The Atoms Premium: Why Software’s Moat Moved

Robots raised more money in the first half of 2026 than in any full year on record, and almost everyone is reading the number wrong.

Crunchbase counts $18.8 billion into robotics startups by mid-year, already past all of 2025 at $15 billion and the old 2021 peak of $14.1 billion, with half the year still to run. On a broader definition Dealroom tracks $55.8 billion, close to double what the same method captured in all of last year. The headlines call it a robot bubble, a humanoid mania, dumb money chasing sci-fi. That is the fun story. It is also the story that will cost you the most if you believe it.

Here is the part that matters for anyone building software. This capital is not chasing robots. It is fleeing something. The same force that made it trivially cheap to build software, AI, also quietly destroyed the two things that made software valuable in the first place: its margin and its moat. Money is not stupid. When the old high ground floods, capital climbs to new high ground. Right now the new high ground is made of atoms. I call the gap between what the market will pay for a durable position in atoms versus a position in pure software the atoms premium, and understanding why it opened is worth more to you than any robotics round.

I run two companies that live on frontier models, and I have no interest in building a humanoid. I am not going to tell you to. This is a map of why the premium moved, what actually holds it in place, and the specific, non-robot moves a software founder can make to stand on the right side of it. The robots are the flare. The signal is underneath.

What this post covers

The $18.8 billion tell

Start with the flow of money, because it sets the frame for everything else. A record like this is not an opinion. It is thousands of professional investors, who get paid to be early and punished for being wrong, moving in the same direction at once.

The rounds are not small experiments either. Skild AI raised close to $1.4 billion at a valuation above $14 billion, roughly tripling its price in about seven months. Saronic pulled $1.75 billion at $9.25 billion for autonomous ships. Germany’s Neura Robotics closed up to $1.4 billion, backed by Tether, Nvidia, Amazon, Bosch and the European Investment Bank. Apptronik stacked its Series A past $935 million at $5.5 billion with Google, Mercedes-Benz and John Deere in the room. These are the checks people write when they think they are looking at the next platform, not the next gadget.

The tell is not that robots got funded. It is who is funding them and what they abandoned to do it. The same venture firms that spent fifteen years perfecting the pure software playbook, near-zero marginal cost, 80 to 90 percent gross margins, ship-in-a-weekend, are now writing nine-figure checks into supply chains, actuators and factories. That is not a side bet. That is a portfolio reallocation, and it is the loudest signal in the market. People vote with capital before they explain themselves in essays. The essays are only now catching up, with venture investors openly describing a rotation from bits to atoms because AI has compressed the returns available in bits.

This has happened before, and the rhyme is useful. Every platform shift reprices where value sits, and it always catches the incumbents of the last layer flat-footed. When cloud arrived, value moved from selling boxes to renting compute, and the box makers who did not move got stranded. When mobile arrived, it moved from the desktop page to the app and the sensor in your pocket. Each time, the founders who won were not the ones who defended the old high ground hardest. They were the ones who read where the water was rising and built there first. I keep a running view of where that water is rising now in the AI opportunity map. The atoms migration is the current chapter of a very old book.

So the real question is not “is this a robot bubble.” The real question is: what changed about software that made the smartest capital in the world start climbing out of it? Answer that, and you stop staring at the flare and start reading the signal.

The two commoditizations

AI did two things to the economy at the same moment, and almost every founder is tracking only the first one.

The first is famous. AI made software cheap to produce. A capable builder with a model can now generate in an afternoon what used to take a team a quarter. Everyone celebrated this, because it felt like pure upside. But cheap to produce and valuable to own are opposites. When anyone can build your feature over a weekend, the feature stops being a moat and becomes a commodity. I have written before about how fast that clock runs in the AI commoditization clock. The short version: the thing that is easy for you to build is easy for everyone else to build too, and the market prices it accordingly, which is to say near zero.

The second commoditization is quieter and it hits the balance sheet. Traditional software scaled with almost no cost per extra user, which is where the fat 80 to 90 percent gross margins came from. AI broke that. Every query now runs a model, burning real compute, memory and energy each time. Bessemer’s 2025 read put model-native gross margins around 65 percent. ICONIQ’s early-2026 snapshot found the average AI product margin at 52 percent, with inference alone eating about 23 percent of revenue. The 90 percent benchmark that built the entire cloud valuation model is gone, and on current pricing it does not come back. I walked through that math in detail in the AI gross margins playbook.

Put the two together and you get the whole story. AI compressed software’s moat toward zero and compressed software’s margin toward the middle, at the same time. The moat fell because building got cheap. The margin fell because running got expensive. And here is the part nobody says out loud: those same AI advances, cheap perception, cheap control, cheap planning, are exactly what finally made atoms buildable. The identical breakthrough that hollowed out software from both sides is what turned robotics from a twenty-year science project into a fundable business. One force, two opposite effects.

Sit with how strange this is if you built your instincts in the last decade. For fifteen years the advice never changed: stay asset-light, avoid inventory, never touch hardware, let the cloud carry the weight. That advice was correct, right up until the moment the cloud got so easy that being asset-light stopped protecting anyone. When every competitor is equally asset-light, asset-light is not an edge. It is table stakes. The founders clinging hardest to the old asset-light gospel are the ones most exposed, because they optimized for a scarcity, cheap software, that AI turned into an abundance.

This is why the robotics number is a tell and not a fad. Capital is not being seduced by shiny machines. It is following the value as it drains out of a flooded position and pools in a new one. If you only track the first commoditization, the cheap-build one, you will keep optimizing features while the ground you are standing on quietly loses its worth. You have to see both effects to see where the premium went.

The premium migration

Here is the core idea of this whole post, and it is worth saying in one line. Durable advantage did not disappear. It migrated. It slid along the spectrum that runs from pure bits to pure atoms, and it kept sliding toward atoms as AI capability rose. The premium is not gone. It moved addresses.

Think of a single axis. On the far left sits pure software, code and nothing else. On the far right sits pure atoms, physical systems in the real world. Every business lives somewhere on that line. For fifteen years the durable premium, the part of a company’s value that competitors could not easily copy, sat far to the left, because software was where scale was free and distribution was instant. AI moved the marker. As models got good enough to write the code, the left end lost its protection, and the safest place to hold value shifted right, toward the things models cannot generate on demand.

The mistake most software founders make is to treat this as a binary: stay in safe old software, or leap into scary robotics. It is not a binary. It is a spectrum, and the smartest position for most builders is not either end. It is the middle, where software wraps around something physical that a model cannot conjure. You keep the speed and margins of code and bolt it to an asset that takes years, capital or permits to reproduce. That combination is the atoms premium in its most accessible form, and you do not need a factory to reach for it.

To see why the right end of the line holds value while the left end lost it, compare the two economies directly. They behave like different physics.

Read that table one row at a time and you can feel the premium move. Every property that used to make bits attractive, cheap to run, cheap to build, fast to ship, has flipped into a property shared by everyone, which means it protects no one. Every property that used to make atoms unattractive, expensive, slow, hard to copy, is exactly what now makes them defensible. Difficulty became the moat. The very friction that scared founders away from hardware is the reason the premium came to rest there.

The four un-copyables

Saying the premium moved to atoms is too vague to act on. Atoms are not magic. A worse robot is still a worse business. What actually holds the premium is a specific set of assets that physical businesses accumulate and that AI, no matter how good it gets, cannot generate on demand. I count four of them. Call them the four un-copyables.

The first is physical data. Not the scraped text a model trains on, but the messy stream that only comes from a machine touching the real world: how a gripper slips on a wet surface, how a warehouse floor changes across a shift, how a specific crop looks the week before harvest. A model can be copied. A dataset gathered by ten thousand of your machines running in real conditions for two years cannot. Every unit you ship pumps more of this exhaust into a well only you own. The people studying physical AI closely already say the data the hardware generates, not the hardware itself, is the real asset. I think of it as physical-data exhaust, and it is the closest thing atoms have to a compounding software flywheel.

The second is the supply chain and the ability to manufacture. Designing a device is the easy part now. Sourcing ten thousand reliable actuators, holding quality across a factory, surviving the version where the first batch fails in the field, that is a capability measured in years and relationships, not in prompts. A model will not hand a competitor your vendor list, your yield curve or the scar tissue from your first three production runs.

The third is regulation and permits. A drone that flies in controlled airspace, a device that touches a patient, a machine that shares a factory floor with people, each one carries approvals that took time, lawyers and a safety record to earn. That paperwork is a wall, and walls that took you two years to climb are two years of protection against anyone who follows. Software founders have spent a decade treating regulation as an annoyance. In the atoms world it is a feature of the moat.

The fourth is installed presence. Once your machine is bolted into a customer’s operation, running their line, trusted with their uptime, it does not get swapped for a slightly cheaper rival the way a SaaS tab gets closed. Physical switching costs are heavy in a way software switching costs stopped being. Presence is sticky. This is a cousin of the point I made about the human core of a business in the incompressible core: some value simply refuses to compress, and the market pays a premium for exactly that refusal.

There is a trap hiding inside this, and it is worth naming. Not every physical business has all four un-copyables, and a physical business missing them is worse than a good software business, not better. A commodity gadget with no data moat, a generic supply chain and no switching cost is just hardware with thin margins and none of the protection. Atoms are not a free pass. The premium attaches to the four assets, not to the mere fact of being physical. This is the same reasoning I used about being trapped by a supplier in the note on vendor lock-in and switching costs: the direction of the switching cost decides who holds the power. Own the un-copyables and the switching cost runs in your favor. Skip them and you have taken on all of the hardness of atoms with none of the moat.

Notice what these four have in common. Each one takes time that cannot be prompted away, and each one gets stronger the longer you run. That is the opposite of a weekend feature, which is strongest the day it ships and weaker every day after as rivals copy it. When code is free, the moat is everything code is not. The four un-copyables are the map of everything code is not.

The capital already voted

If the argument so far is right, we would expect two fingerprints in the data: the cost of building atoms should be falling fast enough to make them fundable, and capital should be pouring in ahead of the crowd. Both are there.

Take the cost curve first, using humanoids as the clearest case. Goldman Sachs revised its humanoid market forecast up sixfold, from about $6 billion to $38 billion by 2035, and the reason it gave was blunt: manufacturing costs fell 40 percent in a single year, where analysts had modeled 15 to 20. Unit prices dropped from a $50,000-to-$250,000 range to roughly $30,000 to $150,000. Bank of America pegs a China-built bill of materials near $35,000 in 2025 and below $17,000 by 2030. Bain models a 60 to 70 percent decline across the decade. And then Unitree shipped an R1 humanoid at $5,900, a number that was supposed to be years away. When a cost curve bends that hard, a market that was fantasy last cycle becomes a business this one.

Now the capital fingerprint. Beyond the $18.8 billion into robotics specifically, deep-tech venture investment hit $48 billion in 2025, up from $18 billion in 2020. China alone put $5.6 billion across 176 robotics deals through mid-May and now accounts for more than 43 percent of global robotics venture money. Analysts describe physical AI companies commanding the valuation multiples that used to belong exclusively to software. That last phrase is the whole thesis in five words. The premium that lived in software multiples is now being paid for atoms. The market has already voted. Most founders just have not read the ballot.

The geography of the money says the same thing in a different accent. China put more than 43 percent of global robotics venture dollars to work because it can move an idea from design to a shipping product faster than anyone, and in atoms that manufacturing speed is itself a moat. That is a warning and an opening at once. It is a warning that the hardware layer will be fiercely contested by players with structural cost advantages. It is an opening because the software layer sitting on top of all that hardware, the tooling, the data, the fleet intelligence, is far less bound by where the factory is. A founder anywhere can own that upper layer over machines built on the other side of the world.

I want to be careful here, because a rising number is not a guarantee, and I will attack this thesis directly in a moment. But the pattern is not one data point. It is a cost curve, a capital rotation, a margin reset and a multiple migration all pointing the same way at once. When four independent measurements agree, you stop calling it noise and start calling it the terrain.

You don’t have to build a robot

Now the part that matters for you, because I am guessing you did not come here to raise $1.4 billion for actuators. The good news is that the atoms premium is not a members-only club for hardware founders. It is a spectrum, and there are rungs on it that a software builder can reach without ever touching a soldering iron. I call it the atoms-adjacency ladder.

The idea is simple. You do not have to move all the way to pure atoms to capture some of the premium. You just have to move one rung to the right of where pure software now sits, toward something a model cannot copy for free. Each rung up trades more capital and slower cycles for a deeper moat. You pick the rung that matches your appetite, not the top of the ladder.

Walk the rungs. Rung one is where you probably are: pure software, shallow moat, easy to copy. Do not stay here by default. Rung two keeps you fully in software but attaches you to a stream of physical data nobody else can get. If your product sits on top of machines, cameras, vehicles or sensors, you can own the exhaust they produce and turn it into a moat that compounds, without manufacturing anything. This is the single highest-return move for most software founders, because it costs almost nothing and buys real defensibility.

Make it concrete. Say you run a scheduling tool for cleaning crews, pure rung-one software that a capable builder could clone in a fortnight. Rung two is right there and almost free: the phones your crews already carry are sensors. Log where jobs actually run long, which sites get rescheduled, how routes drift by weather and traffic, and in a year you hold a dataset about the real economics of that trade that no new entrant can match, because they were not in the field collecting it. The product looks the same to the customer. Underneath, it went from copyable to compounding. That move, from selling a workflow to owning the exhaust the workflow throws off, is open to a startling number of software businesses that never thought of themselves as anywhere near atoms. It is the most under-used play in the current AI-native founder playbook.

Rung three is the classic gold-rush play: sell picks and shovels to everyone building physical AI. Simulation environments, data labeling for robots, fleet management, safety and verification tooling, the software layer that every robotics company needs and none wants to build twice. You never ship a robot, but you get paid every time someone else does. Given that thousands of robotics companies are now funded, the tooling underneath them is a large and growing market that is almost entirely software.

Rung four is hardware-enabled software, a thin, cheap device whose whole point is to be a foothold for expensive, sticky software. The device is a delivery truck for the real product. Rung five is full atoms, the deep moat that also demands the deep capital and the long cycle. Almost no reader of this post needs rung five. Almost every reader would benefit from moving to rung two. The distance between rung one and rung two is small in effort and enormous in defensibility, and that gap is the practical meaning of the atoms premium for a software founder.

One more thing about the ladder. It pairs cleanly with a defensive point I have made before. In the wrapper trap I argued that a thin layer of software over someone else’s model is a rented position. The ladder is the escape route from that trap. Moving up a rung is precisely how a thin wrapper turns into something a model provider cannot absorb, because a provider can copy your prompt but not your fleet’s data, your factory or your permits.

The contrarian take: most of this money loses

Now let me argue against myself, because the honest version of this thesis makes it stronger, not weaker.

Most of that $18.8 billion is going to be lost. Hardware is still brutal. Building a good robot is hard, building a reliable one at volume is harder, and building a profitable business around it is harder still. A large share of these humanoid startups will not return their rounds. Some of these valuations are the kind of froth that shows up at the top of every hype wave, and a few of the names getting nine-figure checks today will be case studies in a graveyard post two years from now. If you read this essay as “robots are a sure thing,” you read it wrong.

But watch what that objection does to the thesis. The claim was never that any given robot company is a good bet. The claim is that the durable premium moved from bits to atoms. Those are different statements. Capital being willing to lose most of it chasing atoms is not evidence against the migration. It is evidence for it. Investors do not tolerate that failure rate for positions they think are worthless. They tolerate it because the winners, the few companies that assemble the four un-copyables, will be nearly impossible to dislodge and will hold their value in a way no software company can anymore. The high loss rate is the price of admission to a moat that lasts. That is exactly what you would expect if the premium were real.

And here is the twist that should reassure a software founder rather than scare one. The brutality of hardware is the reason the smart move is not to build the robot. It is to stand next to it. The picks-and-shovels layer, the physical-data moat, the tooling underneath the whole gold rush, those capture a slice of the atoms premium while sidestepping most of the atoms risk. You get the durable-position benefit of the migration without betting your company on a manufacturing miracle. The founders who lose are the ones who either ignore the migration entirely and keep polishing a copyable feature, or overreact and bet everything on a machine they are not equipped to build. The winners read the map and take the rung that fits them.

One more honest caveat, because I have watched founders overreact to exactly this kind of essay. Moving up the ladder is not free even when it looks cheap. Owning a data stream means you now have to store it, clean it, protect it and actually turn it into something a model can use, and plenty of teams collect exhaust for years and never refine it into an advantage. A moat you own on paper but never operate is not a moat. The efficiency story cuts both ways here, and I pushed on that tension in the efficiency trap: cheaper to build is not the same as valuable to own, and cheaper to collect is not the same as valuable to hold. Pick the one rung you will actually operate, not the three you can describe.

There is a real version of the counterargument worth granting too. Software is not dead, and some software moats survive: deep workflow integration, proprietary non-physical data, brand and taste, the things I wrote about in the taste moat. The migration is a shift in the average, not a law of physics that voids every software business. But the direction is not in doubt, and building as if the old 90-percent-margin, feature-is-a-moat world still exists is the most expensive mistake on offer right now.

What to do Monday morning

Enough theory. Here is the exercise I would run on my own companies this week. Call it the atoms-adjacency audit. It takes an hour and it is uncomfortable in the useful way.

First, time your copy. Take your core product and ask honestly: how long would a competent builder with a good model need to clone the part customers actually pay for? If the answer is measured in days or weeks, you are sitting on rung one, and your position is renting, not owning. Write down the number. It is the single most clarifying figure about your business right now, and it is the same discomfort I pushed on in the commoditization clock.

Second, find your physical-data exhaust. List every place your product touches the real world through a machine, a sensor, a camera, a vehicle, a device, a logistics flow. Each of those is a potential stream of data that only you can collect. Most software teams are throwing this exhaust away without noticing. Circle the richest stream. That circle is your cheapest path to rung two.

Third, pick one rung and one move. You are not rebuilding the company. You are choosing a single step to the right. Maybe it is instrumenting your product to capture and own a real-world dataset. Maybe it is turning an internal tool into a picks-and-shovels product for the robotics wave. Maybe it is attaching a thin device to your software to win presence. One rung, one move, this quarter.

Fourth, set a moat tripwire. Decide in advance what would tell you your current moat has been commoditized, a competitor shipping your headline feature, a model provider launching it natively, your pricing power slipping. Write the tripwire down with a date to check it. The failure mode is not moving too slowly. It is never noticing the ground moved. This is the same instinct behind reading a platform’s incentives before it acts, which I covered in the piece on agent distribution.

Do those four things and you will have converted a scary macro trend into a concrete plan. You will know your copy time, you will have found your exhaust, you will have picked a rung, and you will have a tripwire. That is the entire atoms premium reduced to a Monday to-do list.

Frequently asked questions

What is the atoms premium?

The atoms premium is the extra, durable value the market now pays for a defensible position built on physical systems (atoms) versus one built on pure software (bits). AI made software cheap to build and cheaper to copy, which drained software’s moat, while the same AI advances made physical systems buildable. Capital followed, so a hard-to-copy physical position commands a premium that a copyable software feature no longer does.

Does the atoms premium mean I should build a robot?

No. For most software founders the right move is to climb one rung of the atoms-adjacency ladder, not to jump to full hardware. Owning a stream of real-world sensor data, or selling tooling to the companies building physical AI, captures part of the premium without the capital, risk and long cycles of manufacturing. Very few founders need to build the robot itself.

Why did software gross margins fall?

Traditional software had near-zero cost per extra user, which produced 80 to 90 percent gross margins. AI products run a model on every query, so compute becomes a real, recurring cost of goods sold. Industry snapshots put average AI product margins around 52 percent and model-native companies near 65 percent, with inference alone consuming roughly 23 percent of revenue. On current pricing that margin does not return to the old ceiling.

If most robotics startups will fail, why is the trend real?

A high failure rate is normal for any position with a deep, durable moat. Investors accept losing most of the capital because the few winners, the ones that assemble physical data, supply chains, permits and installed presence, become nearly impossible to dislodge. The willingness to tolerate that failure rate is evidence that the durable premium is real, not evidence against it.

What are the four un-copyables?

They are the four assets a physical business accumulates that AI cannot generate on demand: physical data (real-world sensor streams from your own machines), supply chain and manufacturing capability, regulatory approvals and permits, and installed presence inside a customer’s operation. Each takes time that cannot be prompted away, and each gets stronger the longer you run, which is the opposite of a copyable software feature.

What is physical-data exhaust and how do I capture it?

Physical-data exhaust is the stream of real-world data your product generates wherever it touches a machine, sensor, camera or vehicle. Most software teams discard it. To capture it, list every point where your product meets the physical world, instrument those points to log and store the data, and treat the richest stream as a proprietary asset. It is the cheapest way for a software founder to build a compounding moat.

Is pure software a dead end now?

No, but it is a weaker default. Some software moats survive, deep workflow integration, proprietary non-physical data, brand and taste. The atoms premium describes a shift in where durable advantage concentrates on average, not a rule that voids every software business. The expensive mistake is building as if features still create moats and 90 percent margins still exist.

How much has robotics funding actually grown?

By mid-2026, Crunchbase counted $18.8 billion into robotics startups, past all of 2025 ($15 billion) and the prior 2021 peak ($14.1 billion), with half the year remaining. On a broader definition, Dealroom tracked $55.8 billion globally, close to double the prior year. Alongside this, deep-tech venture investment reached $48 billion in 2025, up from $18 billion in 2020, and humanoid manufacturing costs fell about 40 percent in a single year.