The Cost of Distribution in the AI Era
AI has collapsed the cost of building product. But winning — building a global brand and earning enterprise trust — is getting harder, not easier. When everyone can build the product, only distribution differentiates.
AI is collapsing the cost of building product. Inference costs are dropping 10x per year for equivalent model performance. Small teams are shipping at enterprise scale. Cursor reached $100 million in ARR with 20 people. Midjourney built a $200 million business with 40.
But the cost of distribution — brand building, enterprise sales, customer acquisition, implementation services — is not declining. If anything, it is increasing.
The numbers are striking
OpenAI spent $2 billion on sales and marketing in the first half of 2025 alone — nearly doubling what it spent in all of 2024. That $2 billion is against $4.3 billion in revenue, meaning sales and marketing consumed roughly 47% of every dollar earned. The company is now hiring hundreds of AI consultants to build a technical consulting team for enterprise deployment, recognizing that the technology sells itself in demos but implementing at scale requires an entirely different skill set.
Anthropic grew from $10 million in revenue in 2022 to $14 billion annualized by early 2026 — and now commands 40% of enterprise LLM spend, overtaking OpenAI. To get there, they grew headcount from 192 to over 4,000, built a full enterprise sales organization across five segments, tripled their international workforce, and spent $16.5 million on television advertising in a single year.
Perplexity agreed to pay Snap $400 million over one year — cash and equity — to integrate its AI answer engine into Snapchat. This is a pure distribution play: paying $400 million not to build product, but to access 940 million users.
Generative AI platforms collectively spent more than $1 billion on digital ads in the United States in 2025, up 126% from the prior year. The AI advertising arms race is accelerating, not decelerating.
This is not unique to AI companies
Snowflake, at $3.6 billion in revenue, still spends approximately 47% of every dollar on sales and marketing — roughly $1.7 billion per year. Salesforce spends $13 billion per year on sales and marketing. Thirteen billion dollars. That is larger than most AI companies' total revenue.
Customer acquisition cost across B2B SaaS has increased 60% over the past five years. The median company now spends $2 to acquire $1 of new annual recurring revenue. Average B2B SaaS sales cycles have lengthened from 107 days to 134 days in a single year. And 75% of software companies reported declining retention rates in 2024, creating a feedback loop that demands still more acquisition spending.
The paradox
This is the paradox of AI-driven efficiency: it is easier than ever to build, but harder than ever to become the category-defining company that accrues the majority of value in a market.
The fastest-growing AI startups — what Bessemer calls "AI Supernovas" — are running gross margins of approximately 25%. Some have negative gross margins. They are deliberately trading profitability for distribution, investing aggressively in customer acquisition and implementation services in the hope that scale and switching costs will eventually create defensibility.
Palantir pioneered this model: embedding implementation engineers directly with customers, trading margin for moat. Job postings for "forward-deployed engineers" across AI startups are up hundreds of percent in 2025. Andreessen Horowitz calls this "services-led growth." It works — but it is expensive.
Why this matters
In an era where product development costs are plummeting, the scarce resource is not engineering. It is distribution. Enterprise trust. Brand. Customer relationships built over years, not quarters.
The companies that win will be those that invested early and aggressively in these assets — companies with hundreds of millions in revenue, evangelical customer bases, and go-to-market machines that took years and billions of dollars to build. These moats are not replicable by a small team with a coding assistant.
This has always been true in enterprise software. But the AI era makes it more stark: when everyone can build the product, only distribution differentiates. We invest in the companies that have already built it.