NomadicML builds software that turns autonomous vehicle and robotics footage into searchable training data. The NomadicML funding news is that the San Francisco startup has raised an $8.4 million seed round to attack a very real bottleneck: companies collect absurd amounts of fleet video, but a lot of the most useful footage still sits in archives because humans can’t review it fast enough. Founded in 2024 by CEO Mustafa Bal and CTO Varun Krishnan, the company was built around a problem they kept running into in earlier jobs at places like Snowflake and Lyft.
What is NomadicML and how does it work?
NomadicML is a motion-video analysis platform for autonomous systems. Customers upload footage through a web app or use a Python SDK. The system analyzes the video, detects motion events, flags possible safety or operational issues, and returns a list of events with AI-written analysis and recommendations. Nomadic frames this as semantic video understanding for fleets and robots. It also targets compliance teams and autonomous systems builders.
The query layer is the more interesting part. A user can ask for a specific scenario in plain English the docs show an example for finding every instance of an ego vehicle straddling two lanes and the platform searches the footage for matching events. In the TechCrunch example, that could mean isolating every time a vehicle went under a certain bridge. Or every case where a car moved through a red light because a police officer waved it on.
That’s why Nomadic keeps insisting it isn’t just another annotation tool. Krishnan described it as an “agentic reasoning system” that uses multiple models to understand what’s happening in context, then find the right clips. The end product is a structured dataset that teams can use for monitoring and compliance checks. It also feeds evaluation and reinforcement learning workflows instead of leaving raw video as a giant archive.
The roadmap already goes beyond generic search. Nomadic is building tools that can reason about the physics of lane changes from camera footage. It’s also building tools that estimate robot gripper positions more precisely in video, with lidar and cross-sensor fusion as the next step. That matters because the hard part in physical AI usually isn’t getting more footage. It’s making that footage useful.
Who founded NomadicML and why are they building it?
The founding story
NomadicML was founded in 2024 by Mustafa Bal and Varun Krishnan, who met as Harvard computer science undergrads. Bal is CEO. Krishnan is CTO. Their pitch is simple: both had seen the same data problems show up over and over in previous roles, especially around finding rare but useful events inside huge video collections.
Why these founders fit the problem
Bal’s background is unusually relevant for a company trying to industrialize machine-learning infrastructure. Before NomadicML, he worked on applied ML and platform engineering at Snowflake, and an AutoSens profile says he led ML optimizations there that saved more than $10 million annually. Before that, he worked at Microsoft on deep-learning acceleration libraries including DeepSpeed, ONNX Runtime, and PyTorch. He also studied computer science and linguistics at Harvard.
Krishnan brings the autonomy side. Search results on his background tie him to Lyft in data science and ML roles, and TechCrunch reported that TQ Ventures partner Schuster Tanger called out his status as an international chess master ranking him at No. 1,549 globally. It’s a quirky detail. But it also tells you what investors are buying here: technical obsession, not sales polish.
The team is still small roughly a dozen engineers and Krishnan has said all of them have published scientific papers. That doesn’t guarantee anything. Still, it fits the company’s angle. Nomadic isn’t selling labor-heavy annotation services first. It’s trying to sell infrastructure that lets autonomy teams describe a hard scenario and let the system do the digging.
Early traction and the seed round
The product is already in use. Nomadic says customers include Zoox, Mitsubishi Electric, Natix Network, and Zendar. Zendar VP of Engineering Antonio Puglielli said the tool helped the company scale faster than outsourcing would have, and that Nomadic’s domain expertise stood out against alternatives.
On Tuesday, March 31, 2026, the company announced an $8.4 million seed round at a $50 million post-money valuation. TQ Ventures led the deal. Pear VC and Jeff Dean also participated. Nomadic says the money will go toward bringing on more customers and refining the platform. The startup also won first prize at NVIDIA GTC’s pitch contest in March 2026.
How NomadicML compares with Scale, Kognic, Encord, and NVIDIA
The old-school alternative is still brutally manual: people watch footage, fast-forward it, tag what they find, and hope they didn’t miss the weird stuff. That breaks pretty quickly when fleets are generating millions of hours of video.
The modern alternative is a growing cluster of data infrastructure vendors. Scale’s Nucleus lets teams organize datasets and query specific attributes. It also centralizes labeled and unlabeled data and builds end-to-end data pipelines for autonomy programs. Kognic is more specialized for autonomous driving and robotics, with native camera, lidar, and radar annotation, sensor fusion, 90-plus automated quality checks, and more than 100 million annotations delivered. Encord comes from a broader multimodal angle, pushing search and filtering. It also handles annotation across images, videos, text, audio, documents, and lidar in one interface. NVIDIA, meanwhile, is attacking the stack from the model side with Alpamayo, its open family of reasoning-based autonomous vehicle models and simulation tools.
Nomadic’s wedge is narrower and, honestly, sharper. It’s betting that AV and robotics teams don’t want a generic labeling dashboard as much as they want a system that can reason over motion video, find long-tail events from natural-language prompts, and feed those events straight back into training and evaluation. That’s a more opinionated product. It could also be the right one if physical AI teams keep drowning in unindexed footage.
Why does NomadicML funding matter right now?
This round matters because it gives Nomadic room to move from “helpful workflow tool” into core autonomy infrastructure. Bal has already pointed to new modules for lane-change physics and robot gripper localization, plus future work on lidar and multimodal sensor fusion. If those pieces land, the company stops being just a smarter search layer on top of video. It starts to look more like a data operating system for physical AI teams.
TQ Ventures’ thesis is also pretty clear. Tanger’s argument was that autonomous vehicle companies shouldn’t burn engineering time building this category in-house any more than Salesforce should build its own cloud or Netflix its own delivery network. That’s the bet behind the NomadicML funding round: specialized data infrastructure becomes its own layer, and the winners are the startups that stay focused on the robot rather than the plumbing around it.
For customers, the immediate value is practical, not philosophical. If a team can instantly pull every clip involving a rare driving instruction, a strange bridge geometry, or a robot hand missing a grasp, it can tighten evaluation loops and generate better RL or compliance datasets without sending humans back through weeks of footage. That kind of speed-up is boring on paper. In autonomy, it’s the whole job.
How big is the physical AI market NomadicML serves?
The demand story behind this company is getting bigger fast. Grand View Research estimates the global physical AI market was worth $81.64 billion in 2025, should reach $110.77 billion in 2026, and could grow to $960.38 billion by 2033. North America held the largest share in 2025, helped by heavy spending on robotics, autonomous mobility, and intelligent systems.
The hardware base is already massive. The International Federation of Robotics says 542,000 industrial robots were installed globally in 2024, the second-highest annual total on record, and the total installed base reached 4.664 million units. Pair that with a data-labeling market that Mordor Intelligence sizes at $2.32 billion in 2026, and you get the setup for companies like Nomadic: more machines, more sensors, more edge cases, and a lot more pressure to turn raw data into usable training signals.
Final take on NomadicML funding
A lot of startups say they’re building for robotics or autonomy when they’re really just wrapping a model around a spreadsheet problem. Nomadic isn’t doing that. It’s going after the ugly middle layer between raw fleet footage and model improvement, which is exactly where many physical AI teams still lose time. The next thing to watch after this NomadicML funding round is whether the company can expand from camera-first video reasoning into lidar and true multimodal workflows without losing the speed that makes the product appealing in the first place.
Read how Bachatt Funding Accel Backs $12M AI Wealth Push is accelerating AI-powered wealth management and smarter investment tools.
FAQ
What is the NomadicML funding round?
NomadicML raised an $8.4 million seed round announced on March 31, 2026. TQ Ventures led the deal, with Pear VC and Jeff Dean participating, and the startup said the round values the company at $50 million post-money.
How does NomadicML work for autonomous vehicle and robotics teams?
It works by turning raw motion video into structured, searchable data. Teams can upload footage through a web platform or Python SDK. Then they can use the system to detect events, review AI-generated analysis, and run natural-language searches for rare scenarios that would be painful to find by hand.
Who are Mustafa Bal and Varun Krishnan?
Mustafa Bal is NomadicML’s co-founder and CEO, with prior ML infrastructure work at Snowflake and Microsoft. Varun Krishnan is the co-founder and CTO, previously tied to Lyft in data science and ML roles, and he’s also an international chess master.
Is NomadicML a data labeling startup or a physical AI company?
It’s really a mix of both, but the cleaner description is physical AI data infrastructure. Nomadic sits in the workflow between massive autonomous-system datasets and the teams training or evaluating models. That puts it next to annotation and curation vendors like Scale, Kognic, and Encord while also riding the broader growth of robotics and autonomous systems.
