What Happened
Confidential financial documents from OpenAI and Anthropic reveal the staggering economics behind AI's biggest labs ahead of their IPOs. OpenAI expects to spend $121 billion on computing power for AI research by 2028, projecting $85 billion in losses that year even after nearly doubling sales. Anthropic's forecasts tell a similar story of mounting training costs. Both companies report two measures of profitability — one excluding training costs (where both are near breakeven) and one including them (where OpenAI does not expect to break even until the 2030s). Revenue is exploding for both, driven by enterprise adoption and coding tools, but inference costs still eat more than half of revenue. Both companies will burn through massive amounts of cash and are counting on IPO investors to sustain their businesses.
My Take
This is the most important article about AI economics published this year, and every builder should read it carefully. The numbers tell a story that the product demos never will: the AI models you depend on are being sold at a loss, subsidized by venture capital and soon by IPO investors who are betting that scale will eventually bend the cost curve. OpenAI burning $85 billion in a single year is not a business model. It is a faith-based bet that intelligence gets cheaper faster than the appetite for it grows. And the two-books accounting — profit if you ignore training, massive loss if you include it — is the kind of financial framing that should make anyone paying attention uncomfortable. For builders, this changes how you should think about platform risk. When your AI provider is losing money on every query you send, your pricing is not stable. It is subsidized. And subsidies end. The companies that survive the next five years will be the ones that understood this: the tools are artificially cheap right now, the providers are racing to lock in market share before they have to charge real prices, and the moment IPO investors start demanding profitability, the economics of every AI-powered product change overnight. Build with that in mind. Understand your actual compute costs. Do not architect your business around pricing that cannot last. And pay close attention to which provider reaches sustainable unit economics first — because that is the one most likely to still be standing when the music stops.
Read Original Source