For the last two decades, investors naturally gravitated towards bits businesses: software, apps, and online platforms. The logic was straightforward; software scales cheaply, distributes instantly, and generates strong margins once built. Atoms businesses, involving physical assets, factories, logistics, or energy, were seen as slower, capital-intensive, and less venture-friendly.

That balance is starting to shift, and AI is a significant part of why.

As Joe Fath, a partner at Eclipse Ventures who invests in physical industries, puts it: "For the first time, AI is making physical industries meaningfully programmable, opening the door for faster scaling with less capital and labour." At the same time, AI is putting real pressure on traditional software business models as the barriers to entry are plummeting and once competitive moats are threatened. 

The underlying logic is one of scarcity. AI is dramatically increasing the supply of digital goods: code, text, images, analysis. As that supply rises, the relative value of digital work tends to fall. Physical work, the business of building things, managing infrastructure, and maintaining assets, is considerably harder to automate in the same way. "As AI drops the cost of computer work to near zero," as Rohan Pandey, a former researcher at OpenAI, recently put it, "the bottleneck, and therefore the capital, returns to the physical world."

This is not a particularly fringe view; Travis Kalanick recently relaunched his next venture under the name Atoms, focused on manufacturing, logistics, and robotics. Jeff Bezos is reportedly pursuing a similar thesis at scale. Even the largest technology companies, including Microsoft, Google, Amazon, and Meta, are now pouring hundreds of billions into data centres, chips, and energy infrastructure. The biggest bits companies are, in a sense, becoming atoms companies.

We think that similar dynamics are playing out at the early stage, and it shapes how we are thinking about the next few years of investing.

Much of the high-profile capital in physical AI is concentrated in the United States, oriented around US national priorities: reshoring, defence procurement, and energy independence. Andreessen Horowitz's $600m "American Dynamism" fund is a stark expression of this.

The deployment opportunity is not uniquely American, though. The UK and Europe have deep industrial heritage, world-class engineering talent, and a real policy urgency around productive capacity and energy transition. Britain started the first industrial revolution and we think it should have a serious role in leading the next. The talent and infrastructure that made that possible, in aerospace, automotive, energy, construction, and engineering, is largely still here and in many sectors remains world-leading.

That said, there are real headwinds for the UK too, and it is worth calling them out. Physical AI is energy-intensive, and UK industrial energy costs are among the highest in the developed world. This is a constraint on the economics of automation at scale. Whilst policy is key, we also believe that the convergence of physical AI with onsite solar and battery storage has real potential as industrial facilities that can decouple from the grid change the cost equation materially. We’re backing companies on both sides of that equation.

The scale of investment flowing into physical AI is striking. Physical AI scaleups raised over $16bn globally in the first three quarters of 2025 alone, and worldwide venture funding in the category reached $26.7bn by early 2025. It is worth noting that a single company, Waymo, accounted for $16bn of that total. Strip that out and the rest of the field is still early, still fragmented, and in many areas wide open. Much of the remaining capital is concentrated in the most capital-intensive spaces: humanoid robots, autonomous vehicles, advanced manufacturing. Figure AI raised $675m whilst Agility Robotics is deploying humanoid workers in commercial warehouses. These are exciting areas that may produce transformative companies over the coming decade, and some of the most sophisticated funds in the world are backing them with conviction.

Where we focus is somewhat different, shaped by where we can find the best founders and where we are well-placed to add value.

There is a parallel wave in physical AI, quieter and less headline-grabbing, that we find equally compelling. Three things have converged to make this the right moment for it:

Models can now work with messy, unstructured data. Information is not neatly logged in industrial businesses. For example, maintenance logs, engineering documents, procurement records, planning files are fragmented, inconsistent, and unstructured. Until recently, making sense of them at scale was enormously expensive. Modern language and reasoning models can now extract signals from this kind of data in ways that were not practical even two or three years ago, opening up industries that previously resisted digitisation.

Integration time has collapsed. Connecting to a legacy system once meant years of expensive enterprise implementation. The best companies we are seeing today are integrating with decades-old infrastructure in weeks rather than months or years, unlocking value fast enough to build a commercial case at an early stage. That shift in the economics of deployment changes what is possible for a well-funded seed company.

The ROI case has become hard to ignore. The cost of downtime in physical industries is concrete and quantifiable in a way that productivity gains in knowledge work often are not. When a critical asset fails unexpectedly, the cost is measurable and immediate. When a construction project overruns, the margin impact is felt directly. That creates a commercial pull for these tools, not persuading customers to change speculatively, but solving a problem with a clear price tag attached.

We are not the right fund for hardware moonshots or nine-figure rounds. We are well-placed to back software-first businesses, typically at pre-seed or seed, building AI tools that are deeply embedded in the workflows of physical industries.

The robots get the headlines and tend to attract the most capital, from humanoid workers on factory floors, autonomous vehicles navigating city streets and warehouse robots picking orders at speed. What gets less attention is the enabling stack beneath them. This is the software that makes physical AI buildable, deployable, and manageable at scale. The tools that help engineers design faster, operators predict failures before they happen, and asset managers run portfolios of physical infrastructure without armies of people doing it manually. It appears that the real moat in physical AI is rarely the model itself. It is the data you can only generate by being embedded in the real world: proprietary, hard-won, and extremely difficult to replicate from the outside. The companies accumulating that data, quietly and at scale, are the ones we think become very hard to displace.

What tends to distinguish the companies we back most comes down to a few things. 

• The founders have usually lived inside the industry before building the tool: as engineers, operators, or domain experts who understand legacy systems and procurement dynamics from the inside rather than the outside. 

• The product sits inside existing workflows rather than asking people to adopt new ones. 

• The business is accumulating real-world operational data that a later entrant simply cannot buy. Unlike generative AI, which could train on the internet, physical AI has no equivalent corpus. Every environment is different, which means deployment data is scarce, valuable and unique. The companies that accumulate it earliest tend to build positions that are hard to displace.

In terms of specific solutions, here are just some of the areas that we are actively looking to meet founders right now:

• Predictive maintenance for critical infrastructure. AI that ingests data from legacy systems to forecast failures across energy, manufacturing, logistics, and transport. The pattern that works in one heavy industry tends to be applicable across several others.

• AI for engineering design and simulation. Tools that compress the cycle from concept to manufacturable product, particularly in structural, civil, and mechanical engineering. The incumbents in this space are decades old, and the opportunity to rebuild for an AI-native world is significant.

• Orchestration layers for fragmented engineering workflows. AI that connects the disconnected tools that engineers already use, surfaces the downstream impact of design changes, and gradually automates the coordination overhead that currently sits with people.

• Grid edge intelligence. Software for managing distributed energy assets as a coordinated fleet rather than isolated units. The hardware is being deployed at pace; the intelligence layer is materially behind.

• Construction site operational intelligence. Real-time visibility across materials, labour, and progress on site. Construction projects consistently run over time and over budget, and the data to change this is increasingly available.

• Voice and ambient AI for physical workplaces. Tools that capture ground-truth operational intelligence from frontline workers, surfacing context that dashboards and metrics tend to miss.

• Autonomous inspection. Computer vision and drone-based monitoring of physical assets: wind turbines, solar farms, bridges, pipelines, buildings. The data generated tends to compound in value over time.

• AI for land, planning, and property data. Planning permissions, land registry, environmental surveys, infrastructure records. Largely fragmented and unstructured, but foundational to construction, energy, nature, and finance.

Our downstream investor network spans hundreds of high-net-worth individuals and family offices with deep roots in real estate, construction, infrastructure, and finance: the industries that physical AI is most likely to transform first. We have backed companies at the intersection of AI and physical operations across our portfolio, in solar asset management, construction procurement, nature restoration, and engineering software, and we have seen first hand what tends to work in these environments. We believe in the productive potential of UK industry, and we want to back the founders who are rebuilding it.

The shift from bits to atoms in technology investment is not a passing theme. It reflects something more structural: as AI drives down the cost and value of purely digital work, scarcity and therefore value tends to move back towards the physical world. The most interesting early-stage expression of that shift, in our view, is not the humanoid robot or the autonomous vehicle, but the software company that becomes quietly indispensable to how physical industries actually operate.

If you, or someone you know, are building in any of these areas, we’d love to chat.