ExpertsFMA rational conversation on where AI is actually going | Benedict Evans
Lenny's Podcast: Product | Career | Growth · May 31, 2026
5 Selected Lessons:
reasoning_pattern · 2 case · 1 framework · 1 heuristic · 1
Diagnosing value capture in a stack: commodity core vs. differentiated edge
Source
Lenny's Podcast: Product | Career | Growth — A rational conversation on where AI is actually going | Benedict Evans (May 31, 2026) · at 31:33 ▸ · at 34:49 ▸ · at 37:04 ▸ · at 39:12 ▸
When this applies
whenever a new critical input (infrastructure, raw material, platform layer) emerges and you need to find where pricing power/profit will actually sit in the resulting stack
The takeaway
To find where value will concentrate in an emerging stack, test each layer for (1) differentiation durability, (2) switching cost/lock-in, and (3) whether the buyer cares which specific provider sits beneath the layer they interact with.
Why it works
If multiple competing providers of the core layer lack durable differentiation (no network effects, no lock-in, no winner-take-all dynamic) and buyers of the layer above don't care which provider powers them, then competition drives the core layer's margins toward commodity/utility levels, pushing value capture to whoever controls distribution or builds the differentiated, sticky application experience.
One level up
Infrastructure that becomes essential does not automatically retain pricing power; essentiality and profitability are separate questions.
Where it stops holding
This stops working if one core-layer provider achieves genuine winner-take-all dynamics (strong network effects or lock-in), in which case Windows-style leverage up the stack becomes possible — Evans flags this as the open, unresolved question.
Detail
procedure:
- Ask whether there are network effects among competing providers of the core layer (model labs) — Evans sees none.
- Ask whether buyers of the upper layer care which core-layer provider sits underneath (AWS-style indifference) or must standardize on one (Windows-style lock-in) — Evans argues it looks like AWS, not Windows.
- Conclude: if indefinite competition + no real differentiation + buyer indifference, margins compress to utility levels and value moves to the app layer and to distribution.
analogy: Telecoms: mobile industry had ~$1 trillion revenue, ~$200B/yr capex (15-20% of revenue), data usage grew ~1500-2000x since 2010, yet stocks went nowhere in 25 years because it became a low-margin commodity utility while 'all the cool stuff' (and value) was made by companies further up the stack (Apple, app developers).
rationale: Sam Altman's comparison of AI to electricity/water implicitly concedes a utility margin structure, which historically means low margins, not the margins model labs currently enjoy during today's 'radical price disequilibrium.'
In their words
If you have competition indefinitely, you don't have different primary, like really radical differentiation of what the product is, then why would you have pricing power?
Jevons paradox check before assuming job loss
Source
Lenny's Podcast: Product | Career | Growth — A rational conversation on where AI is actually going | Benedict Evans (May 31, 2026) · at 12:29 ▸ · at 17:21 ▸
When this applies
whenever a technology makes some unit of output cheaper, before assuming fewer workers are needed
The takeaway
When evaluating 'will X get automated away,' ask whether demand for the underlying output is elastic — if cheaper production unlocks dramatically more volume, headcount can rise even as productivity per task rises.
Why it works
Price elasticity means cheaper production can increase total demand (more work gets done, ROI thresholds that were previously unreachable now clear), offsetting or exceeding the labor saved per unit.
One level up
Supply-side efficiency gains are not conclusions about demand; always check what happens to total volume before concluding about total employment.
Where it stops holding
This stops working when demand for the output is inelastic/saturated — i.e., there's a fixed, capped need (truly the elevator-attendant case) and cheaper production doesn't create more work to do.
Detail
approach: Check historical employment charts for the profession across past automation waves rather than reasoning from the automation event alone.
signals:
- Does cheaper production let people do 'more for the same money' or 'more for more money because of new ROI'?
- Has employment in this field historically risen despite prior waves of tooling (spreadsheets, IDEs, cloud)?
rationale: Before Excel, junior bankers worked brutal hours; after Excel, the same roles still exist but work differently — automation of the task didn't eliminate the job, it changed what's demanded of it.
In their words
If you make it cheaper to do something, what happens? Do you do the same for less money? Or do you do more for the same amount of money? Or do you do more for more money because you've got a new ROI?
You can't predict which jobs are exposed from first principles
Source
Lenny's Podcast: Product | Career | Growth — A rational conversation on where AI is actually going | Benedict Evans (May 31, 2026) · at 1:02:02 ▸
When this applies
when forecasting which specific jobs or industries a new general-purpose technology will or won't disrupt
The takeaway
Treat 'this job/industry obviously will/won't be affected' claims as unreliable — the historical pattern (newspapers, taxis) shows the actual disruption mechanism is usually not the obvious one people first imagine.
Why it works
People anchor predictions on the technology's most visible, literal application (newspapers = printing costs; internet = online booking for taxis) and miss the technology enabling an entirely different mechanism (digital distribution killing print; an app replacing dispatch and ultimately the entire taxi relationship).
One level up
The most confidently stated predictions about a new technology's scope tend to be wrong precisely because confidence substitutes for the humility the situation requires.
Where it stops holding
This stops working when a technology's application really is as narrow and obvious as initially assumed — not every confident prediction is wrong, just unreliable as a class.
Detail
approach: When someone confidently declares a job 'safe' or 'exposed' based on a surface read of the technology, ask what the literal mechanism of disruption would need to be, and consider less obvious alternate mechanisms.
signals:
- Is the stated reasoning anchored only to the most literal, narrow use of the technology (e.g., 'online booking') rather than a deeper restructuring?
- Has a similar confident prediction been wrong before in an earlier platform shift?
rationale: In 1997, people said newspapers would be fine because the internet would only save printing costs, and taxis wouldn't be affected because 'that's got nothing to do with the internet' — both predictions were badly wrong.
In their words
You can't predict which things are going to be exposed necessarily. Or, you know, a lot of the big companies are things that didn't look like that would work and didn't look like that was exposed.
Distribution matters most when the underlying product is a commodity
Source
Lenny's Podcast: Product | Career | Growth — A rational conversation on where AI is actually going | Benedict Evans (May 31, 2026) · at 41:58 ▸ · at 44:00 ▸
When this applies
whenever competing products in a market have converged to being functionally indistinguishable to the end customer
The takeaway
Once a market commoditizes at the product level, competitive advantage shifts entirely to distribution and default-status; the historical browser-engine market and chatbot market both show this pattern.
Why it works
When users can't perceive meaningful quality differences between competing products, the deciding factor becomes whichever product reaches them first/most by default — and incumbents with existing reach (billions of devices, existing app placements) can win on reach even with an 'adequate' rather than best product.
One level up
The location of competitive advantage migrates to whichever layer of the stack still has meaningful differentiation — and distribution/default-status is itself a form of differentiation when the product layer has none.
Where it stops holding
This stops working when the product layer still has genuine, perceivable quality differences for the target user (e.g., in Evans' framing, this logic doesn't apply to power users who do perceive differences between models).
Detail
rule: If the product layer has commoditized (no perceivable difference to normal users), expect distribution and brand to determine market share, not product quality.
threshold: Test: can a normal (non-power) user tell the difference between competing products? If not, distribution wins.
rationale: Meta AI ranked near the top of LLM usage surveys despite being 'written off' by tech insiders, simply because it was 'sprayed on every surface'; Microsoft won browsers via distribution even though winning browsers ultimately didn't matter because value sat further up the stack.
In their words
If the product is a commodity, then the distribution is what matters.
Manufacturing vs. the 'actual thing' — newspapers and record labels steamrolled by the internet
Source
Lenny's Podcast: Product | Career | Growth — A rational conversation on where AI is actually going | Benedict Evans (May 31, 2026) · at 14:48 ▸
When this applies
when an incumbent conflates their physical/logistical delivery mechanism with their actual value proposition
The takeaway
Identify whether your organization's defensible value is the thing you think you sell, or merely the physical/logistical wrapper around it — because the wrapper is the part most likely to be commoditized by new technology.
Why it works
Firms whose self-identity was bundled with a now-replaceable physical process (manufacturing plastic discs, printing and trucking paper) were blindsided when that process was decoupled and became worthless, revealing they had no other moat.
One level up
Incumbents fail not from blindness to disruption generally, but from misidentifying which part of their stack is actually their moat.
Where it stops holding
This stops working when the physical/logistical layer actually IS the differentiated value (e.g., a company whose moat genuinely is superior manufacturing or distribution — not every physical layer is replaceable).
Detail
situation: Record labels and newspapers believed their business was music/journalism, but their actual operational core was manufacturing and distributing physical plastic/paper.
decision: Neither industry had split apart or diversified away from their physical-layer dependency before the internet made that layer free.
outcome: When manufacturing/distribution was decoupled from the end product, both industries were 'screwed' because that was actually what they did, not what they thought they did.
lesson: Audit which layer of your value chain is actually scarce/defensible versus which is commodity logistics dressed up as strategy.
In their words
Record companies do not think of themselves as being in the business of manufacturing small pieces of plastic. But that was what they actually did. And when that went away, they were screwed.