For AI ventures, compute has moved from an operating expense to a structural constraint that shapes dilution, margin and defensibility. What investors are actually underwriting.
For most of software's history, infrastructure was a line item that shrank as you scaled. AI inverted that. For a company whose product is inference, cost of goods rises with usage and does not obviously fall with scale.
That single change has consequences that reach the cap table.
The margin problem nobody likes discussing
A conventional SaaS business serves the marginal customer at close to zero cost. An AI product serves the marginal customer at a real, recurring compute cost that scales with how much they use it.
Price on seats and heavy users destroy your margin. Price on usage and you cap your own adoption. Most companies end up somewhere uncomfortable in between, and many discover their best customers are their least profitable.
The founders who handle this well treat unit economics as a product decision — routing cheap queries to cheap models, caching aggressively, designing the product so the expensive path is the rare path. The ones who handle it badly discover the problem at Series B, when growth has made it structural.
Capital intensity changes who can compete
If your venture requires significant training compute, you are running a capital-intensive business wearing software clothing.
That changes the funding shape. More capital, earlier, with dilution that compounds — and a strategic question about whether you take investment from the party that also sells you your inputs. Compute-for-equity arrangements are real and sometimes rational, but they entangle your cost base with your cap table in ways that matter at exit.
Companies that build on top of models rather than training them avoid most of this. They also give up most of the defensibility, which is the other side of the trade.
The defensibility question
"We have a proprietary model" was a credible moat for a shorter window than most people expected. Capability diffuses, open weights improve, and the gap between frontier and good-enough keeps closing for most applications.
What has held up better:
Proprietary data that cannot be recreated. Not scraped data. Data generated by the operation of your product, or by access nobody else has.
Workflow depth. Systems embedded far enough into how work actually gets done that replacing them means changing the work.
Regulatory position. An authorization, a certification, an approval. Slow to get and slow for a competitor to get.
Notice that two of those three have nothing to do with the model.
What investors are underwriting
The sharper diligence in this sector has stopped asking about benchmarks and started asking about cost structure:
What does it cost to serve your median customer for a month, and your heaviest?
What happens to gross margin if usage doubles?
How much of your cost base is a vendor who could reprice you?
If the frontier model you build on improved tenfold tomorrow, are you better off or obsolete?
That last question separates two kinds of company. One is a wrapper that gets commoditized by its own supplier's progress. The other is positioned so that better models make its product better.
The read
The interesting AI companies right now are not the ones with the best model. They are the ones whose economics improve as the technology commoditizes, and whose position does not depend on maintaining a capability lead they cannot afford to defend.
For founders, that means designing for a world where the model is cheap. For investors, it means underwriting the business underneath the model rather than the model itself.
In person
Come see the Alliance in person
Members meet at our summits, fireside chats and coffee chats — smaller rooms than the name suggests, and the conversations are the point. Founders can also enter our next pitch session.
Tell us what interests you and we will send details on the next one.
Summit calendarFireside chatsPitch sessionsThe Alliance Dispatch