Investors have put $60 million behind Mecka AI’s promise to turn petabytes of video into robot training data at a margin. One engineer, four months into the job, owns the number that decides whether the promise holds.
Money is moving toward robots again. In June, Mecka AI, a two-year-old company that sells training data for physical AI, announced $60 million in funding: a $25 million Series A that closed quietly in November 2025, led by Framework Ventures, and a $35 million extension months later. Menlo Ventures, SV Angel and Kindred Ventures joined, along with the roboticist Ted Xiao.
At Mecka, the assumption under those numbers has a name. Basil Yusuf, a software engineer who joined in March from Google, owns the two layers of the stack that set the company’s cost per hour of video: the hardware encoding path and the pipeline orchestration everything else runs on. Investors priced the company against a cost curve. Yusuf keeps the curve where they priced it.
The round fits a pattern. Meta paid more than $14 billion for a stake in Scale AI, which built the labeling workforce behind the language-model boom, on the theory that whoever owns the data layer sets the pace of the field. The same bet is moving to the physical world. Wayve collects driving video at fleet scale. MicroAGI and a crowd of smaller firms chase the same robotics customers Mecka serves. Every one of these businesses rests on a claim almost nobody outside them examines: that a company can ingest, process and verify petabytes of first-person video cheaply enough to sell at a margin.
Mecka, founded in February 2025 and split between Toronto and New York, holds the picks-and-shovels spot in that rush. Its contributors film themselves doing ordinary work, cooking, cleaning, assembling, repairing, and the company turns the footage into the labeled, verified episodes robotics labs train on. Its open dataset, Egoverse, has become a calling card. The product sounds simple. The machinery under it is not.
Yusuf talks about the cost claim the way a structural engineer talks about load, as something the field keeps taking for granted and he is paid never to. “More GPUs is a band-aid, not a strategy,” he said of the reflex answer to processing costs. Teams scale sideways on the same architecture instead of asking whether the architecture fits the work, and the habit holds until the cost curve stops being linear, which at petabyte scale comes sooner than anyone budgets for.
Video is the raw material of physical AI, and video is costly. Before a robot can learn from a person folding laundry or plating a dish, hours of raw footage have to be encoded, split, stored, labeled and verified. The field runs on general-purpose GPUs by default and books the cost as a fixed line item, the way airlines book fuel. Yusuf refused the entry. He moved production video encoding off GPUs onto dedicated video-transcoding ASICs, chips built for one narrow job, and rebuilt the encoded path so the work could survive the move.
The results are the kind that decide whether a data business has margins at all. Encoding an hour of video now costs about a quarter of what it cost on the GPU path. Across the platform the saving compounds to a large multiple, and processing a ten-hour block runs at a fraction of the old price.
What makes the work matter at an industry level is less any single number than where the saving sits. Encoding is upstream of everything. A data company’s labeling models, its verification systems and its delivery pipeline all eat processed video, so a cut in the cost of processing runs through the whole business. A dollar saved at encode time is saved again at every stage below it. Yusuf built the system so the cheaper path is the default rather than an option: an autoscaler treats the ASIC route as primary and holds GPU capacity as automatic spillover, so the saving reaches every job, not a favored few.
He did it fast. Yusuf joined on March 14 and ran a staged rollout, phasing the older GPUs out as primary workers over roughly two to three weeks. He arrived at a company moving fast, almost everything breaking at once, everyone working flat out to fix it. Winning over the infrastructure team took one thing: tests, run before the rollout, that settled the question.
The bet has a shape investors recognize from other infrastructure businesses. Specialized hardware buys margin and costs flexibility, and the trade pays only if the volume arrives. Mecka’s signed contracts say it will.
For buyers, the curve is hard to inspect. A robotics lab signing a data contract sees a price per hour and a delivery schedule, not the encode path behind them. Two vendors can quote the same number and run on cost structures that pull apart as volume grows. The difference surfaces later, at renewal, when one of them has to reprice and the other does not.
The economics matter because the market is about to test them. Robotics companies buying training data are buying someone else’s cost curve. As collection climbs from thousands of hours to millions, a vendor whose costs grow in a straight line watches its margins vanish just as demand lands. “The data business only works if the cost of processing an hour of video keeps falling faster than the price we charge for it,” said Jason Chong, Mecka’s co-founder and chief technology officer. “Basil owns that curve.”
The physical AI wave will mint its own versions of the Scale AI story: big rounds, bigger customers, a race to lock up supply. The quieter race is the one Yusuf runs, against the cost of the pipeline itself. The companies that survive the next phase will be the ones whose cost curves stay sublinear as volume climbs. Investors wrote Mecka a $60 million check on the belief that its curve will. The engineer who has to keep that promise arrived four months ago and went straight to the layer where promises like it hold or break.