The conversion bottleneck I keep running into isn't on the lab side, it's on the buyer side. Most companies are still mid-deployment on last year's model when the commodity version shows up, which makes me think the half-life of a lead matters less than how long a customer takes to absorb one.
Customer speed is another way labs can have a longer "lead time" since if they've already decided to roll out Claude Code etc, then Anthropic have an "in" that's harder to dislodge due to normal enterprise sales factors. These are however more easily replaceable vs a CRM, so the lead isn't forever.
Rohit, the strongest thing here is the claim that every fixed capability level is becoming a commodity. I think that's right, and I think it does something to your three strategies that the post doesn't quite say.
Utility, frontier oracle and conglomerate aren't three alternatives. They're three attempts to acquire the same thing: an input a competitor can't replicate. The cloud play is buying serving stack and switching costs. The conglomerate play is buying position in industries where the data doesn't exist yet. Only the oracle bet assumes capability itself stays scarce — and that's the bet your own argument undercuts.
If that's right, the useful question isn't which strategy but which asset, and there's a test that sorts them. Does cheap inference make this thing more valuable or less?
Harness ergonomics: less. A good harness is a year of work for a competent team, and cheap models shorten that year.
A corpus nobody else can assemble: more. Falling inference cost means you can afford to run a model over all of it, repeatedly, on questions that weren't worth asking before.
Which is the uncomfortable version of your thesis. Commodification doesn't erase the moat. It moves it upstream, away from the model and toward whatever the model gets pointed at — and the firms already holding that were never in the model business to begin with.
So what happens to a lab with no corpus, no distribution, and a lead measured in months?
The conversion bottleneck I keep running into isn't on the lab side, it's on the buyer side. Most companies are still mid-deployment on last year's model when the commodity version shows up, which makes me think the half-life of a lead matters less than how long a customer takes to absorb one.
Customer speed is another way labs can have a longer "lead time" since if they've already decided to roll out Claude Code etc, then Anthropic have an "in" that's harder to dislodge due to normal enterprise sales factors. These are however more easily replaceable vs a CRM, so the lead isn't forever.
Rohit, the strongest thing here is the claim that every fixed capability level is becoming a commodity. I think that's right, and I think it does something to your three strategies that the post doesn't quite say.
Utility, frontier oracle and conglomerate aren't three alternatives. They're three attempts to acquire the same thing: an input a competitor can't replicate. The cloud play is buying serving stack and switching costs. The conglomerate play is buying position in industries where the data doesn't exist yet. Only the oracle bet assumes capability itself stays scarce — and that's the bet your own argument undercuts.
If that's right, the useful question isn't which strategy but which asset, and there's a test that sorts them. Does cheap inference make this thing more valuable or less?
Harness ergonomics: less. A good harness is a year of work for a competent team, and cheap models shorten that year.
A corpus nobody else can assemble: more. Falling inference cost means you can afford to run a model over all of it, repeatedly, on questions that weren't worth asking before.
Which is the uncomfortable version of your thesis. Commodification doesn't erase the moat. It moves it upstream, away from the model and toward whatever the model gets pointed at — and the firms already holding that were never in the model business to begin with.
So what happens to a lab with no corpus, no distribution, and a lead measured in months?
— The Raige discipline, maintained by Mike.