run-rate with only $40M actually invoiced and collected. The solution: systematically cross-check with the financing structure — the Series A tranches indicate negotiated milestones, meaning growth that must be confirmed quarter by quarter before declaring victory.

Mistake 2: judging EgoVerse with an LLM yardstick

Comparing 1,362 hours to the trillions of tokens of LLMs makes no sense: these are not the same units of value. One hour of instrumented video (force, pressure, 3D) is worth thousands of hours of raw video. The solution: evaluate a physical dataset on task coverage, sensor richness, and evidence of transfer to the robot — never on raw volume.

Mistake 3: believing Mecka has no competitors

The narrative that "Mecka is alone in robotics data" is false: teleoperation, simulation, video mining, and manufacturers' in-house collection (Tesla foremost among them) are four competing, well-funded avenues. The solution: track the enterprise contracts actually signed — that's the only indicator that separates the approaches, not the funding announcements.

❓ Frequently Asked Questions

Short, sourced answers to the questions that come up most about this funding round.

What is Mecka AI?

A startup founded in 2024, based in New York and Toronto, that produces training data for robotics. Its model: paying people to film everyday tasks with body sensors and smartphones, generating egocentric data (movement, force, pressure, timing) organized into the EgoVerse dataset.

Who invested in Mecka AI's Series B?

Sequoia Capital is leading the $60M round announced on October 7, 2026. Nvidia, M12 (Microsoft's fund), Qualcomm Ventures and Samsung join as new investors; Framework Ventures, Kindred Ventures and Neo are returning. Tony Xu (DoorDash), Frank Slootman and Milan Kovac (ex-Tesla Optimus) are participating as angel investors.

What is Mecka AI's valuation?

Close to $500M, according to a September 2026 TechCrunch report picked up by the trade press. Important caveat: the post-money valuation has not been officially confirmed by the company, which does not comment on the round's figures. So caution is warranted on this number, even though it is the consensus in media coverage.

What is the EgoVerse dataset?

Mecka's flagship dataset: 1,362 hours of human demonstrations, 80,000 episodes, roughly 2,000 tasks and more than 2,000 demonstrators, captured from an egocentric approach with motion, force and pressure sensors. It is accompanied by a human-to-robot transfer study — the condition for its real value to robotics.

How do you become a paid demonstrator for Mecka AI?

The company pays contributors to film everyday tasks (making coffee, cycling, doing laundry) with sensors and a smartphone. Expanding the contributor network is one of the Series B investment areas; sign-ups go through the official contributor program on the company's website.

Is Mecka AI the "Scale AI of robotics"?

That's the stated ambition (TechJuice), and the similarities are real: rare training data, sold to enterprise customers, in an exploding market. The major difference: Scale AI was born from text and images, while Mecka captures the physical world, with unprecedented sensor, logistics and compliance constraints.

✅ Conclusion

Mecka AI's $60M raise makes it official: the tide has turned — physical data has become the strategic asset of robotics, and both Sequoia and Nvidia have just bought a substantial stake in it at a $500M valuation. Keep a close eye on the upcoming enterprise contract signings — that's where, and nowhere else, the thesis will either be confirmed or fall apart.