Analysis
XDOF is in late-stage talks for a Series B led by 8VC at a valuation of roughly $1.2 billion, TechCrunch reported on September 4. Terms are not final and the round size has not been disclosed.
The company collects real-world teleoperation data used to train general-purpose robots -- operators wear rigs or drive robot arms through physical tasks, and XDOF turns the resulting streams into labeled training sets. It sells the pipelines, collection tooling and annotation systems to AI labs and robotics companies that would rather outsource the messy physical-data problem than build a collection operation in-house.
XDOF was founded in 2024 by UC Berkeley researchers Philipp Wu, now CEO, and Fred Shentu, CTO. Wu's doctoral work centered on how robots learn from large datasets, and identified the shortage of real-world manipulation data as the thing holding the field back. The company raised a $70 million Series A in June 2026 led by Thrive Capital, with Andreessen Horowitz, Lux Capital and Spark Capital participating -- a round that already looked aggressive at the time. It reports annualized revenue approaching $50 million across 20 customers, several of them frontier labs.
“XDOF was founded in 2024 by UC Berkeley researchers Philipp Wu, now CEO, and Fred Shentu, CTO.”
That revenue figure is what makes the mark defensible. At roughly $50 million annualized, a $1.2 billion valuation is about 24x revenue -- expensive, but not in the same universe as the 70x-plus multiples showing up in enterprise AI rounds this month. The comparison set:
- Scale AI -- data-labeling business effectively repriced when Meta paid $14.3 billion for a 49% stake in mid-2025, implying a $29 billion valuation
- Micro1 -- grown fast in expert data, chasing the same enterprise AI training buyers
- Mecka AI -- newer entrant competing directly for robotics data customers
- Physical Intelligence and Skild -- raised at multi-billion valuations to build the robot foundation models themselves; XDOF sells the fuel to all of them
Pulse has previously covered the robotics funding surge that set the stage for deals like this one -- the category topped $18.8 billion in H1 2026 alone.
The Concentration Risk
The risk is customer concentration wearing a different hat. Twenty customers producing $50 million means a handful of frontier labs are most of the revenue, and those labs are actively building internal data-collection fleets -- Tesla, Figure and 1X all run their own. Data vendors get squeezed when their largest buyers vertically integrate, which is exactly the arc Scale AI's enterprise business followed.
For seed investors, the pattern to note is the compression: stealth to $70 million to $1.2 billion inside a year, off real revenue rather than a demo. That is now the fastest path to a unicorn mark in 2026, and it runs through selling picks and shovels to labs rather than building an application.
Why the Data Can't Be Scraped
The market XDOF sells into barely existed 24 months ago. Robot foundation models -- Physical Intelligence's pi-series, Skild's general policies, Google DeepMind's RT and Gemini Robotics work -- all consume demonstration data measured in thousands of hours per task family, and unlike text, none of it can be scraped. Every hour has to be produced by a human operating hardware, which turns data into an operations business with recruiting, facilities and quality control rather than a scraping pipeline.
That operational character is also the moat argument. Wu and Shentu have said the hard part is not the rigs but the throughput: keeping operator error rates low, catching sensor drift, and delivering datasets that transfer between robot embodiments. Buyers who tried building it internally generally underestimate the failure rate of raw teleoperation footage, which is why 20 customers were willing to pay for it in year two.
Watch whether 8VC's round closes at the reported number or resets. Late-stage talks at a tripled valuation three months after a Series A are exactly the deals that get re-cut in diligence.
