What Frontier Labs Actually Buy When They Buy Data
Nobody buys 'data.' They buy specifications, delivery formats, QA guarantees, and the right of first refusal on the next capture cohort. Understanding the actual purchase order is where vendors either win or lose the account.
The word 'data' does a lot of work in press releases. In the actual purchase orders that frontier labs send to data-infrastructure companies, the term is much more specific, and understanding that specificity is the difference between a vendor that closes deals and one that keeps losing to a competitor for reasons it cannot quite name.
What is actually on the purchase order
A task specification, negotiated in detail, what the operator does, what the acceptable variation is, what counts as a failed session. A delivery schedule with hard weekly milestones. A data format specification, file structure, tensor shapes, coordinate frames, force units, timestamp conventions, that plugs into the lab's existing training pipeline without a translation layer. A QA guarantee, what fraction of delivered hours is guaranteed to pass the lab's internal quality gate. And typically, a right of first refusal or exclusivity clause on the specific task family.
Price is one line item among many. Vendors that lead with price and treat the rest as boilerplate lose deals to vendors that lead with the format and QA spec and treat price as an outcome.
Where wearable and teleop vendors compete on the same terms
Both have to hit the same delivery schedule. Both have to meet the same QA bar. Both have to write to a format the lab's pipeline can ingest. The underlying capture technology is not the deciding factor at the contract level, the operational maturity of the vendor is.
This is why the vendors that win are the ones with real operations teams, not just great hardware. A brilliant sensor stack backed by a chaotic delivery process loses every RFP it enters. A merely good sensor stack backed by an operations team that never misses a milestone wins deal after deal.
Where they differentiate
On tasks that fit cleanly into a cell, teleop wins on native action space and shorter time-to-first-hour. On tasks that need in-situ capture, embodiment diversity, or fingertip-level force, wearable wins. On volume at low cost per hour, wearable wins at scale. On single-embodiment platform-specific tuning, teleop wins.
The mature buyer knows this and buys from both. The immature buyer picks one and regrets it a year later.
Building or training robots?
We license manipulation datasets and run custom capture programs. Get in touch to see what fits.
