Korva is a scroll of the internet’s best ideas.
A million tokens costs 1,000x less
Falling prices
The quality of intelligence available in an ordinary chat app today would have been almost unthinkable in 2020, and the price keeps collapsing. That makes intelligence abundant.
Early in the curve
This sits in the early stretch of the curve. More breakthroughs are coming that will return more and more intelligence for the same cost and put it within reach of everyone.
Pointed at science
Intelligence at that price becomes an input for other fields. It can be pointed at the edges of science and technology, including new materials discovery and personalized medicine.
Why did AI prices fall so fast?
Stanford's 2025 AI Index reported that the price of querying a model at the level of GPT-3.5 on a standard benchmark fell from $20 per million tokens in November 2022 to $0.07 in October 2024. Faster chips, more efficient algorithms, and smaller models trained to imitate larger ones all pushed costs down, and competition among providers pushed prices down further. That is a decline of more than 280 times in under two years.
How does a small model learn from a large one?
The technique is called distillation. A large teacher model's output probabilities are used as training targets for a smaller student model, and those probabilities carry richer signals than a single right answer. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean described it in a 2015 paper titled Distilling the Knowledge in a Neural Network.
Why did early chatbots cost so much to run?
Generating each token requires running the input through all of a model's parameters on specialized chips. OpenAI's GPT-3, released in 2020, had 175 billion parameters. Models that size had to be spread across several GPUs, each costing thousands of dollars.
More from Lightspeed Venture Partners
