Strong piece — the wedge vs. delivery-engine distinction is useful, and the commoditisation warning is well-earned. But I think the entire framework has a blind spot that comes from looking at this through a SaaS investor's lens. (So does Sequoia and everyone else BTW!).
The debate as framed here is: "Do AI services become software?" That's the wrong question. The right question is: "What happens to a physical asset when intelligence becomes free?"
The piece assumes value accrues to systems of record and embedded software — and that was true in the SaaS era. But it ignores the single most important equation in asset-heavy verticals:
Output = Capacity × Utilisation × Yield
AI doesn't just reduce cost. It increases utilisation of constrained assets. A clinic, a trucking fleet, a salon chair, a lab machine — these are capacity-constrained assets running at low utilisation because of human bottlenecks in diagnosis, routing, conversion, and retention. AI removes the bottleneck not by replacing the service, but by controlling the flow of work through the asset. Same clinic. Same staff. But utilisation up, throughput up, cash flow way up.
This isn't a "services vs. software" outcome. It's a re-rating of the underlying asset.
Your Fab Five analogy is telling — those firms had no asset underneath. They were pure labour arbitrage. But the interesting AI-services plays today aren't consulting firms. They're operating against physical constraints where the bottleneck isn't in software — it's in the real world. Doctor time. Equipment hours. Chair capacity. That's where pricing power actually sits, and that's what gets repriced when intelligence goes to zero.
The piece is right that if you're just delivering a task cheaper, you get commoditised. But the operators who use AI to systematise flow through a physical asset aren't cheaper vendors — they're turning fragmented service businesses into production systems. The moat isn't the software layer or the service wedge. It's control of the bottleneck.
So I'd push back on the closing line. The next giant outcome might not have software margins or software moats. It might have infrastructure-grade cash flows generated by AI-optimised physical assets — bought at services multiples, operated at systems-level efficiency. Not AI tools. Not SaaS. Repriced assets.
Good push. Utilisation is the right correction to a SaaS-only frame, and “control of the bottleneck” is the right abstraction.
Worth separating two cases though. In throughput-constrained assets like clinics, fleets, chairs etc. the bottleneck is utilisation. AI routes more work through the same asset and the cash flows get repriced. Agreed.
In safety-critical assets like process plants, LNG and refining (where I’m at), the bottleneck isn’t throughput. It’s verification of human execution, and regulation locks the human in place. You can’t optimise the operator away, and you wouldn’t want to. The moat there isn’t asset utilisation, it’s owning the data layer for work that was never digital to begin with.
Same underlying argument, i.e: value accrues to whoever controls the physical bottleneck, not whoever wraps the LLM. Different flavour of bottleneck.
Curious whether Euclid sees those as one category or two or possibly a third????
In delivery models, the primary challenge is not intelligence work but coordination, compliance, liability ownership, and the physical-world handoff between the software-powered output and the customer’s desired outcome. The service provider is effectively an embedded partner and owns the outcome. Owning the outcome means the last mile is structurally lengthy, at least today. The work is typically outsourced, but can also be insourced.
BTW Euclid, I’m a new subscriber and thoroughly enjoying your fresh & different content
Thank you Neil! We’re glad you have you onboard :)
YC also talked about it. would be interesting to actually see the status report on ai-native services these VCs actually invested in, not just talks
Such a strong piece!
Thank you, very kind of you
Strong piece — the wedge vs. delivery-engine distinction is useful, and the commoditisation warning is well-earned. But I think the entire framework has a blind spot that comes from looking at this through a SaaS investor's lens. (So does Sequoia and everyone else BTW!).
The debate as framed here is: "Do AI services become software?" That's the wrong question. The right question is: "What happens to a physical asset when intelligence becomes free?"
The piece assumes value accrues to systems of record and embedded software — and that was true in the SaaS era. But it ignores the single most important equation in asset-heavy verticals:
Output = Capacity × Utilisation × Yield
AI doesn't just reduce cost. It increases utilisation of constrained assets. A clinic, a trucking fleet, a salon chair, a lab machine — these are capacity-constrained assets running at low utilisation because of human bottlenecks in diagnosis, routing, conversion, and retention. AI removes the bottleneck not by replacing the service, but by controlling the flow of work through the asset. Same clinic. Same staff. But utilisation up, throughput up, cash flow way up.
This isn't a "services vs. software" outcome. It's a re-rating of the underlying asset.
Your Fab Five analogy is telling — those firms had no asset underneath. They were pure labour arbitrage. But the interesting AI-services plays today aren't consulting firms. They're operating against physical constraints where the bottleneck isn't in software — it's in the real world. Doctor time. Equipment hours. Chair capacity. That's where pricing power actually sits, and that's what gets repriced when intelligence goes to zero.
The piece is right that if you're just delivering a task cheaper, you get commoditised. But the operators who use AI to systematise flow through a physical asset aren't cheaper vendors — they're turning fragmented service businesses into production systems. The moat isn't the software layer or the service wedge. It's control of the bottleneck.
So I'd push back on the closing line. The next giant outcome might not have software margins or software moats. It might have infrastructure-grade cash flows generated by AI-optimised physical assets — bought at services multiples, operated at systems-level efficiency. Not AI tools. Not SaaS. Repriced assets.
Good push. Utilisation is the right correction to a SaaS-only frame, and “control of the bottleneck” is the right abstraction.
Worth separating two cases though. In throughput-constrained assets like clinics, fleets, chairs etc. the bottleneck is utilisation. AI routes more work through the same asset and the cash flows get repriced. Agreed.
In safety-critical assets like process plants, LNG and refining (where I’m at), the bottleneck isn’t throughput. It’s verification of human execution, and regulation locks the human in place. You can’t optimise the operator away, and you wouldn’t want to. The moat there isn’t asset utilisation, it’s owning the data layer for work that was never digital to begin with.
Same underlying argument, i.e: value accrues to whoever controls the physical bottleneck, not whoever wraps the LLM. Different flavour of bottleneck.
Curious whether Euclid sees those as one category or two or possibly a third????
Great one. Thanks for sharing. Definitely AI + Services is the generational opportunity for entrepreneurs
Love this debate, its getting better daily!
I disagree in our context, acknowledging the wedge vs moat:
https://unbroker.com/why/
This paragraph is showing up twice, I think:
In delivery models, the primary challenge is not intelligence work but coordination, compliance, liability ownership, and the physical-world handoff between the software-powered output and the customer’s desired outcome. The service provider is effectively an embedded partner and owns the outcome. Owning the outcome means the last mile is structurally lengthy, at least today. The work is typically outsourced, but can also be insourced.