Training gets the headlines. Using AI gets cheaper every year.

Building a frontier model gets about 2.4× more expensive every year. Running a model of a given quality goes the other way: it gets about 10× cheaper every year.

Cost to build a frontier model · share of total

Hardware chips, servers, networking47–67%
Research staff29–49%
Electricity2–6%

2.4× / year

Growth in the cost of the largest training runs since 2016.

$1B+

What the biggest single training run may cost by 2027 if the trend holds.

1,000×

Price drop for GPT-3-level output since 2021: from $60 to $0.06 per million tokens.

Our take for founders: you don't need to train a frontier model, and almost no startup can. With AI getting 10× cheaper to use every year, the opportunity is in putting it to work on hard physical problems: robots, materials, satellites, factories.

Sources: Epoch AI (training costs), a16z, "LLMflation" (inference prices).

More Deep Tech 101: Robotics · AI · Space · Czechia

Submit Your Project →

We back deep tech founders in Europe at the idea and MVP stage. About Dynumerics →