TypeScript has overtaken Python in open-source AI
Python has been the answer to “what is AI written in” for long enough that the question stopped being asked. For projects created in 2026, it is no longer the answer — by the measure most people care about.
TL;DR — the short version
Among projects created in 2026, TypeScript holds more stars than Python — 3,366,531 against 3,138,643, or 28.3 % of the cohort against 26.4 %. It is the first year that has happened.
Python still ships more repositories, but its lead is collapsing. It was the primary language of 54.0 % of the 2024 intake, 46.3 % in 2025 and 36.0 % in 2026. TypeScript went 13.1 % → 21.3 % → 24.6 %.
Rust more than doubled, from 3.2 % of new projects in 2024 to 7.3 % in 2026, and now holds 10.6 % of the cohort’s stars — more than Go, JavaScript or C++.
What follows for you: the models are still Python. The things being built around the models have moved, and a team whose AI competence is entirely Python is now staffed for one half of this field.
Python has been the answer to “what is AI written in” for long enough that the question stopped being asked. For projects created in 2026, it is no longer the answer — at least not by the measure most people actually care about.
Of the 11,902,530 stars held by projects created this year, TypeScript projects hold 3,366,531 and Python projects hold 3,138,643. Python still ships more repositories — 935 against 638 — but the attention has crossed over.
Two curves, three years
Python’s share of new AI repositories was flat for half a decade — 46.6 % in 2019, 48.9 % in 2022, 51.1 % in 2023, peaking at 54.0 % in 2024. Then it fell eighteen points in two years.
Nothing happened to Python. What happened is that the kind of project being started changed. In 2023 a new AI repository was usually something that touched a model: a fine-tuning script, a notebook, a serving wrapper, an evaluation harness. All of that is Python, because the model libraries are Python.
In 2026 a new AI repository is usually an agent, an MCP server, an editor integration or a tool that calls a hosted model over HTTP — and none of that needs the Python ecosystem at all. It needs a type system, an npm ecosystem and somewhere to run. The labels moved in exactly the same direction at exactly the same time.
| Born in | Projects | Python | TypeScript | Rust | C++ |
|---|---|---|---|---|---|
| 2019 | 596 | 46.6 % | 5.0 % | 0.7 % | 11.1 % |
| 2020 | 614 | 50.8 % | 2.8 % | 2.3 % | 8.8 % |
| 2021 | 535 | 47.7 % | 5.2 % | 2.4 % | 9.7 % |
| 2022 | 734 | 48.9 % | 6.8 % | 3.1 % | 5.3 % |
| 2023 | 1,816 | 51.1 % | 12.0 % | 2.1 % | 2.7 % |
| 2024 | 1,231 | 54.0 % | 13.1 % | 3.2 % | 2.7 % |
| 2025 | 1,485 | 46.3 % | 21.3 % | 4.0 % | 1.5 % |
| 2026 | 2,594 | 36.0 % | 24.6 % | 7.3 % | 1.0 % |
The quiet third curve
C++ went the other way and almost disappeared: 11.1 % of new AI projects in 2019, 1.0 % in 2026. That is the inference-engine era closing. The engines that got written — the runtimes everything else now calls — got written, and new work sits on top of them rather than beside them.
Rust is the one to watch. 7.3 % of new repositories is still a small share, but it is 10.6 % of the cohort’s stars, which means Rust projects in this field are landing considerably harder than their count suggests. Runtimes, local model servers and command-line agents are where it shows up.
What this does not say
It does not say Python is declining. Python projects created in 2026 hold 3.1 million stars; the language is having its second-largest year ever in this catalogue by absolute volume. The share fell because the denominator grew faster.
It also does not say the models moved. They did not, and they are not going to: training and inference libraries remain Python, and the measure here is repositories in the catalogue, which weights a one-file agent the same as a framework.
What it says is narrower and more useful. The centre of gravity of new work shifted from the layer that runs models to the layer that connects them to everything else, and that layer is being written in TypeScript and Rust.
Method
All figures come from one snapshot of the olud.ai catalogue taken on 1 November 2026 at 00:03 CET: 11,394 projects carrying 63,822,390 stars between them. The catalogue is a curated subset of open-source AI on GitHub, not a census of it, so every share quoted is a share of that catalogue and not of the whole forge. Star counts move daily and the figures here will drift; the live numbers are on each project’s page in the catalogue.
The primary language is the one the forge reports for the repository, which it infers from the byte count of source files. It is a single label on a polyglot repository: a Python library with a TypeScript front end is counted once. Shares of repository counts and shares of stars are reported separately throughout, because they differ, and the headline crossover is a stars figure. The 2026 cohort covers ten months of the year, which affects absolute counts but not within-cohort shares. Projects with no language reported — 5.2 % of the 2026 intake — are included in the denominator and in the “other” residual.
Data: olud.ai catalogue, refreshed nightly — method on /methodology.html.
Keep reading
- The words open-source AI stopped using — the same shift, seen in the labels instead of the code.
- Python peaked — where we first saw the curve turn.
- 700 MCP servers, mapped — most of them written in the language that just took the lead.