The Economics of Running an AI Company Are Disastrous

"AI companies are spending untold billions of dollars building out data centers to support behemoth AI models - and a return on their investments is still nowhere in sight. Valuations have soared well past the trillion-dollar mark, with AI chipmaker Nvidia becoming the world's most valuable company with a market cap that recently topped $4.5 trillion. But that's all, despite AI company revenues barely making a dent on balance sheets."

"One key problem stems from AI firms' efforts to fight "hallucinations," one of the biggest and stickiest problems plaguing the industry. To coax more reliable and trustworthy answers out of AI chatbots, companies are using techniques like reinforcement learning - which are driving resource needs up, not down."

"That's coming as consumer demand continues to rise - particularly when it comes to even more resource-intensive text-to-video models, like OpenAI's Sora 2 - and companies like OpenAI and Anthropic will have to pick up massively growing bills. For now, investors have played along, as the WSJ points out, continuing to fund major data center projects and hiring sprees. But how long they'll continue bankrolling an industry with meager returns on investments remains unclear."

AI companies are spending huge sums to build data centers to run extremely large models while returns on those investments remain limited. Market valuations have soared, exemplified by Nvidia's market cap exceeding $4.5 trillion, even as AI company revenues barely affect balance sheets. Large firms are losing substantial amounts, raising concerns about overheated valuations. Efforts to reduce hallucinations use reinforcement learning and other costly methods that increase compute demand. Rising consumer interest in resource-intensive applications, like text-to-video, is driving further cost growth. Investors continue to finance expansion for now, but sustained funding is uncertain without clearer returns. Major data centers could unlock compute and potentially lower long-term costs.

Read at Futurism