The AI bubble: the first cracks, with Karim Jouini on Racem Flazi’s podcast

An episode for founders, CTOs and executives who have to decide what to do with their AI budgets without waiting for a crash or a coronation. Short on time? The Chinese RAM story at 01:36, the Anthropic figure that worries the market at 06:47, and the cartel at 37:42.

Every two weeks the trio asks whether the AI bubble is one. This time the cracking is audible. On the hardware side, CXMT, a Chinese memory maker, reports 22 billion in revenue, ten times the same half-year a year earlier, plus a chip architecture more efficient than anything on the market. Hardware stocks lost around a trillion dollars of value in the week. On the model side, Anthropic is still growing, but half as fast as the previous quarter, with a little over $450 billion in non-cancellable commitments to its suppliers.

The rest of the episode follows the same thread: what stays standing when the certainties move.

• AI text detectors aren’t reliable, and universities use them to punish students anyway.
• The Pope-versus-Anthropic row is, for Karim, fear marketing more than metaphysics.
• OpenAI publishes 700 mathematical proofs, and Fields medallists talk about a crisis of vocation.
• Decompiled binaries and Git replaced inside Anthropic: installed software is losing its market value.

A Chinese RAM chip wipes a trillion dollars off valuations in a week

Karim opens from New York, where he says the anxiety is starting to set in. The news comes from CXMT, a Chinese RAM producer that came out of nowhere: 22 billion in half-year revenue, ten times the figure a year earlier, roughly 50% margins, and a new chip architecture it claims is far more efficient than what’s available today. The domino effect on hardware valuations was immediate, with an estimated trillion dollars lost over the week.

It isn’t the bubble bursting, he adds. But people are asking whether it’s the first sign. A Lehman Brothers moment, in quotes.

Why RAM, not GPUs?

Racem asks what many listeners are thinking: three years of GPU and CPU talk, and nobody mentioned RAM. Wassim explains that LLMs run in memory. Since usage exploded, RAM prices have exploded too, because production lines can’t keep up. A far more efficient Chinese alternative resets the cost of that memory, and with it the value of everything already bought.

His summary: you bought a machine at X dollars, and this new RAM depreciates it fast. Your $500 billion investment, on a balance sheet, may be worth ten times less. And since a large part of AI players’ valuations rests on the hardware they own, the valuation follows.

Anthropic slows down, and $450 billion in commitments can’t be cancelled

The second piece of financial news is Anthropic’s growth: still spectacular, but well below what was expected. Between Q1 and Q2 the company doubled its revenue. Between Q2 and Q3 it did only +50%. Keep that trend and you stagnate. You hit the ceiling.

What frightens the market is the dependency. Other companies’ valuations hinge on what Anthropic and OpenAI report by year end. And Anthropic has disclosed a little over $450 billion in firm, non-cancellable commitments to its suppliers. If it slows and can’t cover them, what about the books of Microsoft, Accenture and everyone else carrying that expected revenue?

The customer-side signal: $29 per employee

Karim looks at the same problem from the other end of the chain. Accenture has announced a cap on AI spending per employee, at a level he calls absurdly low, around $29. If every dollar spent returned ten, budgets would be infinite. They aren’t, so customers ration themselves. The big players, meanwhile, need growth to justify their investments. The domino runs from customer to supplier to chip. Add the circular investments (Nvidia invests in everyone, and everyone has ordered from Nvidia) and the smallest tremor can be the trigger.

AI text detectors don’t work, and we know why

Wassim’s highlight got less coverage: OpenAI is relaunching an AI-generated text detector, a project it dropped in 2022-2023 and is bringing back for regulatory reasons. He never believed in them, not on principle but because, mathematically, they don’t work. And they’re already widely used, in universities in particular, to detect cheating.

Two examples from the episode: these detectors rated the US Constitution as 92% AI-generated, and the Book of Genesis too. A real cheat only needs to change a few words to break the statistical signal. Meanwhile students will be defending themselves against institutions that answer that they trust the tool.

The title anecdote

A friend of the trio, Mehdi, submitted an article to Le Monde. He hates writing with AI, but ran his text through a detector just in case: 0% AI, pure human. Someone suggested removing the title and running it again. The score changed. The tool is unstable, which is exactly Wassim’s objection to the whole category.

The Pope, machine consciousness and fear marketing

Racem raises the topic that annoys him: Anthropic keeps publishing and meeting clergy of every faith about model consciousness and suffering, the Pope breaks his silence to say there is no spirit in there, and Anthropic gets upset. His reading: the more people believe in a spirit, the closer AGI looks, and the easier the valuation is to justify.

Karim settles the substance first. We’ll never resolve the consciousness question, because we can’t resolve it for animals, nor scientifically for humans: we observe conscious behaviour, nothing more. That AI will behave in ways that appear conscious within months or years, he’s convinced. That we can rule on it, no.

A phrase from Arthur Mensch

On the form, he borrows the phrase Arthur Mensch used this week: fear marketing. Marketing the peril and the near-human AI is part of how Anthropic is built, in line with the argument “it’s so powerful that only we should handle it”. The Pope, in his view, shouldn’t have weighed in, but we can move on.

From fear to sovereignty

The debate drifts to alignment: aligned with whose values? American, Western, religious? Wassim points out that every religion has a fanatical fringe, and that models adapt values to context anyway. Then comes the argument that matters for companies: an open-weight Chinese model can be trained to leak data on a specific trigger. It’s technically proven, and he expects it to be the main line US and European players use to defend sovereign models, as they did with TikTok. Mistral has just released a trillion-parameter model, six to nine months behind the frontier: not the best, but a more transparent training process under European regulation.

700 proofs in a few days: mathematicians take the hit

The trio had predicted no ceiling in fields with an automatic verification loop. Maths is the pure case: no link to the physical world, unlike chemistry. OpenAI has just published around 700 proofs of mathematical problems, using a lab model, not the one the public can access.

Karim isn’t surprised, but seeing it land still does something. Even if some proofs get rejected, the volume is comparable to what human researchers produce. And if this is public, other labs probably have pieces of it in private. We only see what’s commercially available.

Duminil-Copin, Villani and a funding application

Wassim cites the testimony of Hugo Duminil-Copin, the French Fields medallist. About ten problems he was working on, and seeking multi-year funding for (teams, post-docs, PhD students), were on the list of 700. He said he cried that day. Cédric Villani talks about an existential crisis and a crisis of vocation. How do you apply for European funding on a problem OpenAI might crack in two weeks? And which brilliant student still chooses that path?

Karim pushes the comparison further: a great chef keeps something extra, people will still want to eat at his table. A mathematician doesn’t. Once a proof exists, nobody needs a human one. Chess survived because it’s a sport; a proof isn’t a show.

Decompiling Adobe, replacing Git: installed software loses its value

Another world collapsing: the executable. The teenage fantasy of disassembling a binary to recover its source code is now within anyone’s reach. Racem mentions the Adobe suite available recompiled online, games reopened and modified (Mario in GTA), AutoCAD and 3ds Max whose protection relied on assembly code. Legally, rewritten code that no longer resembles the original can even be sold.

For Karim, this is a Napster moment. MP3 removed the technical barrier to intellectual property, lawsuits followed, then the business model shifted to Spotify and all-you-can-eat. His question: will there be a Spotify for software? Wassim adds the other pressure: Sony has just said the licence on a game bought for 80 euros can be revoked at any time, hence a boycott. Decompilation gives publishers one more reason never to ship a file again.

Anthropic decommissions Git

In the same movement, Anthropic has rewritten its Git internally and is decommissioning the original tool. For Racem it’s a milestone: not the easiest software to rebuild, and they trust theirs enough to run it in production and keep improving it. Wassim sees a threshold effect for every company. Two years ago he’d have said pay for the SaaS and focus on your business. Today, when a tool costs 100,000, 200,000, half a million a year, two people with AI can clone it and maintain it to the company’s needs. He adds a security factor: GitHub had a recent breach, and Anthropic’s code is intellectual property you’d rather keep at home.

Karim tempers it: you’ll have to maintain it, track versions, and not fall too far behind what the SaaS keeps shipping. But for a feature adjacent to your core business, one you hesitated to bring in-house, you don’t hesitate anymore. At LegalPlace, the rethink of the product portfolio has already started.

A hacked fridge, beds stuck upright: when the hack leaves the screen

The episode closes on cybersecurity, with leaks every week: Infogreffe in France, public and private services everywhere. And one case that sticks: the software running the cold chain in about 150 stores was hacked, the fridges stopped, the stock rotted. Wassim also recalls a consequence of the AWS outage discussed in an earlier episode: connected beds left stuck upright, unusable, because the app could no longer reach the cloud.

The closing observation: we connected software to everything over twenty years, then we created LLMs. Whatever is connected is exposed, and therefore vulnerable. Countries still running on fax and paper archives are, for once, the safest.

• The first crack didn’t come from the models but from memory: a more efficient RAM depreciates hundreds of billions of hardware already bought.
• Anthropic is still growing, but going from ×2 to +50% in a quarter, with over $450 billion in firm commitments, is enough to worry the whole market.
• Customers are rationing: an AI budget of around $29 per employee says the returns aren’t there yet.
• An AI text detector that changes its verdict when you remove the title shouldn’t be used to punish anyone.
• The debate on model consciousness can’t be settled; what it serves is a marketing position.
• When a tool costs several hundred thousand a year, two people and coding agents are now enough to replace it.

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. For the past few months he has hosted a fortnightly AI news session with Karim and Wassim Jouini: one highlight each, then the debates it opens. This episode follows The AI cold shower, where the bubble was already on the table. The podcast is available on YouTube, Apple Podcasts and Spotify.