Cheap AI Is Good Enough for Most Work, but the Money Goes to the Smartest Models
Sacks credits Anthropic's jump from about $10 billion to more than $80 billion in annualized revenue to selling the smartest models.
All-In with Chamath, Jason, Sacks & Friedberg · Google’s AI Brain Drain, SpaceX's Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AIJason Calacanis, a startup investor, said he now does 95% of his work with cheaper, non-frontier open source models, meaning publicly released models that trail the smartest systems. For the work he does, he said, the difference between the cheaper models and the smartest ones is negligible. Elon Musk, who runs Tesla and SpaceX, replied to him online, “it's actually a world of difference.”
Investor David Sacks argues the market has split into two tiers. His idea is that there are the smartest models, which he calls frontier intelligence, and commodity models that trail by six to 12 months.
Sacks said “an Android phone would be good enough for me,” but he still pays a premium because he uses it so much. His point is that heavy users will spend extra for the top tool, even when a cheaper one would work. In business, he argues, a hedge fund in a highly competitive industry does not want to risk missing the best intelligence, and a company hunting for new uses for AI wants the smartest model because finding a valuable use matters far more than the small extra cost for each use.
That premium is what lets the frontier sellers charge for the model itself, Sacks argues. Makers of lagging models, by contrast, can charge only for the computing power or for help setting things up. As evidence, he points to Anthropic, which started the year at about $10 billion in annualized revenue and has passed $80 billion, with forecasts rising to $110 billion or $120 billion by year end.
Investor Brad Gerstner sums up the split this way: use of cheaper open models keeps rising, while the share of the profit goes to the frontier labs. Sacks makes the same phone comparison: Android has more users around the world, but Apple takes most of the profit because people pay for the premium experience.
Questions people ask
What does open source mean for an AI model?
It usually means the underlying weights, the numbers that define how the model behaves, are publicly released and can be downloaded and run by anyone. Companies can then run the model on their own computers and customize it, often more cheaply than paying per use to OpenAI or Anthropic. The tradeoff is they must handle hosting, updates, and safety themselves.
Why would Anthropic be worth so much more revenue than most AI startups?
Anthropic and OpenAI sell access to what many businesses consider the most capable models for coding, research, and complex office work. Large customers pay recurring subscriptions and per-use fees, and heavy business use adds up quickly. Investors value that recurring, premium-priced revenue more highly than revenue from reselling computing power.
Do cheaper models actually keep getting better?
Yes. Open models often learn from techniques pioneered by frontier labs and improve within months, and falling chip and hosting costs make them cheaper to run. That is why many people find them good enough for drafting, summarizing, and simple coding. The frontier labs try to stay ahead by releasing still smarter models for harder problems.