Physical AI vs Software AI: who actually wins

AI is infrastructure. Robotics is execution. Only one gets paid.

We are seeing it, breathing it, even absorbing it. AI is accelerating in a singular way.
Models everywhere, costs collapsing, capabilities distributed.
The gap between those who produce AI — building power plants, data centers, infrastructure at near-industrial scale — and those who use it to generate margins is wide, and still widening.

On the production side, costs are certain and increasing. On the usage side, returns are uncertain, hard to measure, even harder to sustain over time.

In one sentence: software AI is growing, adoption is increasing, but competitive advantage is thinning, and the ability to consistently and continuously turn intelligence into money is far from guaranteed.

Meanwhile, robotics is still being read through outdated categories. Expensive, slow, hard to scale. But that picture is dated and incomplete.

Because while software becomes increasingly interchangeable, Physical AI is entering a different phase. It has not become simple. It has become INEVITABLE.

And when it works, it doesn’t improve work — it replaces it, generating economic value that is concrete, direct, measurable, and cumulative.

Intelligence commoditizes. Execution does not.

The cost of accessing artificial intelligence has collapsed. The differences between one LLM and another are shrinking fast. What was distinctive yesterday is baseline today.

The advantage is no longer having access to AI, but being able to monetize it.

Inference costs dropped 280x between November 2022 and October 2024 *(Stanford AI Index 2025)*. Enterprise contracts are growing — from $39K in early 2023 to $530K by the end of 2025 *(Ramp Economics Lab)* — but require longer sales cycles and deep integrations. 88% of organizations use AI regularly. Only 7% have completed real deployments. And only 6% of companies can attribute more than 5% of EBIT to AI.

Intelligence is becoming a commodity. This maturity compresses margins across the value chain. Those who control infrastructure — chips, cloud, foundational models — still hold a massive advantage. But everything in between, the generic application layer, operates in a crowded, fragile competitive zone. Traditional SaaS margins were 80–90%. AI execution today sits at 25–60%, and declining.

Building AI is no longer the hard part.
Capturing value — consistently, over time — is.
But is that also true for Physical AI?

First, a distinction.

Software AI — systems that process, generate, or decide on digital data. They don’t touch the physical world. Their output is text, code, images, recommendations, predictions. ChatGPT, GitHub Copilot, Midjourney, fraud detection, Bloomberg AI, Palantir AIP, AlphaFold.

Physical AI — systems that perceive the physical world through sensors and act on it through actuators. They require hardware. Errors have physical consequences — there is no rollback. Tesla Autopilot, Boston Dynamics Spot, Figure 01, Da Vinci Surgical, Amazon Sequoia, Multiply Labs, D.gree, John Deere See & Spray.

Now let’s see whether and how we can compare the two.

Robotics doesn’t scale. Until it does.

Take a deliberately simple example.

A robot that flips burgers, costing a few thousand dollars per month under RaaS — Robotics as a Service, no capex, pay per use. Now compare it to an advanced cloud AI system capable of writing code, analyzing documents, supporting complex decisions.
The point is not which is more sophisticated.

The real questions are two: which generates margin more directly and defensibly, and how high is the real barrier to entry?

And that second question is not secondary. Because the barrier is not just economic. It is cognitive, organizational, cultural.

How much of humanity actually uses AI productively? How many can write effective prompts? How many generate outputs that can be sold? How many — individuals or organizations — can build a stable, repeatable process around AI that consistently produces margin?

Today,
about one sixth of the global population has used generative AI at least once. Of those, less than 2% pay for it. 7% of organizations have completed a real deployment. And only a fraction of those manage to build a stable, repeatable, margin-generating process around AI. Even getting there — and very few do — is exhausting.

The burger-flipping robot does not have this problem.
A task is executed. A cost is removed. The return is visible within weeks. And most importantly — a fast food operator can make it work without prompt engineering, without change management, without retraining staff on new tools.
The example simplifies. But it holds.

UPS has recorded productivity increases of around 300% in its most automated sites, with over 1,400 robots deployed. Amazon has surpassed one million operational robots in its warehouses. Multiply Labs reports cost reductions up to 74% in pharmaceutical manufacturing. Shake Shack increased operating margins from 20% to 25% through automation — while raising wages by 33% in the same period.

These are no longer experiments. Each robot replaces human labor directly. Each additional unit increases operational capacity. Each automated process becomes more predictable, more continuous.

Repeatable, accounted-for margins.

This — the ability to account for return directly and defend it over time — is what markets are starting to price differently compared to software AI.

The entry threshold for Physical AI has dropped from $50,000–$1,000,000 capex to $3,000–$8,000 monthly opex through RaaS.
Payback in some sectors is under ten months. Break-even is reached with five deployed units. Below that threshold, ROI is negative — above it, it compounds.

And as the threshold drops, the margins of those controlling intelligence inside physical systems rise. In 2020, 60–70% of a robotic system’s value was in mechanical hardware. By 2026, 40–50% of the value — with margins of 60–70% — is captured by the intelligence layer: chips and models. The chassis is becoming interchangeable. The brain is not.

This is the crossover the data makes visible.

With RaaS, logistics and cobots move into the same quadrant as AI SaaS tools — investment under $10K, payback under 10 months. Bubble size indicates ROI ceiling at scale. RaaS delivers the best of both worlds: software accessibility, robotics ceiling.

Two curves intersecting. Software margins declining while physical margins rise. Robotics ROI surpasses software after 2028 — not because software stops working, but because generic software can no longer defend competitive advantage once the model becomes commodity.

Defensibility remains in verticals with proprietary data. Defensibility remains in physical systems where usage data — every pharmaceutical batch, every picking cycle, every surgical procedure — feeds a model that improves over time and cannot be replicated without the same deployment.

Value stops where data cannot be copied.

Large funds do not predict the future. They build it.

When Sequoia says Physical AI is the NEXT SUPERCYCLE, when a16z talks about “80% of global GDP still physical” as territory to conquer, when Kleiner Perkins raises $3.5B focused on chips and model labs — they are not analyzing. They are directing capital, talent, and narrative toward a direction. Their predictions are self-fulfilling by design. It is the privilege — and the power — of those who control the flow.

On this, I have nothing to add.
Acknowledging it is the honest starting point.

The categories they highlight are real. Vertical AI with data moats. Data-driven healthcare. Infrastructure for agents. Physical AI as the nervous system of the physical world. The framework is solid — and anyone building today would do well to use it as a compass.

But fund manifestos speak in categories.
They do not speak about operational mechanics.

They do not describe what happens when a mid-sized company — not a San Francisco startup, not a corporate innovation lab — decides to enter this world next quarter.

I don’t invest.
I talk to people who produce.
To operations directors who need to justify a budget line. To manufacturing companies trying to figure out whether a cobot makes sense on an eight-person line. To people building services on data no one has collected yet. From there, things look different. Not better — different.

And this is what you see from there.

RaaS is the entry point. It lowers the threshold from hundreds of thousands in capex to a few thousand in monthly opex. It makes robotics accessible to those who never considered it affordable. But RaaS is not the point — it is the mechanism that opens the real point.

Because robots in the field generate data. Physical, proprietary data that exists nowhere else. Process data, environmental data, behavioral patterns of machines under real conditions. No foundational model has it. No competitor can buy it. No public API distributes it.

This is the hardest-to-replicate structural advantage in the AI economy today. Harder than patents. Harder than teams. Because it requires physical time in the field to accumulate — and time cannot be bought.

This changes the economic model non-linearly.

Conclusions

VCs build the future from the top. They define categories, direct capital, make their predictions real. The framework they have drawn — Physical AI, vertical AI with data moats, agent infrastructure — is the most reliable map available today for anyone trying to understand where the world is heading over the next five years.
But maps do not tell you how to walk.

Intelligence has commoditized. Execution has not.
And physical execution — the kind that leaves data in the real world, in contexts no foundational model has ever seen — is where competitive advantage is being built now, before the narrative catches up and prices it in.

The mechanism is simple to describe, difficult to replicate once in motion.

You enter through RaaS — low threshold, no capex, payback in months.
You deploy physical systems that work, produce, generate repeatable and accounted-for margins. And in the process, they accumulate data. Data that does not exist elsewhere. Data that feeds models, becomes APIs, turns into a proprietary intelligence layer that eventually becomes more valuable than the hardware that created it.

Suppliers are paid in opex. Customers pay in tokens.

This is already happening — in Amazon warehouses, in Multiply Labs clean rooms, in maritime routes where the first atlases of physical data that did not exist yesterday are being built today.

This is not science fiction. This is not a pitch deck. It is the mechanics of how value shifts — from those who produce generic intelligence to those who embed it in irreproducible physical contexts.

Three signals to watch as this unfolds.

The first is the power grid. The United States projects a 68 GW deficit by 2028 due to data center demand. If that constraint materializes, advantage shifts toward those processing intelligence locally — edge AI, autonomous physical systems — without depending on remote cloud.

The second is the pivot to inference. 2026 marks the transition from training to deployment. Inference requires low latency, dedicated hardware, different architectures. Capital is already moving. Those with proprietary physical data hold a structural advantage over anyone wanting to train vertical models in the future.

The third is VLA models — Vision-Language-Action. Systems that translate natural language directly into physical action, eliminating specialized code. When they mature, the distance between an instruction and a robotic action collapses to a few layers. The boundary between Software AI and Physical AI begins to dissolve — and those who have already accumulated physical data will be best positioned to train those models.

Not predictions. Signals already in motion.

Value stops where data cannot be copied. And physical data — generated by a robot operating in a real, irreproducible, proprietary environment — is the hardest data to copy that exists.