|
A perennial characteristic of Silicon Valley startup companies is that they lose a lot of money, at least at first. That’s what happened to Amazon, Uber, YouTube, etc. But to my knowledge, no tech company has ever burned more cash more quickly than OpenAI.
In 2024, it lost about $5 billion; in the first half of 2025, it lost a reported $13.5 billion; and in the last quarter alone, it lost another $12 billion. For artificial intelligence to ever pencil out, some truly enormous revenue streams will be required—$2 trillion by 2030, according to Bain & Company. As the company at the center of the AI boom (along with Nvidia), OpenAI would represent a sizable chunk of that money.
Faced with this dilemma—where do you get a trillion dollars quick?—OpenAI is getting ready to run hat in hand to the taxpayer for subsidies, like every great Ayn Randian self-created entrepreneur, pulling themselves up by their bootstraps. At a recent Wall Street Journal tech conference, OpenAI Chief Financial Officer Sarah Friar suggested that a government loan guarantee might be necessary to fund the enormous investments needed to keep the company at the cutting edge.
There are very strong reasons to doubt that any company will be able to build the kind of enormous market leverage on which Silicon Valley has historically depended. Indeed, in many ways AI is the opposite of the traditional software or social media business model that prints hundreds of billions of dollars in profits. Microsoft and Facebook, for example, have relatively low capital costs, near-zero marginal costs, and are protected from competition. Each additional copy of Word or new Facebook account costs pennies; government software patents legally forbid people from duplicating the former, while network effects make it nearly impossible to compete with the latter.
ChatGPT, by contrast, has extremely high capital costs, relatively high marginal costs, and is structurally vulnerable to imitation. As I have argued before, its very business model of selling API access to ChatGPT is precisely what you would do if you wanted to sample and thereby recreate the data set used to train it. And any attempt to copyright the data is going to be a stiff ask, since OpenAI did not pay for the data it ingested either. This is apparently what Chinese AI creator DeepSeek did—pay to get ChatGPT to puke itself inside out, and rebuild a very similar model for a fraction of the cost.
|
