Racem Flazi – Le Game

Open source LLMs and the AI cartel, with Karim Jouini on Racem Flazi’s podcast

Nvidia’s CEO steps down, and three French tech founders ask what it means for open source LLMs. Jensen Huang was the biggest backer of open source in AI. For the hosts, his exit looks less like a scandal and more like a power struggle with the closed-model camp they call the AI cartel.

In this episode of Racem Flazi’s podcast, Racem Flazi, co-founder of LegalPlace, talks with his two regular guests: Karim Jouini, our CEO and co-founder, and Wassim Jouini, CTO of LegalPlace. Over an hour and a half they cover AI-found security flaws, software rebuilt by agents in days, a dentist who automated his front desk, prompts as intellectual property and the day Claude went down. This page sums up what they said, with the figures quoted and the moments worth hearing.

Podcast
Racem Flazi – Le Game
 - episode
Host
Racem Flazi
Guest(s)
Karim Jouini (Thunders) and Wassim Jouini (LegalPlace)
Published
November 17, 2026
Duration
86 min
 ·
French
Listen
YouTube
 ·
Racem Flazi – Le Game

Watch the episode: open source LLMs, Claude dependency and the AI cartel

An episode for founders, CTOs and engineering leads deciding how much of their know-how to hand to a single AI vendor. Short on time? Start with the Jensen Huang affair at 01:07, prompts as intellectual property at 42:38, and how Thunders works on its customers’ own models at 58:31.

Chapters

  • 00:00 – Intro
  • 01:07 – The Jensen Huang affair at Nvidia: coup or scandal?
  • 15:20 – Cybersecurity: is everyone vulnerable now?
  • 24:13 – Reverse engineering: the end of proprietary software?
  • 35:02 – Use case: AI in small businesses
  • 42:38 – Intellectual property: can you really trust OpenAI?
  • 52:53 – Too dependent on AI
  • 1:02:03 – Weight-space programming, or brain surgery for models
  • 1:09:30 – A Democratic win in the US: the end of the data centre boom?
  • 1:18:38 – Yann LeCun’s latest research
  • 1:21:39 – Viewer questions: can Chinese models overtake American ones?

Scroll inside the list to see all chapters.

The episode in 2 minutes

The trio now meets every two weeks, because three weeks of AI news turned out to be too much to cover. This time the headline is Jensen Huang’s resignation from Nvidia. The hosts doubt the official story and read it as a coup against the one big CEO funding open source, at a time when a handful of closed-model vendors are pulling up the drawbridge.

The rest of the episode follows one question: what happens to companies that depend on those vendors?

  • AI now finds security flaws faster than anyone can patch them, and a $129 million fraud shows where the weak links are.
  • Agents rebuild software from the outside, from a 200-service AWS emulator to a family business’s old ERP.
  • Prompts and skills may be a company’s real intellectual property, and Karim says half of Thunders’ engineering is prompts.
  • When Claude went down for half a day, teams found out how dependent they were. Revolut had already moved to an open source model.

Jensen Huang’s exit: what’s at stake for open source LLMs

The episode opens on Nvidia. Jensen Huang has resigned after revelations about his private life, and while they were recording, Racem says, parts of the story were already falling apart. Whatever the facts of the scandal, Karim reads the outcome as a coup: companies like Nvidia, Anthropic or OpenAI are now more powerful than most mid-sized states, and the fight for control looks like politics.

Wassim points to what made Huang awkward for some. He was the only dissenting voice against the US “cartel” the trio discussed in earlier episodes, a group of labs heading towards closed, proprietary models. Huang had taken the opposite path and was pushing open models for everyone. Racem draws the conclusion: kill open source in the US and it will thrive elsewhere, with China at its centre.

Who captures the margin: Nvidia or the model vendor

Karim puts numbers on why this matters, using an article from Revolut about its move to open source models. When a company runs open source at scale, around 80% of the margin goes to Nvidia, and the data centre and electricity share the crumbs. With a proprietary model, 70 to 80% of the margin goes to Anthropic or OpenAI instead. The choice of model decides who captures the value AI creates.

AI finds security flaws faster than companies can patch them

Three weeks earlier, a zero-day in PHP shook the web. Wassim notes it hit one of the most tested pieces of software around, which runs about 70% of the internet. In Finland and France, some government services were switched off until they were patched. Karim sums up the situation: it’s as if someone now had a key that opens every lock, and changing all the locks takes time. He recalls the NSA advising PHP site owners to go offline while it was fixed.

Wassim adds a nuance. Many flaws are being found, few are exploited yet, because attackers aren’t organised enough to hit everything at once. The visible targets go first: government sites, retailers, large companies. The rest of the wave will come with cheaper models and better-organised groups.

A $129 million fraud built on a phone number

Karim tells the case of three people charged with stealing $129 million from US and UK mid-sized companies through CEO fraud. The deepfaked voice and face are the part everyone expects. The detail that struck him is smaller: they hacked a law firm’s website just to change its phone number. When the CFO looked up the firm on Google and called to check, he reached the fraudsters. The firm never noticed. For the next 12 to 18 months, the hosts expect attackers to keep the edge.

Reverse engineering: rebuilding software in days, not years

Racem raises Anthropic’s announcement of its own office suite, offered for free, a direct shot at Microsoft. Karim saw it coming. What surprised him that week were two other cases. Amazon used AI agents to build a self-hostable clone of AWS: about 200 services you can run on a laptop to develop and test locally. And an engineer on the Thunders team rebuilt his family business’s old, unmaintained ERP. With Codex, Claude and reverse-engineering prompts, he turned it into a web app his family was due to start using the following week. He isn’t stealing anyone’s IP, Karim says. He has business processes, they live in an outdated tool, and he’s moving them somewhere modern.

“We’re becoming Chinese companies”

Wassim describes how LegalPlace does it: give the agent the existing tool plus a list of non-negotiable user stories, and you have something you can roll out within days. Racem mentions one tool that cost them €30,000 to €40,000 a year for a single business unit. “We’re becoming Chinese companies,” Wassim says, and Karim explains the comparison with what he saw in China this summer: companies there own their stack and buy specialised software with its source code instead of subscribing to a standard SaaS product. Thunders is replacing several small tools the same way, he adds, despite being smaller than LegalPlace.

A dentist, a browser extension and 1% adoption

Racem’s case is a dentist friend whose assistants spent their days on patient admin, with six or seven insurers each running their own process. No code, no development: through Claude’s browser extension, the dentist asked it to text tomorrow’s patients, confirm or remove each appointment based on their answer, then find the right insurer portal for each patient file. It worked the first time. Racem estimates real adoption at 1 or 2% of the market.

Karim sees the other side. Today it’s sold absurdly cheap compared with the value it creates. If it saves the dentist ten thousand a month, there’s no reason he keeps paying 30 euros. Expect prices to move.

Who owns the know-how you hand over to Claude?

The dentist case raises a legal question: an American company now holds European patients’ dental data. Karim describes a legislator outpaced by a tool that is everywhere at once. He cites a French bill amending the Darcos law, which would reverse the burden of proof: if a model produces work that resembles yours, the AI company would have to prove it didn’t train on it. He thinks it’s unenforceable because of the US, but it starts the right debate.

Your prompts as intellectual property

Karim asks whether, with AI, a company’s skills and prompts are its intellectual property of the future. At Thunders, he says, about 50% of the engineering is prompts: hundreds of them, the tricks that make a large account say “this tests well”. If a model vendor took them, it could build a Thunders in a week, and proving it would be close to impossible. Racem adds the case of companies whose employees plug Claude into internal systems without anyone knowing, so the secrets are absorbed anyway.

“Do not store”, unless you call sales

Wassim reminds listeners that on personal accounts, most chat tools are allowed by default to reuse conversations to train their models. Karim gives an example he invites everyone to check: on OpenAI’s Google Cloud offer, he says the documentation states that the “do not store” setting is ignored until you’ve had a call with sales. Wassim’s three levers: switch on every privacy setting, push for zero data retention to be the default rather than an enterprise clause, and keep your skills portable so you can move to an open source model tomorrow.

Claude goes down: why teams are switching to open source LLMs

The episode was recorded on Friday 13 November, and it doubled as a stress test. Claude was down for half the day and Microsoft was partly down too. Wassim checked who pinged him at LegalPlace during an earlier Claude incident, and got buried in messages. As Karim puts it, when Claude cuts out you’re not back two years, you’re back ten: people have uninstalled their IDEs.

Wassim points to Revolut, which moved its coding to a Chinese open source model, GLM, run through OpenAI’s Codex, for about a tenth of the cost. He puts that model roughly at the level of Claude Opus 4.5 or 4.6. He sees it as the first big public case of a trend many companies are preparing quietly. Asked whether Chinese models work in French, Karim says he uses Kimi and DeepSeek in French daily, and they do.

Why Thunders runs on its customers’ own models

Racem asks Karim directly: is testing safe from being absorbed by the big labs? Karim doesn’t think so. Amazon used its sales data to launch Amazon Basics; dominant players move into adjacent markets, and testing is adjacent. He has no doubt they’ll look at it.

That’s also why model choice is already settled for him, and his customers decide it. His large accounts care about personal data and their own IP, so they all run open source models or have special contracts with Google or Anthropic to host the model themselves. One very large US customer self-hosts Gemini. One of Thunders’ differentiators, he says, is evaluating a customer’s model in one or two hours: which scenarios it handles, which it doesn’t. Thunders had just signed a large European defence manufacturer that provides its own model for that evaluation. Testing itself is changing too: you can still force the agent on the tests that matter, and ask it to find bugs and propose test plans with reproduction steps for the rest.

6 things to remember about open source LLMs and AI dependency

  • The model you choose decides who captures the margin: around 80% goes to Nvidia on an open source stack, 70 to 80% to the vendor on a proprietary one.
  • AI finds flaws faster than teams can patch them. Patch what you can, turn on two-factor authentication, and watch the boring weak links like a website phone number.
  • Rebuilding a niche tool from its user stories now takes days, not a project. You still own the maintenance and the security.
  • Your prompts and skills are know-how. On consumer accounts, they may be training someone else’s model by default.
  • Zero data retention should be the default, not something you unlock with a sales call.
  • Large accounts already run open source or self-hosted models, so the tools they buy have to work on their model, not the vendor’s.

Key moments from the episode

  1. “Companies like Nvidia, Anthropic or OpenAI are more powerful than most mid-sized states.” « Des sociétés comme Nvidia, Anthropic ou OpenAI sont plus puissantes que la majorité des États de taille moyenne. » (04:06)
  2. “It’s as if someone now had a key that opens every lock.” « C’est comme si on avait quelqu’un qui a une clé qui ouvre toutes les serrures. » (13:52)
  3. “We’re becoming Chinese companies.” « On devient des boîtes chinoises. » (30:36)
  4. “In the world of AI, aren’t my skills the intellectual property of the future?” « Dans le monde de l’IA, est-ce que ce ne sont pas mes skills qui sont la propriété intellectuelle du futur ? » (43:18)
  5. “Our engineering is still 50% prompts.” « Notre ingénierie contient quand même 50 % de prompts. » (43:50)
  6. “We shouldn’t have to call a salesperson to ask them not to store our data.” « On ne devrait pas appeler un sales pour lui demander de ne pas enregistrer nos données. » (51:27)
  7. “When Claude cuts out, we’re not back two years, we’re back ten.” « Quand Claude coupe, on n’est pas revenu il y a deux ans, on est revenu il y a dix ans. » (55:02)
  8. “All my large accounts run on open source models.” « Tous mes grands comptes sont sur des modèles open source. » (59:44)
  9. “We can evaluate our customer’s model in one or two hours.” « On est capable d’évaluer en une ou deux heures le modèle de notre client. » (1:00:21)

About Racem Flazi’s podcast

Racem Flazi – Le Game is the French-language podcast of Racem Flazi, president and co-founder of LegalPlace, the French platform for starting and running a company. His AI news session with Karim and Wassim Jouini now runs every two weeks. This episode follows The AI bubble: the first cracks, where the AI cartel first came up. The podcast is available on YouTube, Apple Podcasts and Spotify.

About Karim Jouini

Karim Jouini is our CEO and co-founder. An AI engineer trained well before LLMs, he spent seven years at Microsoft before founding Expensya, sold for more than $100 million. At Thunders he builds test agents that have to run on whatever model the customer picks, often an open source or self-hosted one, which is why self-hosting with your own LLM provider is part of the product. When he talks about dependency on a single AI vendor, he’s describing a constraint his customers set for him.

About the other guest

Wassim Jouini is CTO and Head of AI at LegalPlace, holds a PhD in artificial intelligence and previously worked at Microsoft. In the episode he explains the PHP flaw, the Revolut move to an open source model and the three ways to protect your data from AI vendors.

Frequently Asked Questions

Why are companies moving to open source LLMs?

Cost, control and independence. The episode cites Revolut, which moved its coding work to GLM, an open source Chinese model, run through Codex, for about a tenth of the cost. Open source also avoids depending on one vendor that can raise prices or absorb your know-how, and large accounts increasingly require it for data and IP reasons.

Are open source models good enough for real work?

For many uses, yes. In the episode, Wassim places the model Revolut moved to around the level of Claude Opus 4.5 or 4.6, a generation behind the frontier but enough for coding. Karim says he uses Kimi and DeepSeek in French without problems. The frontier models are better, but not by enough to change the game for most tasks.

How can a company protect its intellectual property when using Claude or ChatGPT?

Wassim gives three levers. Switch on every privacy setting, since consumer accounts usually allow training on your conversations by default. Get zero data retention in your contract and push for it to become the default. And keep your prompts and skills portable, so you can move to an open source or self-hosted model without losing your know-how.

What is zero data retention?

It’s a contractual commitment from an AI provider not to store your prompts or outputs. In the episode, Wassim explains that the provider may keep some logs for a week or two to detect abuse, and nothing else. Today it usually comes with an enterprise offer negotiated with sales. He argues it should be the default.

Why is AI making cyberattacks more dangerous right now?

AI finds flaws in old, heavily tested software faster than teams can patch them, as the PHP zero-day showed. Attackers aren’t yet organised enough to exploit everything, so visible targets go first. The speakers expect attackers to keep the edge for 12 to 18 months: patch everything, turn on two-factor authentication and watch for small weak links.