Self-hosted open source AI for software engineering
Engineering teams face high expenses when relying on third-party commercial AI coding tools like Claude Code and Codex. To reduce these high expenditures, companies need a more cost-effective way to power AI-assisted software development for large groups of developers.
7 different people have asked for or struggled with this since 29 Oct 2026, most recently on 05 Nov 2026. 7 of them in the last 30 days.
- “Revolut now does AI coding entirely with self-hosted, open weight models (like GLM-5.3).”
- “Revolut made self-hosted open-weight GLM models the default coding AI for ~2,000 engineers, on 56 Nvidia B200s.”
- “Revolut moved ~2,000 engineers off Claude Code and Codex to self-hosted GLM on 56 B200s.”
7 people in 7 posts across 1 source · 7 in the last 30 days, 0 in the 30 before
The pain
What goes wrong, and for whom
- Who
- 7 people, in Self-Hosted & Open Source
- Since
- 29 Oct 2026 → 05 Nov 2026
The opportunity
from the same postsWhether this pain leaves room for a product
- Demand
- 7 people raised it · 7 in the last 30 days, 0 in the 30 before
- Looking for something else
- 0% want an alternative · 0% ask for a feature · 100% mention price
- What the numbers say
- Mostly complaints, few people asking for an alternative — the need is clear, the buyer is not yet. Not shown by this data: whether they would pay.
The pain — what people say
Their asks, who is writing, and the posts themselves
What people were doing
Every post behind this page, by what its author was doing — and when each one was written
- About price7100%
Sources
7 signals · 1 sources
Evidence
Raw public posts behind this opportunity — click through to the original
Revolut now does AI coding entirely with self-hosted, open weight models (like GLM-5.3). Cost savings look impressive but the article doesn’t factor in how much they had to pay for their Nvidia GPUs. Hmm […]
Wants a free optionRevolut made self-hosted open-weight GLM models the default coding AI for ~2,000 engineers, on 56 Nvidia B200s. Its bill is now about a tenth of what it paid Anthropic and OpenAI. Worth copying: on the same GPUs, tuning cut time to first token from 14.3s to 1.7s while demand rose 76%.
Wants a free optionRevolut moved ~2,000 engineers off Claude Code and Codex to self-hosted GLM on 56 B200s. Bill ~10x lower. The interesting part: their productivity gains showed up once token limits were lifted. Agents work better when tokens aren’t rationed.
Wants a free optionrevolut moved ~2,000 engineers to self-hosted glm 5.3 on 56 b200s 🏦 ~$0.30 per 1m tokens, about a tenth of the old anthropic/openai bill. claude code + codex are now the exception unlimited cheap tokens > rationed smart ones 👀
Wants a free option‼️ Revolut says it cut AI coding costs about 10x by making a self-hosted Chinese model its default in place of Claude Code. It runs Z ai's GLM on its own Nvidia GPUs for about \$0.30 per million tokens.
Wants a free option
The opportunity — market and score
Products already there, how the score moved, and what it is made of
Opportunity history
Every score change is stored as a snapshot — click a sub-score to overlay it
Why did it move?
Score 38.4 → 38.8 · 05 Nov 2026, 23:00 → 00:15 UTC
How the score was measured38 measured values behind the seven sub-scores · each shows the value, then the points it earned out of 100Show ↓Hide ↑
Questions about this idea
- Is Self-hosted open source AI for software engineering a good startup idea?
- It has an Opportunity Score of 38.8, from 7 mentions by 7 different people across 1 sources. Demand is low (12/100), pain low (11/100) and competition moderate (41/100). This measures what people said, not what they paid for — test willingness to pay before building.
- How many people have this problem?
- 7 different people described it independently, in 7 mentions, the earliest from 29 Oct 2026 and the most recent from 05 Nov 2026. The quotes, with dates and links to the originals, are in the Evidence section.
- How crowded is this market?
- Measured competition is moderate (41/100). Low competition can be a real gap or simply thin data — look for competitors yourself before concluding either.
- Is demand for this growing?
- Status: new — first reported in the last 45 days. 7 people raised it in the last 30 days and 0 in the 30 days before; the status compares its share of everything collected from the same source, so collecting more or less cannot create a trend.
Adjacent opportunities
Nearest neighbours by embedding similarity
Would you use this? Have you tried building it?
Written by readers, except comments marked Team, which are ours. Not part of the evidence above and not counted in any score.
Create a free account to answer — it takes a few seconds and brings you back here. Already have one? Sign in
No comments yet
If you have run into this problem, what did you try?