If the models are so good, why is OpenAI still selling access?

If you had a machine that could build a profitable company from a prompt, would you rent it out by the token? That is the uncomfortable question behind a Hacker News thread. The poster asks whether OpenAI or Anthropic could make a lot more money by starting companies that use their own models, instead of selling access to everyone else.

It's a naive question in the best sense, and it has no comfortable answer. We think the "selling access is the tell" reading is mostly right, though it needs some nuance.

The tell: nobody sells the goose that lays golden eggs

When OpenAI launched GPT-5, it said that, when using reasoning, the model was "comparable to or better than experts in roughly half the cases" on an internal benchmark of economically valuable knowledge work, across more than 40 occupations. That is a strong claim. But "half the cases" on a lab-designed test is not "run a company unsupervised." A company is not a task list. It means finding customers, taking on liability, handling the call that goes wrong at 3 a.m., and being right about a market for years.

If the models could do that reliably, the rational move would be to keep them in-house and compound the advantage. Labs do the opposite: they sell the capability to everyone. We read that as revealed preference. It doesn't prove the models are weak. It does suggest that the distance between impressive on a benchmark and autonomous in the market is larger than the marketing implies.

The better explanation is less flattering and more boring

There is a second reading, and we should be fair to it. Labs may not start companies left and right because they are busy with a different bet: being the platform. A platform business and a portfolio of startups are different jobs, with different talent, risks and conflicts of interest. If a lab competed with its own customers, those customers might well switch providers.

Then there is the money. Per the account in a Medium piece, which we'd treat cautiously because it is a blog post built around a chatbot transcript, OpenAI lost nearly $8 billion in 2025 and only 5% of ChatGPT's 800 million users pay. Take those numbers as indicative rather than audited. If they're even roughly right, the company is not sitting on a spare army of founders. It is fighting to cover compute bills, and according to that same piece it has started testing ads in ChatGPT.

Meanwhile the capital is enormous. CNBC reports, citing people familiar with the deal, that OpenAI is raising up to $30 billion at a $1.25 trillion valuation, and that it told investors in February it was targeting roughly $600 billion in total compute spend by 2030. A business spending at that scale needs predictable, recurring revenue now. Metered access to a model delivers that. A portfolio of startups, many of which would probably fail, does not.

What this means for the rest of us

Our read is that the thread punctures two stories at once. The first is the "AI will replace companies" story. If replacement were near, the first companies replaced would be the ones the labs could build themselves. The second is the "AGI makes the economics irrelevant" story. Noah Smith writes that when he raised worries about costs and commoditization with current or former OpenAI people, their response was that none of it matters because OpenAI would be first to reach AGI. That is a faith, not a business plan, and the fact that the company still sells tokens suggests that it hedges too.

The practical lesson is for founders and buyers. If you're building on a model, we'd assume the labs have not found a way to do your job with it, or they'd be doing it. Your edge is distribution, domain knowledge and accountability, none of which a model supplies. And if you're buying, remember that the model is becoming a commodity input: open models reach about 90% of closed-model performance at far lower prices, according to MIT Sloan, though most users still pay for closed ones.

Verdict

The question isn't a gotcha, but it isn't a proof either. Selling access is what a company does when the product is valuable, the costs are brutal and the autonomy is probably not there yet. Treat every claim that models can run businesses by themselves as unproven until a lab bets its own balance sheet on it. When that day comes, we expect they'll tell us by quietly stopping the sales.