The factory director's roadmap

Industry 4.0 never took off. Is physical AI a bubble?

The question deserves a straight answer: the industry has been disappointed before. The concept of "pilot purgatory" is not new. Founding the Global Lighthouse Network in 2018, the World Economic Forum and McKinsey documented that more than 70 percent of manufacturers were stuck with digital pilots that never delivered measurable business impact. Seven years later, the same statistic still circulates.

Two different bubbles coexist. The first is financial: Figure was valued at 39 billion dollars in its Series C round (September 2025, more than 1 billion raised, Parkway Venture Capital, Nvidia, Brookfield), and robotics startups raised roughly 15 billion dollars worldwide in 2025, above the 2021 peak of 14.1 billion. Those valuations bet on a generalist humanoid market that does not yet exist at scale: no multi-site fleet has been publicly verified.

The second reality is one of deployments that already bill customers: GXO (Agility's Digit, robots-as-a-service since June 2024, more than 100,000 totes moved), Schaeffler and Mercedes-Benz (Apptronik, over EUR 100M invested) since 2025, BMW (Figure 02 then Figure 03 from June 2026). In classic robotics, Exotec has passed 1 billion dollars in systems sold, with claimed growth near 100 percent a year, without making its profitability public. These are not demos: these are lines that run.

Three differences separate 2026 from Industry 4.0's failure: model generalisation attacks the real bottleneck, the engineering cost per task, not the data; real, recurring revenue already exists, where Industry 4.0 stayed a dashboard exercise; and the industrial labour shortage is structural, not cyclical. The European Commission (JRC, June 2025) projects the EU will lose roughly one million workers a year until 2050; in France, 43.8 percent of industrial hires are rated difficult in 2026 (above 70 percent in metalworking and maintenance), with roughly 813,000 retirements expected in industry by 2030.

The honest conclusion: a valuation bubble is likely in some segments, notably pre-IPO generalist humanoids. The underlying technology is not one of them. The real risk is not buying a dead technology: it is scoping the first project badly, picking a vendor that will not survive the coming consolidation, or underestimating integration cost. Track the real maturity of humanoid deployments in our humanoid production index.

When to invest: the 2026-2030 timeline

Three horizons, three very different maturities. Confusing them is the costliest mistake a factory director can make in 2026.

For an industrial mid-cap, the most rational posture in 2026 is that of a prepared fast follower: run pilots now on mature use cases that pay for themselves, build internal skills while humanoid technology stabilises at the pioneers, structure active monitoring instead of waiting. Being a fast follower does not mean lagging: it means letting large groups absorb the growing pains of humanoid integration, while building, on the mature use cases, the real advantage: teams who already know how to deploy, measure and evolve an AI-driven robotic system.

The double ROI: money and people

The financial calculation of a pilot (payback period, cost structure, assumptions to test) is covered in our Factory ROI of a physical AI pilot pillar. This section focuses on a second ROI, almost always missing from board presentations: the human ROI.

First item: musculoskeletal disorders (MSDs). They account for more than 80 percent of recognised occupational diseases in France and cost companies roughly 1 billion euros a year in contributions, for more than 11 million lost working days (INRS data, general regime, schedule 57). Automating a repetitive handling task at risk is not only a social gesture: it is a measurable budget line.

Second item: retention and attractiveness. In France, 43.8 percent of industrial hires are rated difficult in 2026, a rate above 70 percent for metalworking and maintenance, with roughly 813,000 retirements expected in industry by 2030. An unfilled physically demanding role is no longer just a turnover cost: sometimes it simply stays empty.

Third item, the slowest: requalification. An operator supervising a robotic cell instead of repeating a physical motion develops diagnostic and first-level maintenance skills, valuable elsewhere in the plant. This gain is real, but it is measured in quarters, not weeks, and requires an explicit training plan.

The hierarchy stays clear: productivity justifies the project, human ROI reinforces it. A pilot sold purely on the human argument, without a measurable productivity gain, rarely survives its first budget committee.

Who these technologies actually sell to in 2026

According to Deloitte's State of AI in the Enterprise 2026 survey (3,235 executives polled in August-September 2025), 58 percent of companies already report at least limited use of physical AI, and 80 percent expect to within two years. Adoption is most advanced in manufacturing, logistics and defense, and it almost always starts in controlled environments before any extension.

On the ground, the real 2026 buyers overlap: 3PLs and logistics (GXO leading), automotive (OEMs and Tier 1 suppliers, with humanoid pilots at BMW and Mercedes-Benz), large manufacturers facing a labour shortage (Schaeffler), defense, and integrators who sell application cells rather than bare robots.

What is not yet for sale in 2026: a generalist humanoid for an SME. Every documented deployment to date is carried by groups with in-house integration teams, seven-figure pilot budgets and a direct agreement with the manufacturer. An SME contacting a humanoid maker today would run into a waiting list of large accounts, not a product catalogue.

The starter kit: hardware and software

Three budget tiers correspond to three different intentions. The figures below are 2026 orders of magnitude, to be confirmed with an integrator.

None of these building blocks are theoretical: NVIDIA launched its Jetson AGX Thor developer kit at $3,500 in August 2025, and LeRobot, Hugging Face's open library for imitation-based robot learning, is actively used by researchers and industrial teams in the exploration phase.

One point deserves honesty: there is no universal "ready-to-use" package in 2026. The closest offers are application cells assembled by integrators and robots-as-a-service, which turns the purchase into a subscription: the GXO-Agility Robotics model is the best-documented example on the humanoid side, and players such as Formic offer, on the classic robotics side, a pay-for-productivity model with no upfront investment. Compare cobots, AMRs and other platforms in our dedicated comparator.

The people: who you need, and how to build the skills

Contrary to popular belief, a factory does not need a robotics PhD to get started. It needs four roles, two of which often already exist in-house: an automation or maintenance technician trained on the new tool; a designated "robot pilot", the cell's daily point of contact; an IT/OT lead who connects the cell to the information system; and an executive sponsor who shields the project from short-term arbitrations. At the start, the integrator carries the rest: advanced programming, commissioning, performance guarantees.

Skill-building paths already exist. On the vendor side, Universal Robots Academy offers free e-learning in several languages including French, and NVIDIA's Deep Learning Institute offers free self-paced courses on robotics and computer vision. On the French institutional side, Cnam offers training in robotics, cobotics and industrial vision plus a professional degree in industrial robotics, and AFPA trains senior automation technicians including cobotics. Federations such as UIMM also structure regional roadmaps for industrial employment.

A training budget of roughly 10 to 15 percent of the total pilot budget is a reasonable order of magnitude, to adjust (a declared estimate, not drawn from a published study). Learning happens mostly by doing: the best trainer remains the pilot itself, provided people get protected time on it.

The classic trap: management buys the technology but frees up no human time to tame it. The best-integrated cobot in the world sits idle if the operator meant to run it never gets a dedicated hour to practise.

The 90-day method for deciding

An operational checklist, built to produce a decision, not one more report.

Days 1 to 30: map.

1. List 5 candidate roles, cross-referencing physical strain, repetitiveness and current error rate.
2. Cost the status quo for each: loaded labour cost, cost of poor quality, turnover.

Days 31 to 60: confront reality.

1. Visit 2 real deployments, ideally at an integrator's customer site rather than a showroom.
2. Ask 2 different integrators to quote the exact same scope.

Days 61 to 90: decide.

1. Pick a single use case, the most boring and repetitive one on the list.
2. Write down, before any launch, the quantified criteria for scaling up or stopping the pilot.

Alongside these 90 days, check whether your project qualifies for co-financing: our funding your project pillar details the Bpifrance, regional, tax and European schemes available in 2026.

This method is no substitute for dedicated support, but a factory director who follows it seriously walks into a leadership meeting with an argued decision rather than a hunch. If scoping this first pilot or choosing between integrators calls for an outside perspective, D-Fairy Consulting supports this kind of decision day to day.