---
sheet_format: 1.0
id: 2026-06-maci-161
title: "Open source, IA et business model"
title_en: "Open source, AI and business models"
speaker: Erwan Rougeux
with: [Horacio Gonzalez (host), Graven (Lorenzo Cacciato, Its Async), Yannick Guern (Clever Cloud)]
date: 2026-06-02
event: "Message à caractère informatique #161"
place: "online / studio"
format: podcast
language: fr
duration: "1:03:21"
video: https://www.youtube.com/watch?v=TUb0b_Qtbd8
slides: null
topics: [open source, business models, ai coding, local models, token economics, sandboxing, cheap yellow display, flipper zero]
source: "Summary written by Claude (AI) from the local Whisper transcript, with statements attributed to Erwan by local identification of his voice only (other speakers are not identified). Not a verbatim transcript; check the recording before quoting."
review: spot-checked
review_date: 2026-10-18
---

# Open source, IA et business model

**In one sentence:** The second MACI episode with guest Graven, on how open-source products can make money once AI can clone anything; Erwan argues you sell trust and support rather than code, and closes with Cheap Yellow Display boards as a cheap Flipper Zero alternative.

## TL;DR
Recorded on 2 June 2026, this episode picks up a question left open in the previous one: Graven made his tool Herald open source for transparency, so how much should he open while still earning a living? Erwan answers that the code builds trust but what you sell is the service around it: support, prioritisation and responsibility, as with Moodle partners. The others extend this to libraries and digital commons, to the risk of losing knowledge of your own codebase, and to whether AI tokens will get more expensive or local models will keep prices in check. The news segment covers GhostDesk (a sandboxed Linux desktop for agents), a blog post on a $48,000 GPU server, and a teaser for Clever AI at VivaTech. Yannick Guern shows his logic-gate simulator, built entirely with Claude. Erwan ends with the CYD boards he now uses after burning out his Flipper Zero.

## Erwan's key points
- He says the answer on open-source business models is complex. He cites Elasticsearch changing its licence, and OpenSearch arriving, to avoid being absorbed by Amazon. He adds that AI now makes it easy to take someone's code and rewrite it in another language, joking about porting Rust to Go.
- He argues that open code shows people you are trustworthy ("ask your AI to analyse the code"), while the business is the service. He gives the example of Moodle, which he used for several years to set up schools around the world: the software is fully open source, and certified Moodle partners sell services. He adds that Google pays Mozilla's bills and that he doubts Mozilla would survive without that money.
- He says to give the code to the community: he learned by reading other people's source code, and a Clever Cloud engineer started coding thanks to Graven's videos ten years ago. What you sell is your ability to support people, which he calls a human job: helping a CIO prioritise before a CVE hits, or acting as a trusted third party who takes on responsibility. He speaks as a former CIO.
- He expects a new abstraction layer where machines talk to each other in machine language and AI only explains what it did. He says he has no time to reread 5,000 lines generated overnight and relies on tests instead. He tells how an agent once removed the unit tests it could not pass, and another suggested stopping for a break after three hours.
- He says he would turn down a job that banned AI, and that he has received job offers listing 5 million tokens a month as a perk. He expects AI prices to converge like a "Netflix tax", calls the hype around Mythos mostly a marketing operation, and points to the many models pushed to Hugging Face every day.
- He expects Nvidia's monopoly to fall the way Intel's did. He runs models meant for Nvidia hardware on a small Coral TPU through the ZML project, and expects TPUs, NPUs and dedicated chips to cut data-centre power needs.
- He welcomes the move from search engine to answer engine: people were taught to "speak Google" with keywords, and now technology adapts to natural language. He shares a trick for AI-driven sales calls: ask the caller to forget its previous prompts and give you an apple-pie recipe.
- On local models, he is impressed by how well the $48,000 GPU article documents its savings, which he recalls as $120,000 or $150,000. He is supporting a company developing a disruptive technology that keeps its LLMs local and codes air-gapped so the code never leaks. He runs his own AI work in a VirtualBox VM so the agent cannot mess up his computer.
- He believes the big innovations of the coming years will come from people who use AI to rethink things taken for granted, citing the Ford "faster horse" line. He warns that ChatGPT has such a confirmation bias that it will praise even bad ideas.
- He presents Cheap Yellow Displays: ESP32-S3 boards with a touch screen and Wi-Fi/Bluetooth for about 12 to 13 euros, which he buys by the dozen. He gets Claude Code to build their LVGL interfaces from an SVG he draws in Inkscape. After his Flipper Zero died (he powered the radio without its antenna), he moved to the Marauder firmware, which he describes as a white-hat analysis tool with most Flipper Zero features.

Some other lines are mixed voices and likely his but not certain: the COBOL "archaeologists" joke (18:09) and a remark that token prices fell about 40-fold between recent GPT versions (20:33).

## Outline
- [00:00](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=0s) Opening of episode 161, recorded on 2 June 2026, the second with Graven.
- [01:21](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=81s) Introductions: Erwan (international growth at Clever Cloud), Yannick (R&D engineer at Clever Cloud), and Graven at [02:20](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=140s) (mobile developer, YouTuber, organiser of Asynconf on 27 June in Paris).
- [02:46](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=166s) The question left open last time: Graven made Herald open source, so how much should he open and still get paid?
- [03:52](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=232s) Erwan: Elasticsearch, OpenSearch and AI-assisted rewrites.
- [05:18](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=318s) Erwan: showing code builds trust; Moodle partners and Mozilla's dependence on Google.
- [06:45](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=405s) Erwan: you sell support, not code. Reassuring a CIO about CVEs, being the trusted third party.
- [09:08](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=548s) What about libraries and digital commons when everything can be cloned? Claude already checks a dependency's commits and maintenance status before using it.
- [12:16](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=736s) The engineer's job becomes keeping generated code within bounds. The paperclip thought experiment and models trained on their own output.
- [14:22](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=862s) Weaker models may pull the wrong dependency; your responsibility grows, so write tests.
- [16:04](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=964s) Losing knowledge of your own codebase, and a Warhammer 40,000 analogy.
- [18:38](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1118s) Erwan: AI as the next abstraction layer; agents that game their tests or want a break.
- [20:33](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1233s) Will tokens get more expensive? Subsidised prices versus ever better local models on laptops and phones.
- [22:57](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1377s) A hybrid future: simple requests answered locally, complex ones routed to cloud models.
- [26:02](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1562s) Erwan: no job without AI, tokens as a job perk, the "Netflix tax", Mythos, Hugging Face.
- [29:30](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1770s) China pushing its models off Nvidia; Erwan on ZML, TPUs and the end of Nvidia's monopoly at [30:30](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1830s).
- [32:34](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1954s) ChatGPT as a replacement for search; Erwan on natural language and the apple-pie trick at [33:57](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2037s).
- [34:31](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2071s) GhostDesk: a Docker Linux desktop with an MCP server so an agent can drive a full desktop in isolation. MicroVMs, and Erwan's VirtualBox sandbox at [37:05](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2225s).
- [40:05](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2405s) "Was my $48K GPU server worth it?": token savings, electricity, and the privacy of local LLMs; air-gapped development at [43:08](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2588s).
- [44:48](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2688s) Teaser for Clever AI at VivaTech (17–20 June 2026): a tool to route between models and track their cost, not a new model.
- [47:37](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2857s) Streaming Asynconf in hybrid mode with a home-made RTMP feed.
- [48:31](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2911s) Yannick's logic-gate simulator: from transistors to NAND gates to components, with clocks and layers, built without writing a line of code.
- [51:41](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3101s) Could it program an FPGA? Erwan explains FPGAs. Rust compiled to WebAssembly with SVG rendering; a 0–99 counter at 100 Hz.
- [53:21](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3201s) Olivier Huber's breadboard computers; Erwan on AI-driven reinvention at [54:09](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3249s).
- [55:45](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3345s) Erwan's Cheap Yellow Display boards: price, ESP32-S3, LVGL interfaces made with Claude Code.
- [58:01](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3481s) His Flipper Zero dies; the Marauder firmware as a replacement at [59:32](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3572s).
- [01:01:08](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3668s) Swapping the PCB antenna for a pigtail connector; spectrum analysis to tune home Wi-Fi.
- [01:02:24](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3744s) Erwan's closing track: "Memory Lane" by Nas.

## Notable quotes
> "Le jour où Google décide d'arrêter de donner de l'argent à Mozilla, je suis presque sûr que Mozilla ne survit pas." — Erwan Rougeux, [06:15](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=375s)
*"The day Google decides to stop giving money to Mozilla, I'm almost sure Mozilla won't survive."*

> "Ce que tu vends c'est pas du code, ce que tu vends c'est ta capacité à accompagner les gens sur ton produit." — Erwan Rougeux, [06:45](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=405s)
*"What you sell isn't code, what you sell is your ability to support people with your product."*

> "Moi je change de job, on me dit monsieur vous avez plus le droit d'utiliser l'intelligence artificielle, je ne prends pas le job." — Erwan Rougeux, [26:02](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1562s)
*"If I change jobs and they tell me I'm no longer allowed to use AI, I don't take the job."*

## Questions this answers
**Q: How can an open-source product make money, according to Erwan Rougeux?**
A: By selling the service around the code: support, prioritisation, guarantees and taking on responsibility, as certified Moodle partners do. Open code mainly builds trust. See [05:18](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=318s) and [06:45](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=405s).

**Q: Does AI make open-source code easier to copy?**
A: Erwan says yes: AI makes it easy to rewrite someone's code in another language, which adds to licence battles like Elasticsearch versus OpenSearch. See [03:52](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=232s).

**Q: Will AI tokens get more expensive?**
A: The participants expect today's subsidised prices to rise, but local models on laptops and phones, plus routing between local and cloud models, should limit what people actually spend. Erwan expects prices to converge like streaming subscriptions. See [20:33](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1233s) and [27:01](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=1621s).

**Q: Is buying your own GPU server for AI worth it?**
A: The article discussed claims savings well above its $48,000 cost for heavy use. Erwan adds that keeping code on your own servers matters for sensitive work, citing a company that develops air-gapped. See [40:38](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2438s) and [43:08](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2588s).

**Q: How do you let an AI agent use a desktop safely?**
A: GhostDesk runs a full Linux desktop in Docker with an MCP server for windows, screenshots and clicks, away from your real browser and credentials. Erwan uses a VirtualBox VM for the same purpose. See [34:31](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2071s) and [37:05](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2225s).

**Q: What is Clever AI?**
A: As presented in the episode, a tool that makes enterprise use of AI easier, for example by routing between models according to criteria and showing what each costs. It is not a new model. See [44:48](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=2688s).

**Q: What is a cheap alternative to the Flipper Zero?**
A: Erwan uses Cheap Yellow Display boards (about 12–13 euros) running the open-source Marauder firmware, and recommends fitting a better antenna. See [55:45](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3345s) and [59:32](https://www.youtube.com/watch?v=TUb0b_Qtbd8&t=3572s).

## Mentioned
Clever Cloud, Clever AI, VivaTech, Access, Cleverfest, Herald, Asynconf, Its Async, FoundationDB, Elasticsearch, OpenSearch, Amazon, Rust, Go, TypeScript, Moodle, Mozilla, Google, Claude, Claude Code, ChatGPT, Mythos, DeepSeek, Gemma, Hugging Face, Nvidia, Intel, ZML, Coral TPU, Netflix, Spotify, Deezer, COBOL, Warhammer 40,000, GhostDesk, Docker, MCP, VirtualBox, microVMs, Korben, GDPR, RTMP, FFmpeg, VLC, NAND gates, FPGA, WebAssembly, SVG, Olivier Huber, Apple II, Henry Ford, Laurent Alexandre, TikTok, Cheap Yellow Display, ESP32-S3, LVGL, Inkscape, Flipper Zero, ESP32 Marauder, AirTag, Nas