Physical AI: why the application layer is the next breakout segment

First software was the new services, now physical AI is the new software.

European robotics startups raised €2.67bn in the first six months of 2026, across 87 deals – the most Sifted has recorded in any six-month period. Globally, this year is up 25% on the whole of 2025, with four months still to go. These numbers are fattened by a few massive rounds for defence-drone companies, but at 11% of all AI deals, there’s no doubt this is the breakout segment.

So far, capital has gone into two things: machines, and the models that drive them. Very little has gone into the software layer – but that’s changing. According to Pitchbook, momentum is shifting towards the software that underpins autonomous systems.

Everyone should be paying attention to this, because that is now the layer with all the untapped value.

What is the untapped value?

• To automate knowledge work, you need two things: intelligence, and applications that apply that intelligence to an organisation’s context.
• To automate a factory, you need three things: intelligence, hardware and applications that apply that intelligence to an organisation’s context.

We’re just getting started on the first two categories, so the obvious place for the next unicorns to emerge is in the application layer, just as it has been in knowledge work. At Emerge, we believe the physical AI application layer is the next big thing.

Why do we care?

Manufacturing employs one in seven working Europeans, most of them without a degree, and this is the first automation technology with the potential to raise the ceiling on what they can do rather than lowering the floor on what they are paid for.

Application-level AI has already rapidly transformed how humans interact with computers, from copilots in the enterprise to personal assistants in everyday life. Just as AI agents are redefining productivity for knowledge workers, physical agents will redefine labour.

Our manifesto asks whether robots will steal our jobs, and answers that they might – but if we can adapt, we’ll get a better one. The application layer is what makes that possible, putting AI in the hands of operators and technicians so they can move into more skilled, better-paid work.

Read on for:

• The physical AI stack
• Three categories of AI-powered innovation – which to build in, why and how 
• Why Europe is leading the way

* * *

The stack

Let’s take a closer look at that three-tier stack we just described.

At the bottom are the models – foundation models, and the newer class of world, vision-language-action (VLA) and behavioural models that map what a camera sees onto what a machine should do. This is where the biggest cheques are going: Physical Intelligence was valued at $5.6bn in November; Skild AI raised $1.4bn in January at more than $14bn.

In the middle is hardware – robot arms, drones, mobile robots, humanoids, and the sensors that supply them. This category is gathering momentum because costs here have fallen fast. Industrial arms that once cost high six figures now sell for $50,000 to $100,000, and Unitree’s R1 humanoid ships at $5,900. Goldman Sachs puts the market for humanoid robots alone at $138bn by 2035.

At the top is the application layer: the software people use to run autonomous systems. At first glance, this looks unglamorous by comparison. Tulip raised a $120m Series D in January at a $1.3bn valuation, led by Mitsubishi Electric, and that is the biggest number in the layer. MaintainX raised $150m in July 2025; XOi raised $230m from KKR for field service. But we think there is massive untapped potential in this space – if you know what to build.

There’s another twist here, too.

What connects the application layer back to the models underneath it is training data. The application layer is also the capture surface for both layers underpinning it. 

Bessemer suggested recently that data is the biggest barrier in robotics and getting there will take more than $3bn of industry spend. NVIDIA has built a harvesting machinery to match, claiming its Cosmos pipeline can curate and label 20m hours of video in a fortnight. This data has to come from work as it is actually done, by people, on machines that already exist.

What’s new about this? 

Historically, the biggest paradigm shifts in manufacturing have been driven by automation: the use of machinery or computerised control systems to perform tasks with reduced human intervention. Beginning in the late industrial revolution, three successive waves of automation have unlocked new opportunities and capabilities.

Wave 1 – Mechanical automation (late 19th/early 20th century)

The first wave of automation mechanised the movement of materials through the production process. Using hard-coded machines bolted to factory floors – conveyors, cranes, assembly lines, distributed electric motors – these systems excelled at repetitive, large-scale motions in controlled environments: lifting, carrying and transferring.

The most famous example: in October 1913, Ford’s Highland Park plant put a moving line under the Model T chassis and the time to build one car fell from 12.5 hours to 93 minutes; the price fell from $850 to $260 within a decade. 

What’s most interesting about this wave is how slowly its gains arrived overall: electrification took about forty years to show up in productivity, because factories had to be redesigned around motors on each machine rather than a single steam shaft.

Wave 2 – Computer-controlled automation (1950s-2000s)

Computers introduced digital control of complex, precise movements through software rather than mechanical design. This enabled the automation of skilled tasks like machining intricate parts, welding joints and coordinated motion.

MIT’s Servomechanisms Lab cut the first numerically-controlled part on a punched-tape mill in 1952; Unimate, the first industrial robot, went onto a General Motors die-casting line in New Jersey in 1961; the Modicon 084, the first programmable logic controller, replaced GM’s relay cabinets in 1969. Japan took the technology furthest: FANUC held half the world’s NC market by 1982 and ran a showcase plant at the foot of Mount Fuji where robots built parts for robots with a hundred staff instead of five hundred.

Overall, however, these systems were useful only within narrow industrial applications and were largely rules-based – they followed explicit instructions rather than pursuing goals.

Wave 3 – AI-powered automation (2010s to now)

AI is unlocking automation of new kinds. New models add perception and decision-making capabilities that allow machines to handle variability and unstructured environments. In addition to following pre-programmed sequences, AI systems can recognise patterns, adapt to variations in materials or conditions, and optimise their own performance. 

A camera on an assembly line at BMW Regensburg now writes an individual inspection list for each of the 1,400 cars that pass it per day. Sereact’s picking robots in Stuttgart have made more than a billion picks in production, with one human intervention every 53,000 picks. Universal Robots passed 100,000 cobots sold in February 2025. 

But there is massive potential. Eurostat puts AI use at 17% of EU manufacturers with ten or more staff. In the US, manufacturing is below the national average of 18%.

The first two waves of automation transformed manufacturing by mechanising physical actions of increasing complexity and detail. 

But AI is the first automation technology that can operate effectively in messier, less structured territory – messy SOPs, images, sensor streams and natural language communications that are the native artifacts of manufacturing work. This extends automation into tasks that previously required human judgement and real-time assessment.

Three categories of AI-powered innovation

We see three areas of opportunity within the application layer, which we arrange from “most mature” (founders: don’t build here!) to “most speculative”.

Remember, the application layer does two crucial things:

1. Enables the diffusion of intelligence across hardware (and people), within and between organisations.
2. Captures training data that feeds back into the models and hardware.

So each of these three categories offers opportunities in both; they afford increasing capabilities of AI augmentation in the work itself, and they also offer surfaces with rising data yield.

1. AI as context provider

Manufacturers have been promised “big data” since the 1990s. The promise kept failing for two prosaic reasons: 

• To use data, you first had to standardise it, which meant asking every plant, shift and supplier to fill in the same forms in the same way. LLMs have removed this standardisation step.
• Critical data is scattered across maintenance logs, engineering drawings, quality records, manuals, sensors and more. Tacit knowledge is locked in people’s heads. Video capture and LLMs help here, but integration remains hard; factories run machines which are 20 or 30 years old and literally run in 1s and 0s with no APIs.

What to build

We’ll cut to the chase: don’t!

OK, here’s a bit more nuance… In our pipeline, by far the largest grouping of early-stage startups is building a knowledge layer over manuals, past tickets and the maintenance system, sold to ops.

Is there a big need for a manufacturing system of record? Absolutely, and building this provides significant value to customers early on, but it is not interesting to VCs!

It’s not defensible and your business will stall at information retrieval and insights generation, not getting work done. This is the same situation we saw with knowledge work circa 2023. Since then, the context layer is where almost all of the incumbent systems of record’s AI features have landed and we expect the same pattern to hold in this space.

So these things are simply table stakes:

• Data sovereignty: European plants are refusing US-hosted tools for their operating data, which makes on-premise or EU-hosted inference a sales feature.
• Strong ingestion: To be useful, the system needs to be able to ingest pretty much every valuable piece of data. Big context windows, performant inference and effective caching are vital. Data is only a moat when it is hard to replicate and changes the product.

“At McDonalds, I owned the operating manual. It was a 4,000 page document, published once a year and shipped to every restaurant, so you can imagine, it costs millions of dollars to do that every year. We moved it into an index-searchable document database where if somebody needed to know how many blueberries go on a parfait, they could natural language query it.

At Yum! Brands – KFC, Pizza Hut, Taco Bell – at one point there were 27 different ways to wash your hands. But there’s really only one! With AI at the organisational level, you need be able to sit there and say, examine this database and tell me where I have duplicative or conflicting procedures.”

Rob Lauber, former CLO @ McDonald’s and Emerge VP

2. AI as Force Multiplier

In contrast to AI that merely synthesises, interprets and presents data, augmentation involves using AI to drive actions. AI-enabled tools can dramatically increase both the quantity and quality of an individual’s output.

Some workflows will always need a human-in-the-loop and operate collaboratively, augmenting human capacities by automating specific subroutines and tasks. 

But there are also cases where AI agents or groups of agents can take on full jobs autonomously: the human specifies a desired outcome, and the AI handles the entire workflow end-to-end – devises a plan of action, executes against it, and then recursively evaluates whether it meets the success criteria. This approach changes the nature of human work in two ways: performance of automated routines becomes orchestration of AI agents, and time is freed up to work on strategic and relational tasks.

As with enterprise AI, what’s exciting about this category is that we simply don’t know yet where this line is ultimately going to fall.

What to build

As investors, we see two big risks in this category.

The biggest risk is that the product can never become more than a feature of the connected-worker platform, especially if you have no answer to distribution beyond one plant at a time.

Most of the commercial traction sits in this category because it attaches to budget lines that already exist, but that also means strong incumbents and unsurprisingly these incumbents are buying up point solutions to bolster their own offering. Since 2023 alone, Poka was acquired by IFS and Redzone by QAD, followed by SwipeGuide (→ L2L), Parsable (→ CAI Software), Acadia (→ Epicor) and Drixit (→ Rombit). 

“Most factories have a capex timeline. Every year they will try to find more process to be automated. It used to be a big capex project where you install a new conveyor belt or a packaging machine. Now it’s: ‘we’re going to deploy a robot hand between this workstation and this workstation’. It’s just another form of capex.”

James Kujareevanich, CEO @ Vision Labs

Second, a key challenge is that people may struggle to adapt. Make UK finds only 2% of British manufacturers have AI widely embedded, and more than half identify their own skills shortage as the reason.

The upside is that video-to-procedure tools open a second revenue stream: frontier labs buying training data. In one approach we’ve seen, the lab pays the capture company, the capture company shares that revenue with the factory, and the factory passes it on to the workers being filmed. So founders in this space can sell the same data twice – once to the plant as a procedure, and once to a lab as training data.

“You see this wave of humanoids arriving. These humanoids will also need work instructions, they will also need the procedure with all the quality specifications, in the same way that humans need it right now.”

Batist Leman, co-founder @ Azumuta

• Build read-first, write-later. Plant managers will buy an agent that diagnoses and drafts; they will not yet buy one that changes a setpoint. Earn the write permission with a track record on the read side.
• The real value in this space lies in verticalised expertise: creating unique, specialised data sets and fine-tuning models to handle the nuances of specific industry or factory processes.
• Differentiation comes from annotation and verification. Cameras and sensors are cheap, albeit hard sometimes to implement in EU settings; the scarce thing is the expert’s explanation of why. HD-EPIC, the hardest public benchmark for understanding filmed hand-work, has Gemini Pro at 37%; the model won’t infer the why for a while yet.

“Capturing the workflow is the easiest part of this, with videos or sensors. What’s really important is annotation that’s necessary from some level of expertise – why you did it this way, or why is that the best way. Experts aren’t really good at telling you in the moment why they did something. They can be much better at it by looking at the video and then narrating why the person’s doing what they’re doing.”

Brandon Carson, former VP of Learning @ Walmart and Emerge VP

 3. AI-native factories

The third category is the one the biggest cheques are betting on: not AI in the factory but a factory built around AI. When an entire tech stack comes together – when context flows freely, data becoming insight becoming action; when tools multiply our capabilities; when AI co-pilots work beside us – the factory becomes a system.

That’s the promise of AI-native manufacturing. In this world, what exists is a handful of vertically integrated companies who own the full stack – the models, the machines, the software and the customer – because that is the only way to get the whole feedback loop.

Today, this is nascent, but it is possible to imagine how linking together context, action and AI agency can create totally new kinds of manufacturing environments.

• Hadrian raised $1.37bn in August at a $7.9bn valuation to build automated precision-parts plants for defence and space; its Arizona factory went from ground to production in under six months. 
• Divergent’s $290m Series E funds a design-print-assemble line that has made 600 distinct aerospace parts. 
• In the UK, Isembard raised $50m from Union Square Ventures in March to open 25 software-run machine shops by the end of the year.

How does this shake out? Will only mega VCs fund this and a few players will win? There are significant barriers to achieving full vertical integration at scale, not least (as you can tell from the size of these numbers) capital intensity. The upfront investment to get into machines, facilities and supply chains simultaneously is substantial and the payback period long compared to software-only plays.

Will we see the same pattern as in enterprise, with frontier models dominating adoption? Or will industry- or factory-level applications win, which means there is a role for VCs to play here?

Our take

The AI-native factory is a game for a handful of mega-funds, and it will produce a handful of winners. But it will take at least 10 years.

Most of Europe’s manufacturing output for the next two decades will come from plants that already exist, many of them family-owned, high-mix and running machines older than their apprentices. Those plants will buy AI one workflow at a time, from whoever understands their line.

So the enterprise pattern will only partly repeat. Frontier labs moved into knowledge-work applications fast because the work was already text on a screen. On a factory floor, a model has to be fitted to a specific machine, tolerance and shift pattern, and every iteration happens in the physical world. That gives factory-level applications a longer window before the labs arrive. 

“The physical world has so many constraints. The digital world is pretty clean, everything is in code, so the iteration speed will be like ten times lower. If it takes one year for the LLM to improve this much, it might take five years.”

James Kujareevanich, CEO @ Vision Labs

How will AI affect jobs?

The history of how automation affects employment is complex and varied.

By now, everyone has heard the parable of the bank teller: how the ATM was supposed to replace the bank teller but instead it made branches cheaper to run, so banks opened more of them, and teller employment actually increased.

Channelling this Jevons paradox, many predict a blue collar “renaissance” on the horizon. Certainly, we think there are good reasons to believe that AI-driven job creation is not only possible but highly likely, for a few reasons:

Eroding economies of scale. Earlier generations of robotic automation incurred huge fixed costs and setup times. These economics favour large factories and mega-companies that can amortise investment across massive production runs. AI can bring those costs down, making smaller companies more competitive and enabling a more distributed manufacturing landscape. 

• We’ve already seen this happen. French data proves that plants that automated became more productive, cut export prices, won market share and hired more people, unskilled workers included. The job losses came from the firms that did not automate and lost position to the ones that did. 

Improving job quality. Traditional manufacturing has a barbell structure: high-paid engineers and low-paid workers, with little in between. AI will eliminate the most routine and dangerous work. At the same time, the interface to manufacturing automation is more accessible than it’s ever been. This will make more challenging work – process optimisation, quality analysis, design iteration – more accessible, potentially rebuilding the middle of wage distribution and attracting more talent to manufacturing work. 

The shape of the shortage. The urgent labour story in manufacturing is not hiring cliffs but retirement. 

• BCG puts the average maintenance technician’s age at 50, with about a third over 55, and finds junior technicians taking 3.5x longer than experienced ones on the same task, costing roughly a quarter of plant availability. 
• The European Labour Authority lists welders in shortage in 22 member states, machinery mechanics in 16. 
• Make UK counts 48,000 open vacancies and £6bn a year of lost output. 
• The Burning Glass Institute found that almost a third of UK technician roles will see a quarter or more of their workforce reach retirement within a decade. 

The customer is not actually even trying to remove a person; they are trying to keep what a person knows after she leaves. Every product in the three categories above is, underneath it all, a substitute for the retiring expert.

Why Europe will lead the way

Europe is under-automated. According to the International Federation of Robotics, which counts the number of industrial robots per 10,000 manufacturing workers, Germany is the only European country in the top tier and carries the EU average.

Installations tell the same story: Europe put in 8% fewer robots in 2024 than in 2023. (For context, despite the risk of stoking geopolitical angst, the US put in 9% fewer while China installed 295,000, more than half the world’s total.) 

So the continent with the most expensive labour, the oldest technicians and the strongest worker protections is also the one that has automated least. 

That’s not a contradiction. Fixed automation was built for long runs of identical product, and European manufacturing is disproportionately high-mix, mid-volume and family-owned; a robot that has to be re-programmed by a systems integrator every time the part changes has historically never paid its way. 

AI-powered automation that learns from watching a person and software that runs a mixed line without an engineer is the version Europe has been waiting for. Which is why change starts here.

“Exploration is up and appetite is up. Has consumption increased? I don’t have empirical data to say that, but we see more inquires, dialogues and deal flow with our partners and clients”

Mike Clementi, former CHRO @ Unilever and Emerge VP