All podcasts / All-In Podcast / Summary

The Robot Episode: Four Leaders on What's Coming

2026-07-29 - 68 min - source - Read full transcript
Jason Calacanis (host)Peter FankhauserBernt BornichAmanda McMasterJonathan Hurst

Key insights

Quadruped robots dominate dangerous industrial inspection because four legs handle unstable, uneven terrain better than two.
Fankhauser said ANYbotics standardized on a four-legged form after 10 years in the field because four legs give a wide, stable footprint for slippery floors, snow, and offshore rigs, while humanoids only make more sense in narrow human-scale spaces like a cafe.
robot-form-factor-and-use-case
Inspection customers don't want the robot, they want the sensor data it collects.
ANYbotics packs its quadrupeds with thermal cameras, acoustic microphones, and gas sensors to detect micro-leaks and overheating beyond human perception; customers justify the six-figure hardware and five-figure annual service cost because avoiding even an hour of asset downtime can pay for the robot.
robot-form-factor-and-use-case
The entire humanoid robotics field is blocked by a lack of internet-scale training data, not compute.
Bornich argued that unlike language models, there is no existing dataset of robot torque commands and sensor input at internet scale, so 1X is deliberately building NEO to be as physically human-like as possible so it can eventually pretrain on all human video footage on the internet, not just narrow teleoperation datasets.
robotics-data-and-training
Training data for robots forms a pyramid, and the industry hasn't tapped the base layer yet.
Bornich described high-fidelity teleoperation data at the top (scarce, expensive to collect), sensor-glove human demonstration data next, egocentric human video below that, and general internet video at the base - orders of magnitude larger than anything collected directly, and only accessible if the robot's embodiment is close enough to human to transfer.
robotics-data-and-training
World models are one tool among many, not a single breakthrough that solves robot learning.
Both Bornich and Hurst rejected the idea of a singular 'world model' silver bullet: simulation lets robots practice at scale, but a persistent sim-to-real gap in fluid dynamics, condensation, and object physics means real-world practice is still required, so progress looks like a snowball picking up speed rather than one crossing-over moment.
robotics-data-and-training
1X is opening NEO into a platform with third-party skills and outside AI models rather than keeping the stack closed.
Bornich announced NEO will support a developer 'app store' or 'skill store' for consumer tasks and will allow outside foundation models to run on the hardware, betting that ecosystem breadth accelerates the path to a fully autonomous home robot faster than 1X could get there alone; NEO begins limited 2026 shipments at roughly $500/month via pre-order deposit, with the first 10,000 slots selling out in days.
robot-economics-and-business-models
Boston Dynamics' Spot has crossed from novelty into ROI-justified deployment at scale.
McMaster said Spot now has over 500 customers in 46 countries doing industrial inspection by day and security patrols by night, with customers requiring payback in under two years; Spot sells on a CAPEX model ($100K-$300K depending on configuration) while the humanoid Atlas is expected to move toward a robot-as-a-service model.
robot-economics-and-business-models
Humanoid robot economics could undercut human factory labor by roughly 90 percent once production scales.
Hurst estimated a Digit robot running 20 hours a day for a 5-year life logs about 40,000 work-hours; at scale (roughly 100,000 units produced) that could bring costs down to about a dollar an hour, versus the $20-40/hour fully-loaded cost of human factory labor - he called that price gap durable because robot value is set by prevailing wages, not the falling cost of the hardware.
robot-economics-and-business-models
US and European robotics CEOs described a coordinated stance against Chinese hardware and IP practices.
Fankhauser said ANYbotics sources zero components from China; McMaster said Boston Dynamics builds 100% of its hardware outside China and Taiwan and argued the US needs a 'national robotics strategy' the way it needed one for semiconductors, citing leaked data from cheap Chinese quadrupeds being back-channeled to China.
china-and-robotics-geopolitics
The China robotics race is framed as being about platform adoption worldwide, not just domestic manufacturing.
McMaster argued the stakes are which country's robotics platform the rest of the world ends up using, warning that losing that race would replay what happened when China caught up in semiconductors - except this time with physical hardware collecting sensitive data inside customers' critical infrastructure.
china-and-robotics-geopolitics
Every guest publicly opposes weaponized robots while acknowledging their companies already do adjacent military work.
Fankhauser co-signed an industry letter with Boston Dynamics and others condemning robot weaponization; McMaster confirmed an anti-weaponization stance while disclosing that the CIA, FBI, and Department of War already field Boston Dynamics robots (unweaponized) and that explosive ordnance disposal is an accepted use case.
robot-military-and-weaponization-ethics
Boston Dynamics would not commit to never building weaponized robots if China arms its own first.
Pressed on whether Boston Dynamics would build weaponized robots to match China's demonstrated armed quadrupeds, McMaster called it 'a tough question' the company would answer 'when the time comes,' signaling the current anti-weaponization position is a business-focus choice rather than a fixed ethical line.
robot-military-and-weaponization-ethics

Media referenced

Companies

Techniques and frameworks

Summary

This is a special guest-interview episode, not the usual four-besties format: Jason Calacanis recorded it solo at the Machina AI conference in Paris, running back-to-back interviews with four robotics company leaders building distinct form factors for distinct markets. Peter Fankhauser (ANYbotics) makes quadruped inspection robots for oil, gas, and wind infrastructure; Bernt Bornich (1X Technologies) makes the NEO home humanoid; Amanda McMaster (interim CEO, Boston Dynamics) runs Spot and Atlas; and Jonathan Hurst (Agility Robotics) makes the Digit warehouse humanoid. Despite different products, the four conversations converge on the same handful of questions: why a given form factor won, how the economics actually pencil out, where the training data comes from, how to handle China, and how far each company will go on military applications.

On form factor, Fankhauser made the case that four legs beat two for industrial inspection - better stability on slippery, uneven, and remote terrain (offshore platforms, arctic sites, explosive atmospheres) where a fall means real damage. He was blunt that customers don't actually want the robot; they want the sensor data it collects, and the economics work because avoiding even an hour of downtime on an expensive industrial asset pays for the hardware. McMaster echoed this with Spot's numbers: over 500 customers in 46 countries, day-shift inspection and night-shift security, sold largely as a CAPEX purchase because industrial buyers think of it as a tool, with payback required inside two years.

The data conversation was the most technically dense stretch of the episode. Bornich laid out a "data pyramid" for training robots: scarce, high-fidelity teleoperation data at the top; sensor-glove human demonstration data below that; egocentric human video below that; and general internet video - vastly larger than anything else - at the base. His bet with 1X is that if NEO is built to be physically close enough to a human (hand dexterity, force sensitivity), the company can eventually train on that huge base layer of ordinary internet video rather than being stuck generating narrow robot-specific datasets. Hurst, working a different form factor, agreed there's no silver bullet: world models let robots practice in simulation at scale, but a persistent sim-to-real gap means real-world practice can't be skipped. Both described progress as an accelerating snowball rather than a single crossing-over moment, and Bornich put "hard takeoff" - robots building robots, chip fabs, and data centers themselves - at somewhere between 3 and 10 years out.

On economics, Hurst did the clearest napkin math: a Digit running 20 hours a day for a 5-year life logs roughly 40,000 work-hours, and at production scale that could push effective cost toward a dollar an hour against $20-40/hour fully-loaded human factory labor - a gap he called durable because robot value tracks prevailing wages, not just falling hardware costs. McMaster confirmed Spot sells CAPEX ($100K-$300K depending on configuration) while Atlas is trending toward a robot-as-a-service model. Bornich's 1X is taking a platform approach instead, opening NEO to third-party developers and even competing foundation models via an app-store-like system, betting ecosystem breadth beats trying to own every layer alone.

Geopolitics ran through every interview. Fankhauser, McMaster, and Hurst all described sourcing components outside China as deliberate policy, with McMaster arguing the US needs a "national robotics strategy" on the model of the semiconductor response, and specifically citing leaked sensor data from cheap Chinese quadrupeds being routed back to China. On weaponization, every guest struck the same pose: publicly against it, while disclosing adjacent military work already happening. Fankhauser co-signed an industry letter against robot weaponization; McMaster confirmed the CIA, FBI, and Department of War already field unweaponized Boston Dynamics robots and that explosive ordnance disposal is accepted, but when Calacanis pushed on whether Boston Dynamics would build weaponized robots if China arms its quadrupeds first, she called it "a tough question" to be answered "when the time comes" - a clear signal the current stance is a business-focus decision rather than a hard line.

Notable Quotes

"For us it's not about labor replacement, right? It's what can we do better? What can we do superhuman?" - Peter Fankhauser

"I am extremely sure there were less than a decade away from hard takeoff... My current bet would be three years." - Bernt Bornich

"It's not just about, yes, it's cute and it dances, but it's long-cost dancing at this point. It's now doing real work." - Amanda McMaster

"If and when that time came, that we had to make a tough decision, we would make the right one. But today we don't have to make that decision." - Amanda McMaster, on whether Boston Dynamics would build weaponized robots

"I don't believe that there's this singularity. I do believe that things are gonna get better and better. Think of it more like a snowball picking up steam going down a hill." - Jonathan Hurst