Capgemini surveyed 1,700 organizations and 79% are already engaged with Physical AI. NVIDIA open-sourced GR00T N1.7 on Apache 2.0. Japan Airlines signed a 2-year deal at Haneda. Here is the week that moved Physical AI from pilot to platform.
A survey question was sent to 1,700 organizations across industries. The question was essentially: are you working with Physical AI yet? 79% said yes - not planning to, not evaluating it, but already engaged. 67% of CEOs in that same group called it a game-changer.
That is the Capgemini number from this week. And it lands differently than a funding headline or a robot demo. Funding can be speculative. Demos are controlled. A 1,700-company survey with 79% active engagement is a market temperature reading - and the temperature this week was unmistakably high.
Here is everything else that happened alongside it, and why the pieces fit together in a way that matters beyond each individual story.
| Metric | Value |
|---|---|
| Companies already engaged with Physical AI | 79% of 1,700 surveyed (Capgemini) |
| Training data for GR00T N1.7 EgoScale | 20,854 hours of human POV video |
| Japan Airlines Haneda commitment | 2-year operational deal with Unitree G1 |
| GR00T N1.7 license | Apache 2.0 - fully commercial open source |
79% Is Not a Forecast. It Is a Survey Result.
There is a meaningful difference between "X% of companies plan to adopt AI" and "X% of companies are already engaged." Planning is cheap. Engagement means teams, budgets, and at least one robot somewhere doing something real.
The Capgemini survey of 1,700 organizations found that 79% are already working with Physical AI and 67% of executives consider it a genuine strategic game-changer. BCG and Deloitte both published separate analyses this week reaching the same conclusion: the industry has crossed from a pilot phase into a strategy phase. These are not the same thing. Pilots have escape hatches. Strategy has budget lines.
Bessemer Venture Partners offered the most precise framing: this is the "GPT-2.5 moment" for robotics. Capabilities are real and demonstrably scaling. But the gap between current performance and the 99.9% production reliability required for full industrial deployment still exists. The analogy is useful because it tells you where we are on the curve: past the "does this work?" question, not yet at the "we can depend on this completely" answer.
What this means for your organization: If you are in the 21% not yet engaged, you are not safe - you are late. The companies currently running pilots are building institutional knowledge that compounds. The cost of catching up in 2027 will be higher than the cost of starting in 2026.
NVIDIA Just Open-Sourced the Brain of a Humanoid Robot. Here Is What That Changes.
On Friday, NVIDIA released Isaac GR00T N1.7 on Apache 2.0 - a fully commercial open-source vision-language-action model for humanoid robots. The license matters: Apache 2.0 means any company can use it in production, modify it, and ship products built on it without royalties or restrictions.
The technical story behind N1.7 is called EgoScale: NVIDIA pre-trained the model on 20,854 hours of video recorded from a human first-person perspective, covering 20+ task categories. From this, they derived the first observed scaling law for dexterity - increasing training data from 1,000 to 20,000 hours more than doubles manipulation accuracy. That is the same kind of predictable scaling that made large language models investable. When you can plot a curve and extrapolate it, you can plan a roadmap.
Humanoid, LG Electronics, and NEURA have already announced they are building on GR00T N1.7. Expect that list to grow fast. An open foundation model reduces the barrier for every robotics company that was previously spending resources on training from scratch. The gravitational effect is deliberate: NVIDIA is building the same ecosystem strategy for physical AI that it built with CUDA for GPU computing.
Why this matters: Open AI foundations accelerate the entire field. Companies that adopt GR00T N1.7 can focus engineering resources on application-layer differentiation rather than foundation model training. The cost curve for capable robots just dropped again - this time at the software layer.
Asia Is Moving Faster Than Western Boardrooms Realize.
Three separate Asia-Pacific moves this week tell a coherent story about who is treating Physical AI as infrastructure, not experiment.
Japan Airlines committed to a 2-year humanoid robot program at Haneda Airport using Unitree G1 robots (130cm, 35kg, 2-hour battery life). Tasks: baggage loading, cargo transport, cabin cleaning. Partner: GMO AI and Robotics. The driver is explicit - Japan's aging population is cutting labor availability while tourist traffic hits records. JAL is not deploying robots because it wants to. It is deploying because the demographic math leaves no alternative at scale.
Mitsubishi Electric and Chiba Institute of Technology signed a 3-year co-creation agreement to build Japan's own Physical AI stack from scratch: multi-legged walking robots, humanoids, and drones for infrastructure and emergency response. Mitsubishi brings precision motion control from its MELFA industrial robot line. Chiba brings large physics models for unpredictable environments. Japan is not licensing Physical AI from US companies - it is building sovereign capability.
In Singapore, IntBot and Certis Group announced a strategic partnership to deploy social robots in hotels, airports, hospitals, and shopping centers. Certis operates over 25,000 workers across Singapore, Australia, and Qatar. IntBot's layer is called General Social Intelligence - robots that recognize intent, hold conversations, and navigate crowded unpredictable spaces.
The pattern: Demographics, sovereignty, and dense urban environments are driving faster adoption in Asia than market analysis typically accounts for. The companies watching this from Western boardrooms should also be tracking the Taiwan supply chain data: Q1 2026 order books for humanoid actuators, gearboxes, and sensors are growing faster than projections across suppliers for Unitree, Figure AI, and 1X Technologies.
What to Watch Next
- GR00T N1.7 adoption velocity - how many companies announce builds on Apache 2.0 in June will signal how fast the open ecosystem forms.
- Robotics Summit Boston follow-through - the "State of Humanoids" panel had Boston Dynamics, Agility, and Schaeffler setting standards together. Watch for joint announcements in the weeks after.
- JAL Haneda operational data - the first real performance data from a 2-year commercial airport deployment will be the most honest benchmark yet for humanoid reliability.
- Capgemini 79% breakdown by industry - the aggregate is striking; the sector distribution will tell you which industries are leading and which are genuinely behind.
- Physical Intelligence $1B close timeline - the round has been in negotiation; a close at $11B valuation resets the entire sector's comparable set.
FAQ
Q: What does "GPT-2.5 moment for robotics" mean in practice?
A: Bessemer's framing refers to the stage GPT-2.5 represented in language AI: capabilities were clearly real and scaling, but the technology was not yet reliable enough for most production use cases. For Physical AI in 2026, this means robots can handle structured tasks in controlled environments at meaningful scale, but the 99.9% reliability required for unsupervised industrial deployment is still a gap. The implication: invest and build now, because the reliability curve is predictable and the companies entering late will find the gap harder to close.
Q: Why does NVIDIA releasing GR00T N1.7 as open source matter for non-robotics companies?
A: Apache 2.0 means any company - including yours - can build a physical AI application on top of the same foundation model that Humanoid and LG Electronics are using, without licensing fees. The practical implication: the cost of building a capable task-specific robot application just dropped significantly. If your industry involves structured physical work, the barrier to prototyping a robotics solution in 2026 is lower than it has ever been.
Q: Why is Japan deploying humanoid robots faster than most other markets?
A: Japan has the most acute combination of demographic pressure and industrial precision culture of any major economy. The population is aging faster than any comparable country, labor availability in physical service roles is already constrained, and Japanese industrial culture has decades of comfort with robotics in manufacturing. Physical AI is not a disruption in Japan - it is a continuation of a 40-year automation trajectory, now applied to tasks that previous generations of robots could not handle.
Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.
Top comments (0)