The world’s biggest tech companies and many new startups are spending billions of dollars to build infrastructure to create large language AI models, but a prominent VC feels there’s a chance the investments never pay off.
Netscape founder and Venture Capitalist Marc Andreessen has said that it’s possible that creating LLMs (Large Language Models) might be a race to the bottom, and the companies creating LLMs never make profits. He said that the investments into LLMs were being made assuming that the company with the biggest AI infrastructure would have the best model, and that model would answer all questions better, and the companies would be able to charge whatever they wanted for the model. He said that in this scenario, LLMs would be have very high margins, and could be a great business.
Andreessen said that there was precedent for such a situation to arise. “That’s what happened with search, where Google ended up in that position,” he said.
But Andreessen seemed to caution that it wasn’t a given that this is how things would play out. “Are all those companies actually in a race to the bottom, in which it actually turns out selling intelligence is like selling rice or something,” Andressen said. “(It could) turn out that anybody can make an LLM — there’s open source LLMs, and there are new LLM startups every day,” he said.
“It (can) turn out anybody can go scrape the Internet, anybody can buy the GPU’s, and then anybody can basically get the same result,’ Andreessen said. “There was a famous Google memo that leaked called “We have no moat, and neither does anybody else”, where they kind of articulate that if you have the same data, you get the same results,” he said.
“(With LLMs) everybody has the same data. In that case, is it a race to the bottom? And by the way, there’s some evidence that it is a race to the bottom because the price per token of generated LLM output has dropped 100x in the last year, which is much faster than Moore’s law. This is the exact opposite of what you’d expect if somebody was going to get a lot of profit out of that business,” Andreessen said.
Tech giants including Google and Microsoft have been spending billions of dollars buying GPUs and building infrastructure to train LLMs. They have been joined by many AI companies, including OpenAI and Anthropic, which are also spending large sums of money in buying infrastructure that’ll let them build LLMs. Brand-new companies are spending massive sums too — SSI, founded by former OpenAI Chief Scientist Ilya Sutskever raised $1 billion to build its AI infrastructure to achieve superintelligence. But while these companies are spending tons of money, other companies, like Meta, are giving away their LLMs for free, and there are other free open-source LLMs that are also producing similar results. As such, there might not be enough differentiation in the LLM space — and Marc Andreessen thinks that this could ensure that there is very little potential to make outsized profits with LLMs.