The Humanoid Data Layer Just Went Commodity: What Open-Sourced Pretraining Means for the $39B Figure Cap Table
For the last twenty-four months the humanoid robotics story has been a foundation-model story. Figure raised at a $39B post-money on the thesis that proprietary vision-language-action models would compound into an insurmountable moat. Apptronik’s $935M Series A and its reported $5.3B valuation followed the same logic. Neura’s $1.4B Series C at a $7B valuation announced earlier this month was priced on a variant of the same claim. The implicit promise: own the model, own the embodiment, own the category.
The last two weeks made that promise harder to hold.
NVIDIA open-sourced the control layer. The GR00T-WholeBodyControl release on NVLabs ships 142,000-plus annotated human motions covering roughly 288 hours of full-body trajectories, with G1 MuJoCo simulation traces, a deployable C++ inference stack, and teleoperation and planning hooks. This is not a research artifact. This is the kind of asset that a humanoid startup would have spent eighteen months and a meaningful portion of its Series B building internally. NVIDIA put it on a public docs page.
Alibaba released a manipulation foundation model. The Qwen-RobotManip paper posted to arXiv on June 17 describes a VLA model built on Qwen-VL, pretrained on a 38,100-hour corpus of open-source robot data and human video, with claims of emergent zero-shot instruction following, error recovery, and cross-embodiment transfer. The technical claims will be debated by the research community. The market signal is unambiguous: the leading Chinese AI lab is positioning manipulation as a foundation-model category, with weights and methodology described in enough detail to reproduce.
An open VLA stack beat a respected closed one on a public benchmark. The Hy-Embodied-0.5-VLA work documented in research walk-throughs on June 19 hit a 90.9% success rate on the RoboTwin 2.0 fifty-task simulation suite, against 65.9% for the previous-generation closed flow-based model, with deployment demonstrated across three different robot morphologies. After three rounds of post-training with a preference-optimization variant, the same stack hit 99% on high-precision tasks like USB insertion. The numbers will be re-tested, but the architecture, delta-chunk action representation, and inference cache strategy are public.
What that means for the cap table
Three of the four most-cited reasons to pay a private-market premium for a humanoid leader are now under pressure.
The data moat is shrinking. A founder pitching a humanoid round in 2025 could credibly claim that proprietary teleoperation data and proprietary motion capture were a multi-hundred-million-dollar replacement cost. With GR00T-WBC’s 288 hours of motion data and Qwen-RobotManip’s 38,100-hour pretraining corpus public, the replacement cost is a much smaller number, and the same number is available to every funded competitor.
The control-stack moat is shrinking. A C++ inference stack that runs whole-body control with teleoperation and planning hooks is a year of senior robotics-engineering work. NVIDIA gave it away. The remaining moat is hardware integration, not algorithm.
The model moat is shrinking. A 90.9% benchmark from a publicly described VLA stack does not mean the closed Figure or Apptronik models are inferior. It does mean the floor under any closed model just rose. A buyer of a private humanoid round in mid-2026 is no longer paying for a possibility that a competitor cannot match the model. The buyer is paying for execution at the deployment layer.
What is left on the moat side is the part that has always been the hardest to fund and the most expensive to verify: factory output, customer deployment data, and supply chain. Figure’s claim of one robot per hour off the BotQ line, Apptronik’s reported pilot deployments with Mercedes-Benz and GXO, and Neura’s articulated plan for real-world training environments are now the assets that determine valuation, not the model card.
Four questions for any humanoid round priced in the next six months
This is the diligence frame a community member should put around a private allocation in the humanoid stack, the simulator stack, or any company whose pitch leans on a proprietary VLA claim.
What is the deployment count and the deployment kinetics? Not pilot count. Not contract count. Robots running production shifts at a customer site, week over week. If the founder cannot answer in those units, the round is mostly priced on the model story that just got commoditized.
What is the marginal cost per additional motion-hour of training data? Pre-GR00T, this was a defensible high number. Post-GR00T, it is a much lower number that anchors to compute and engineer time. A round priced before June 16, 2026 is not directly comparable to a round priced after.
What does the cap table look like if a 2027 entrant ships an open-stack humanoid at half the price? The Hy-Embodied result demonstrated cross-morphology deployment without target-robot data collection. A well-funded second-mover is now a credible threat to specific embodiment bets, even if the leader retains the model edge.
Is the round being priced against the pre-commoditization world or the post-commoditization world? Several humanoid valuations on the public tape were anchored in early 2025 or earlier. The information environment has shifted. Marks struck against the older environment carry more downside than the headline number suggests.
Historical parallel: the 2014 deep-learning commoditization wave
The closest historical parallel is not the autonomous-vehicle cohort. It is the 2014-2016 period when Google released TensorFlow, Facebook released Torch, and ImageNet pretraining went from proprietary to commodity. The CV-startup cohort that had raised on the strength of a model moat repriced violently, while the cohort that had raised on the strength of customer integration and data flywheels held value and compounded. The same pattern is the working hypothesis for humanoid robotics in the second half of 2026 and into 2027.
Educational takeaway
The first investable question in humanoid robotics is no longer ‘whose VLA model is best.’ It is ‘whose deployment pipeline is real.’ AdValorem Research covers AI and robotics as one of our recurring education verticals because the underwriting frame for this category is shifting on a quarter-by-quarter basis and the public information is moving faster than most private-market commentary. The community discussion we host on these structural questions is the kind of read that retail commentary skips. Position before you predict.
Get Weekly Research
Analysis, education, and market intelligence — delivered to your inbox.
Subscribe
Join 586+ members for weekly research. Unsubscribe anytime.
Sources
Want to discuss how these trends connect to our research?
Schedule time with the team to explore these topics further.
This article is informational and educational. It is not an offer to sell or a solicitation to buy any securities. References to AdValorem research verticals describe published education topics, not investment offerings.