In 2025, every AI conversation boiled down to one thing: "I can't get enough GPUs."
But a16z partner Will Bitsky put it bluntly on this podcast. "If 2025's AI zeitgeist was defined by compute constraints and data center buildout, 2026 will be defined by data constraints — and the next frontier in the crusade for data: our critical industries." And that data isn't sitting behind a screen. It's on factory floors and in warehouses.
Everyone thought this was a GPU fight
For the past two years, the AI narrative was simple: whoever locks up the most GPUs and trains the biggest model wins. But this episode, featuring four a16z partners, flips that frame entirely.
Ryan McEntush (the partner covering the electro-industrial stack) put it this way: "Software transformed how we think, design, and communicate. Now it's transforming how we move, build, and produce." Electrification, materials science, and AI are converging for the first time to make the entire physical world controllable by software.
Here's the thing though — the winner of this wave won't be whoever has the best model. It'll be whoever collects "field operational data" first, according to Bitsky. Every truck roll, meter read, maintenance job, and production run in heavy industry is fodder for model training — but until now, there was no industrial vocabulary for capturing, annotating, and pipeline-ing it. The startups filling that gap are becoming the new winners.
The numbers back this up. The cost of collecting high-quality teleoperation data (robot demonstrations captured, labeled, and packaged into a standardized format) fell from roughly $340/hour in early 2024 to $136/hour by Q4 2025 — and for a standard pick-and-place task specifically, it's down to $118/hour as of March 2026. The barrier to entry for building a data asset is dropping fast.
And it's already proving out on the factory floor
If this were just VC optimism, you'd want to filter it out. But this time there are real production-line numbers behind it.
At BMW's Spartanburg plant, two Figure AI Figure 02 humanoids ran for 11 months, loading over 90,000 sheet metal components and contributing to production of more than 30,000 BMW X3s — running 10-hour weekday shifts for roughly 1,250 total operational hours. At Amazon-affiliated fulfillment operations, Agility Robotics' bipedal robot Digit is already moving totes between conveyors and mobile robots without operator intervention, in live production. Closer to home for Korean readers, Hyundai announced plans to build 30,000 Atlas humanoids a year by 2028, prioritizing 25,000+ of them for its own plants, and Figure hit 200 continuous hours of uninterrupted operation in May.
This isn't pilot-stage anymore — the market numbers confirm it too.
| Before | Now | |
|---|---|---|
| Commercial humanoid platforms | 3 (2024) | 12 (2026) |
| Teleoperation data cost | ~$340/hr (early 2024) | $136/hr (Q4 2025) |
| VLA adoption in new deployments | Under 5% (18 months ago) | 40% (2026) |
| Global robotics venture funding | 2024 baseline | $9.4B, +41% YoY in 2025 |
What matters most is that Vision-Language-Action (VLA) model adoption tripled from under 5% to 40% of new deployments in two years. That's a signal that "see, understand, act" robotics has left the lab and entered real production lines. Zoom out further and the physical AI market as a whole is projected to hit $430 billion by 2030 and $1.6 trillion by 2040, spanning nine verticals from industrial automation to autonomous vehicles, healthcare, and defense.
Why this matters even if you don't build robots
At this point you might be thinking, "we're not manufacturing or robotics, so why should we care?" That's where the episode's second concept — physical observability — comes in.
a16z partner Zabie Elmgren points out that U.S. cities already have more than a billion networked cameras and sensors deployed, which is exactly why "understanding what's happening in cities, power grids, and other infrastructure in real time is becoming both urgent and possible." Until now, that data has just been flowing away, unused.
Translate that into plain business terms: your company is already generating unstructured operational data every single day. Customer service logs, in-store camera footage, delivery routes, shift logs — all of it is a potential moat. Just like teleoperation data costs dropped 60%, the gap between companies that start organizing this data now and companies that scramble later isn't going to shrink over time. It's going to widen.
Your data moat starter checklist
- List your company's "messy field data"
Start with what's already piling up that nobody looks at — camera footage, call center recordings, shift logs, delivery routes. - Design the minimum pipeline first
Like the robotics industry does, get a bare-bones "capture → label → store in a standard format" loop running before you build anything fancier. - Start small
Most manipulation tasks in the industry only need 300 to 1,200 demonstrations to train a usable policy. You don't need company-wide data — one team, one process is enough for a pilot. - Build a watchlist of companies to track
Following the quarterly moves of startups like Figure, Agility Robotics, Wandelbots, and Galaxea will give you a feel for when this wave hits your own industry. - Put "data priority" on a recurring agenda
Add "what operational data did we structure this quarter?" as a KPI in your quarterly strategy meeting. That gap compounds into real competitive advantage once you start layering AI on top.
Want to go deeper?
Big Ideas 2026: Physical AI and the Industrial Stack The original podcast where four a16z partners lay out their physical AI thesis a16z.com
State of Robotics 2026 Report A full rundown of robotics industry stats — market size, teleoperation costs, VLA adoption roboticscenter.ai
Humanoid Robotics in 2026: The Race From Pilot To Platform A breakdown of real deployments like Figure/BMW and Digit/Amazon kraneshares.com
2026년 피지컬 AI와 휴머노이드가 이끄는 제조 혁신 A Korean tech media report on Hyundai, Figure, and Enbyus's latest moves etnews.com
Physical AI converges on the warehouse floor A deep dive into how physical AI is actually operating in logistics today marketscale.com
