Thesis
Capability is commoditizing. We fund where value accrues.
Every infrastructure cycle ends the same way. Once the commodity gets cheap, value leaves whoever wrapped it and settles with whoever owns the layer it runs through. AI is there now. This page is the whole argument, and the bar we hold every company to.
01 / The cycle
Value leaves the wrapper and settles in the layer beneath.
It happened with mainframes and the software that ran on them, with the early internet and the rails that carried it, with cloud compute and the applications built on top. In the first years, the scarce thing is the capability and whoever packages it usably captures the value. Then the capability becomes abundant. The package is copied, the price of the underlying unit falls every release, and the margin moves to whoever owns something the abundance cannot reproduce.
AI is following the script faster than any cycle before it. Frontier models are converging on one another, open weights trail the frontier by months, and the price of a unit of intelligence keeps falling. Anything whose value is the model's capability is being repriced toward the cost of the model. The demo that looked like a company two years ago is now a feature of the platform it was built on.
The moat was never the model. It is what the model runs through.
That is the whole thesis in one line. Capability is a commodity input, and a better one is coming next quarter. We back the companies that own the thing the input has to pass through to become an outcome: the data, the record, the license, the distribution, the machine, the settlement rail.
02 / The question
When the next model ships, does this company get stronger, or get erased?
We ask it of every company, before anything else, and we answer it in writing before we look at the terms. It is a stress test that runs on a fixed schedule: a materially better model ships every few months, and each one sorts the market into companies it strengthens and companies it absorbs.
It owns what the model cannot recreate. A better model makes its data more valuable, its workflow more automated, its license more of a moat, and its distribution more profitable. Model progress is a tailwind it does not have to pay for.
It wraps the model. Its product is a capability the next release will include for free. Its traction is real and temporary. Its price is set by the model provider, and the model provider is coming for the margin.
The line rarely announces itself. Both kinds of company have paying customers, good demos and impressive growth in the first year. Telling them apart before the price reflects it is the whole job, and it is a research job, not a pattern-matching one.
03 / Durability
Four things a better model cannot recreate.
Proprietary data that accumulates
Not a dataset that can be licensed by anyone, and not the customer's own data passing through. Data that is generated by operating the function, is rights-cleared, and gets better with every decision the system makes. The test is whether a competitor with the same model and a bigger cheque could rebuild it in a year.
A system of record
The workflow runs through the company. Its output is the thing the customer, the auditor and the regulator look at. Ripping it out means re-running the function, not swapping a vendor. Systems of record change hands once a decade.
A license, an approval or a standing
Regulatory approval, an accreditation, a certification, an audited process, a listed status with a payment network or a government registry. These take years to earn and they do not transfer when the model improves.
Owned distribution
A direct channel into a market that trusts the company and does not switch on price: the integration everyone already uses, the contract with the association, the partner who resells it, the brand the buyer asks for by name.
A company needs at least one of these to be real today and a credible path to a second. A pitch that has none, however good the model behind it, is a wrapper with a runway.
04 / Pricing
Priced on the decision, not the seat.
Software has always been priced on access: a seat, a user, a login. That prices the tool against the IT budget, which is a rounding error next to what the function costs to run. Autonomous systems price the outcome: the claim adjudicated, the code assigned, the underwriting decision made, the route driven, the transaction settled. That prices the company against the labor budget of the function, which is the largest line in most operating businesses.
The market is the payroll, not the software spend
We size a market by the fully loaded cost of the humans who do the function today, and by how much of it the system can run end to end. A category with a small software budget and a large payroll is exactly where we want to be.
The buyer is the owner of the P&L
Outcome pricing is sold to whoever carries the cost of the function, not to the team that buys tools. That shortens the sale, raises the contract value, and makes the renewal a budget decision rather than a feature comparison.
Accountability is the moat
If the system is accountable for the outcome, it needs the record, the audit trail and the controls that go with it. Those are the durable assets from the section above. Outcome pricing and durability are the same thing seen from two sides.
Compute is cost of goods
Model cost sits inside gross margin, so a falling price per token widens the margin of a company that sells outcomes and squeezes the one that sells tokens with a markup. We underwrite unit economics on today's model prices and stress them on the next model's.
05 / The gates
Four gates, before any score.
Some things are not a matter of degree. A company that fails any of these is not investable for us at any valuation, and we say so quickly so nobody's time is wasted.
Civilian
No defense, no weapons, no dual-use systems whose economics depend on a military buyer. Our thesis is about labor and commerce, and we stay inside it.
Fully autonomous
The system runs the function with no human in the loop, or is credibly on a path to it. A co-pilot that makes a person faster is priced against the person, and the next model erodes that price. We fund the system that replaces the loop, not the one that assists it.
Not a thin wrapper
If the product is a prompt, a retrieval layer and an interface over a general model, the next release ships it. We need at least one of the four durable assets to be real today.
Inside our coverage
We only underwrite what we can research to the bottom: the domains we cover, the buyers we can call, the regulation we understand. Being outside that is not a judgement on the company. It is a judgement on our ability to price it.
A gate the material never addressed is a question for the founder, not a fail. Unknown is an honest answer and we treat it as one.
06 / Underwriting
Six questions, one memo, no exceptions.
Every company we take seriously gets the same treatment: a written underwrite against six questions, a score we compute separately from the narrative, and a memo that states the strongest case for and against. The weights are ours and they are stable. The point of a rubric is that it disciplines the conversation, not that it replaces it.
The capability read
The determinative input. Does the next generation of models strengthen this company or erase it? We answer with the model roadmap in view, not the current release.
The moat
Which of the four durable assets does it own, how fast is it accumulating, and what would a well-funded competitor with the same model need to replicate it?
The market, sized by labor budget
The payroll of the function, the share the system can run end to end, and the price per decision the buyer will actually pay.
Autonomy
What fraction of the job runs with no human in the loop today, what the exceptions are, and how the next model changes the ratio.
Commercial traction
Are customers paying for outcomes, is usage replacing labor rather than adding a tool, and do they stay? At the earliest stage this is a design-partner conversation, not a revenue line.
Team
Can they operate inside a regulated market and ship at the model frontier at the same time? Most teams are built for one of those.
The number that disciplines the verdict never arrives in the same breath as the verdict. That is deliberate. A score written after the conclusion is a rationalization, and we have read enough of those.
07 / Stage
Pre-seed to Series A: what good looks like at each.
We invest from the first cheque to the first institutional round, and the bar moves with the stage. What does not move is the question: at every stage we are underwriting the same durability, with more evidence each time.
Pre-seed
The team and the insight. A wedge that is a whole job, not a feature of one. Evidence that the data, the record or the license is obtainable by this team and not by everyone. A first design partner who owns the P&L of the function.
Seed
Early product-market fit for an autonomous system: paying customers whose usage replaces labor, retention that shows the workflow now runs through the company, and a system of record starting to form. Unit economics that hold on today's model prices.
Series A
A repeatable way to sell. Our bar is annual recurring revenue in the low millions and growing, net revenue retention above roughly 110 to 120 percent, and a burn multiple under roughly 1.5 to 2 times. Autonomy that widened, not narrowed, since the seed.
Those numbers are our bar, not a market average, and we hold a company to the bar of its real stage rather than the label on the round.
08 / What changes our mind
The evidence that moves a verdict, in either direction.
A model release absorbs the product
The clearest signal there is. If a platform ships the capability the company sells, we re-underwrite the same week.
Pricing reverts to seats
When a company that sells outcomes starts selling access, the market is telling it what the outcome is worth. We listen to the market.
The data turns out to be licensable
A proprietary dataset that anyone can buy is a cost line, not a moat.
Regulation blocks the loop
If a regulator requires a human at the point of decision and the economics only work without one, the autonomy thesis has a ceiling, and we price the ceiling.
The reverse of any of the above
A license granted, a system-of-record contract signed, autonomy widening after a model release: these move a Watch to a Qualified, and a Qualified to a Conviction.