Physical AI is rewriting the robotics market. Most of the startups doing it have no pipeline.

The machine changed faster than the sales motion did. In about three years robotics went from something you program cell by cell to something you train, and a generation of startups is now selling a capability that did not exist when the incumbents wrote their playbooks. What almost none of them changed is how the thing gets bought. That is why so many physical-AI companies have a stunning demo reel, a serious engineering team, a funded roadmap, and a pipeline that is really just three trade shows and a waitlist.

What physical AI actually changed

Physical AI is the label for models that perceive and act in the real world rather than only in text, and NVIDIA has done most of the work of making the term stick, with Isaac GR00T for humanoids and its Cosmos world models. Google DeepMind put Gemini behind robot arms. Physical Intelligence, Skild AI, Figure, Apptronik and Agility Robotics are all pushing at the same premise from different directions. Behind the branding, three shifts matter commercially, and every one of them moves the sale.

Shift one: from programmed to trained

A traditional industrial cell is instructed. Every task is taught, every gripper pose is defined, and a new part number means an integrator visit. A vision-language-action model is trained instead, and the promise is generalization: the same robot handles the item it has never seen. Commercially this replaces the question can it be programmed for my line with will it hold up on my line. That is a completely different proof burden, and it is one most robotics marketing has never been asked to carry.

Shift two: from a machine you buy to an outcome you rent

Robots-as-a-service moves the robot off the capital budget and onto the operating one: pay per hour, per pick, per unit moved. Agility Robotics putting Digit to work with GXO is the reference point most people cite. It sounds like a pricing decision. It is a go-to-market decision: a different buyer, a different approval path, a much smaller first commitment, and a business that now lives or dies on expansion rather than on the first purchase order.

Shift three: from integrator-led to vendor-led

Classic automation reaches the customer through system integrators and distributors, and time to value is measured in months of engineering. The new wave promises weeks and mostly goes direct, which means the vendor now owns deployment risk, the customer relationship and the renewal. It also means you no longer inherit the integrator pipeline. You have to build your own, and almost nobody budgets for that in the seed deck.

Why the old playbook stalls here

None of this is a technology problem. It is that a company selling a trained, rented, vendor-deployed system is running a go-to-market designed for a programmed, purchased, integrator-deployed machine. Six failure points show up again and again.

01 - Nobody searches for a category they cannot name

If your capability is genuinely new, the search volume for it does not exist yet. Buyers google the problem in the vocabulary they already have: reduce packing line downtime, labour shortage on the third shift, palletizing without a safety fence. A demand program built on category keywords captures the handful of people already educated, usually by a competitor, and misses everyone else. You do not capture demand for a new category. You create it in the buyer language, then capture it in yours.

02 - The buying group doubled and nobody told marketing

A conventional automation purchase can survive with a production engineer as champion. Physical AI cannot. Autonomy on a shop floor pulls in safety, IT and OT security the moment the thing is networked and running models, operations for throughput, finance for a subscription that never ends, and often the integrator who owns the rest of the line. Any of them can stop the deal, and each needs different evidence. Arming one champion with one deck is how a deal everybody liked dies quietly in week six.

03 - The pilot is the sale, and it is ungoverned

In this market the pilot is not a step before the sale. It is the sale. And most are run without a written success metric, without exit criteria and without a contractual path to expansion, which is how a company ends up with eight pilots, no revenue, and an engineering team permanently on site. A pilot with no defined finish line is a free consulting engagement with your hardware attached.

04 - The demo proves the robot, not the deployment

A great video shows the machine doing the hard thing once, in your lighting, with your parts. The buyer is asking whether it survives their items, their dust, their night shift and their workforce. Demo-led marketing wins attention and loses the evaluation. What actually moves a physical-AI deal is unglamorous: the deployment record, the safety file, the integration and security documentation, and the failure modes you are willing to name in public before a procurement team finds them.

05 - The pipeline is three trade shows in a trench coat

Robotics still runs on events, and events still work. The failure is treating the show as the pipeline instead of an input to it. Badges get scanned, the list sits until somebody remembers it, and the interest decays for three weeks. The fix is unglamorous and fast: qualify before the show, book the meetings ahead of it, agree with sales what counts as an opportunity, and instrument the path from badge to tracked deal. That is exactly the work behind the 6 to 18 figure on the front page of this site, the same event turned into eighteen tracked opportunities instead of six, without a bigger booth.

06 - You sell capability where the budget is a line item

General-purpose manipulation is not a budget line anywhere in the world. Throughput on the packing line, cost per pick, third-shift coverage and scrap rate are. Capability language sells to engineers, who are rarely the ones holding the money. Translate the same system into the operating number it moves, and the deal finds a budget it can actually come out of.

What a working physical-AI pipeline looks like

The fix is not more campaigns. It is a demand engine shaped for how this product is actually bought, which comes down to six decisions.

That last row is where the compounding lives, and it is the one most startups leave on the floor. Every deployed site is a demand asset: a running cell produces a measured result, a reference the next buyer will actually believe, and an expansion signal, the second line, the second plant, the second region. Wire that into the demand engine, with an agentic layer watching for the signal and triggering the next move, and expansion stops depending on whether an account manager remembers to ask.

"The robot generalizes now. The go-to-market still does not. That gap is the market."

Fellipe Araujo

Why I am blunt about this one

I have spent fifteen years making technical, long-cycle B2B products generate demand, a large part of it inside robotics, leading web and digital experience for two global robotics brands across markets and languages. I have watched a one million euro business become a nine million euro P&L on the back of a system rather than a launch, and I have seen the same trade show produce three times the tracked opportunities once somebody owned the path from interest to opportunity. The technology in physical AI is genuinely new. The commercial failure modes are not, which is the good news: they are solvable, and they are cheapest to solve now, while the category is still being named.

That is the real cost of delay here. Categories get named once. The company that teaches the market the problem in the buyer own words becomes the one every competitor is compared to, and the discount for arriving late is paid every quarter afterwards in longer cycles and thinner margins. Build the demand system while the robot is still the interesting part of the conversation.

Sources and references: NVIDIA physical-AI platform announcements, including Isaac GR00T and Cosmos; Google DeepMind Gemini Robotics; publicly announced deployments and partnerships from Figure, Apptronik and Agility Robotics; and annual installation data from the International Federation of Robotics World Robotics report. The proof points from my own track record are the ones already published on this site.