In the 1960s, Fairchild Semiconductor produced early versions of transistors for the government. They wanted to expand and wanted to land RCA as their first commercial client. RCA bought vacuum tubes at $1.05 each for their UHF television broadcasting. On the other hand, the transistors were $100 each for Fairchild to produce at low volumes.
The co-founders Robert Noyce (who went on to found Intel) and Jerry Sanders (who went on to found AMD) “anticipated the cheap,” knowing that Moore’s Law (yet-to-be-named by another co-founder of Fairchild and Intel, Gordon Moore) and volume could drastically reduce the production costs over time. Instead of $100, they quoted RCA $1.05, the same price as they paid for vacuum tubes.
“We were going to make the chips in a factory we hadn’t built, using a process we hadn’t yet developed, but the bottom line: We were out there the next week quoting $1.05,” Sanders later recalled. “We were selling into the future.”
– New Rules for the New Economy, Kevin Kelly
They won the contract and were able to produce the transistors for $1.05. They captured nearly the entire UHF tuner market. Two years later, they could reduce the price to $0.50 and still make a profit.
Today, companies like Microsoft are selling GitHub Copilot ($10/month) at a loss. It’s rumored to cost them, on average, $20 a user (source). But, the costs of LLM inference will go down over time. Hardware optimizations, software optimizations, better chip designs, and more chip options will reduce costs. Even with a slowdown of Moore’s law, we’ll be able to run faster, cheaper, and bigger models soon.