Value-Up-The-Stack Test
To find where the money accrues in a platform shift, test the infrastructure layer for network effects, differentiation, and pricing power
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 5
- Confidence
- 88%
A test for locating where profit will concentrate in a technology stack. The instinct is that whoever owns the powerful core layer (foundation models, telecom networks, cloud) captures the value. Evans argues you must instead test that layer for three properties — network effects, product differentiation, and pricing power. Where all three are absent, the core becomes a commodity utility selling at marginal cost and the value migrates up the stack to the applications built on top.
Origin
Benedict Evans's thesis from the capital section of his 'AI is eating the world' deck, drawing on his career as a telecoms equity analyst and the histories of mobile networks, web browsers, and cloud (AWS vs. Windows).
Core principles
- 01Selling objectively amazing, complex infrastructure does not guarantee margins — the global mobile industry proves this
- 02Ask three questions of the core layer: does it have network effects, real differentiation, and therefore pricing power
- 03No network effects means no winner-takes-all, means indefinite competition, means commodity pricing
- 04Windows-style leverage up the stack is the exception; AWS-style commodity-underneath is the more common outcome
- 05The commodity layer powers the value but does not capture it — 'all the cool stuff is made by you'
How to run it
- 1
Map the layers of the stack
Separate the capital-intensive core infrastructure (foundation models, networks) from the application layer that users actually touch.
- 2
Test for network effects
Ask whether one provider's product improves as it gains users in a way that runs away from rivals. If models show no network effect, no single one runs away and competition persists indefinitely.
Watch out A capital moat (hundreds of billions in capex) is not the same as a network-effect moat; do not conflate them.
- 3
Test for differentiation
Ask whether the products are radically different to the user. If a normal user cannot tell Gemini from a rival, the layer is undifferentiated regardless of the Nobel-prize-level science inside it.
Pro tip Flat-panel screens involved Nobel prizes and are still a low-margin commodity — sophistication does not equal margin.
- 4
Ask whether the core builds the apps or others do
If thousands of different applications must be built by many different companies rather than by the core provider, the core looks like cloud (AWS), not like an OS (Windows), and has little leverage up the stack.
Watch out Distinguish today's chaotic price disequilibrium from the steady-state; a $1.5M monthly token bill is like a $50k mobile-data bill in 2010 — temporary.
- 5
Conclude where value accrues
If the core has no network effects, no differentiation, and does not build the apps, expect commodity pricing at the core and value concentrated up the stack in the applications.
In the wild
Global mobile has ~$1 trillion revenue, spends ~$200bn a year on capex, and carries data volumes ~1,500-2,000x their 2010 level. Yet the stocks have gone nowhere for 25 years because it is a low-margin commodity utility — all the valuable stuff runs further up the stack on your iPhone.
→ Objectively amazing infrastructure captured almost none of the value it enabled.
Microsoft used distribution to win the browser war and held it for five or six years, but it 'doesn't matter, it doesn't get them anything' because the value was further up the stack.
→ Owning the commodity layer produced no durable profit; value lived above it.
Common mistakes
Assuming the powerful core layer captures the value
Sam Altman's 'we'll sell intelligence on a meter like electricity' analogy backfires — the TV company doesn't pay the electric utility a share of your bill; utilities are low-margin, so metered-intelligence implies commodity economics, not dominance.
Reading today's price disequilibrium as the steady state
Extreme current token prices and spend are transient, like a $50k mobile bill in 2010; forecasting pricing power from them ignores where the lines settle once the market reaches equilibrium.
Is it for you?
Best for
Investors and strategists deciding which layer of an emerging tech stack (models vs. applications) will capture durable profit
Not ideal for
Layers that genuinely do have network effects or proprietary lock-in, where Windows-style upstack leverage can hold
From the transcript
“the models don't seem to have network effects. So there doesn't seem to be a winner takes all effect where one of these will run…”
“why would the model companies have pricing power and wouldn't all the value be further up the stack?”
“it's an Xgrowth low margin commodity utility where they're selling this inc this objectively amazing piece of global technology infrastructure”
“it should end up looking more like cloud than it looks like Windows”
“They're they're undifferentiated commodity infrastructure providers.”
From the episode
A rational conversation on where AI is actually going
Benedict Evans