By One Investor's Count, AI Labs Get 41 Days Of Profit Where Drugmakers Get 17 Years
High margins must pay for the next half-billion-dollar model before price cuts and cheaper rivals erase them.
This Week in Startups · Why a Do-Nothing Congress could be great for markets | E2359Venture investor Tomasz Tunguz, who leads the firm Theory Ventures, estimates that a leading AI lab keeps its share for about 41 days before another strong model arrives and market share starts to shift.
That brief window is his explanation for why AI looks different from prescription drugs. A drugmaker that spends a half billion or a billion dollars developing a drug gets 17 years of patent protection to sell it without exact copies.
An AI lab that spends a similar sum training a powerful model gets no such shelter. Tunguz said the lab needs a large profit in that short period to finance development of its next model.
Tunguz put leading AI margins at about 85 percent. He cited Anthropic's filing at about 81 to 82 percent gross margin, meaning sales minus direct costs, compared with about 71 percent for a publicly traded software company.
But those large margins invite the response capitalism usually delivers. OpenAI cut some prices by 50 percent, then by 80 percent, while freely available open source models cost 80 to 85 percent less than closed models, creating steady downward pressure.
D.A. Wallach, an investor with Time Bio Ventures, said those roughly 80 percent margins make him uneasy. He questioned Silicon Valley's everyone wins thesis, the idea that memory makers, chipmaker NVIDIA, cloud providers, and frontier labs can all keep monopoly like profits at once.
In normal competition, Wallach said, an 80 percent margin should attract rivals until profits fall. For AI labs, the open question is what a lasting, ordinary profit level looks like when the advantage keeps expiring in weeks.
Questions people ask
What does an 80 percent gross margin actually mean?
Gross margin is sales minus the direct cost of delivering the product, divided by sales. For an AI lab, that mainly means the computing, power, and related costs of answering user requests, not research, salaries, or rent.
Why does it cost hundreds of millions of dollars to build an AI model?
Training means feeding huge amounts of text, images, and other data through thousands of costly specialized chips for weeks or months. The expense comes from renting or buying those chips, supplying large amounts of electricity, and employing scarce researchers.
What is an open source AI model, and why does it push prices down?
An open source model is one whose underlying weights, the numbers that define how it responds, are publicly released for others to copy, modify, and run. Because several labs release them, customers can switch to a cheaper option instead of paying a premium brand price.