AI is shrinking teams

The old Silicon Valley formula was simple: raise a huge round, hire a huge team, and grow into the headcount. In April 2026, Duolingo chief executive Luis von Ahn said the next big product can be built by two people in six months, which is the opposite of that playbook. (youtube.com) That claim landed because Duolingo is not a garage startup talking big. Duolingo is a Nasdaq-listed company based in Pittsburgh, and its 2024 annual report filed with the United States Securities and Exchange Commission shows it operates at public-company scale. (sec.gov) Duolingo had already made this shift visible in April 2025, when von Ahn announced an “artificial intelligence first” strategy and said the company would gradually stop using contractors for work that artificial intelligence could handle. That memo triggered a public backlash because people read it as a hiring story and a jobs story at the same time. (usatoday.com) A year later, the conversation has moved from “can artificial intelligence write code” to “how small can a serious product team get.” The new pitch is that one designer and one engineer, both using models all day, can do work that used to require a product manager, researchers, analysts, and a bench of junior developers. (youtube.com) Y Combinator, the startup accelerator behind Airbnb, Dropbox, and Stripe, is pushing the same idea from the investor side. In a March 2025 interview, Y Combinator president Garry Tan said about a quarter of current Y Combinator startups had 95 percent of their code written by artificial intelligence. (nbclosangeles.com) Tan tied that directly to company size. He said founders no longer need 50 or 100 engineers, and he said some companies were reaching as much as $10 million in revenue with teams of fewer than 10 people. (nbclosangeles.com) That changes the math of a startup before it changes the org chart. If a team can ship with 8 people instead of 40, the same seed round buys more time, the founders give up less equity, and the pressure to hire ahead of demand drops fast. (nbclosangeles.com) The new buzzword sitting under that math is “tokenmaxxing.” In recent creator coverage, the term means treating artificial intelligence token consumption like a scorecard, where more prompts, more model calls, and more automated work are taken as signs of higher output. (youtube.com, youtube.com) A token is the small unit a language model reads and writes, which makes it closer to a meter on a taxi than a seat on a payroll chart. Once companies start paying for work in tokens instead of salaries, managers start asking a different question: not “how many people do we need,” but “how much model usage do we need.” (youtube.com) That is why tiny teams and tokenmaxxing show up together. A smaller team can look incredibly productive if software agents are drafting code, testing features, summarizing meetings, answering support tickets, and generating marketing copy in the background. (youtube.com, youtube.com) But the meter can lie. One recent explainer on tokenmaxxing described cases where heavy users ran up huge bills, including a reported $150,000 month of artificial intelligence compute, and argued that cost per finished task is a better measure than raw token burn. (youtube.com) So the real shift is not that every company suddenly wants fewer employees. The shift is that founders now see headcount as a last resort, and they are trying to buy output with models first, contractors second, and full-time hires third. (youtube.com, youtube.com, usatoday.com)