It’s Not a Bubble. It’s a Flood
Why the market is dumping SaaS while Nvidia triples
A commenter on Reddit joked that we are buying non-existent RAM with non-existent money for non-existent GPUs, to be installed in data centers that haven’t been built, powered by infrastructure that may never appear, to satisfy demand that doesn’t actually exist and to obtain profit that is mathematically impossible.
He got 309 upvotes. People seemed to love it.
But I actually don’t think it’s a scam. It’s more of a thermodynamic overhang.
Last week, Anthropic released a set of new features that automate work across legal, finance and data services. Within four days, $611 billion in market value evaporated. Thomson Reuters had its worst week ever. Morningstar posted its steepest drop since 2009. Salesforce, HubSpot, Atlassian, Zscaler; all down double digits. Hedge fund exposure to software hit a record low.
And here’s the interesting part: the fundamentals for these companies are actually improving. Earnings projections are going up, so they aren’t dying. But the repricing is happening for a real reason. The world is changing and the indicators are flipping, the same way an overheating engine makes all the sensors scream at once. The readings look random if you’re staring at individual gauges. They make perfect sense if you understand that the system itself is under pressure.
To understand where that pressure comes from, it helps to look at this as a stack.
At the base you have companies like Nvidia producing world-class chips. Above them, the data centers and the foundation models. These layers have been extraordinarily efficient. They’ve compressed rapidly into stable primitives: reliable chips, powerful models, clean APIs. Think of this as a reservoir filling up. So far, so good.
Now all that energy needs somewhere to go. The assumption was that the application layer, Salesforce, Adobe, the entire SaaS ecosystem, would act as the turbine. They would take all that compute and turn it into value. Well, last week, the turbine cracked.
Not because the demand for results is fake. It’s very real. But the demand for middlemen is collapsing. And when the reservoir keeps filling and the turbine can’t handle the pressure, the water doesn’t politely wait. It floods. That flood is what we’re watching right now.
The pressure has created two problems, and both are fascinating.
The first is the startup graveyard. New model updates are constantly reshaping what’s possible, which means startups keep chasing product-market fits that dissolve under their feet. Think of these as falling stars: they burn bright and die fast. In startup world this is usually called a “hockey stick.” But the physical truth is that the sharper the hockey stick, the shorter it lasts. A company that had product-market fit in year one might not have it in year two, because the models just got better and made their product redundant.
A good example is the Swedish startup Lovable. Their own growth manager writes openly about this challenge. In her words, “PMF is changing every few months.” They managed to capture one early wave, and a couple of months later it was basically dead. But investors had already poured hundreds of millions into the project and are now betting that the brand itself will sustain. The startups with enough capital to reinvent themselves as a PR operation and survive on momentum are the lucky ones. The rest will die.
The second problem is even more interesting.
For years, tech companies have mass-recruited layers of people who served as redundant overhead. Not just product people; managers, scrum masters, coordinators, facilitators. I remember sitting in a meeting with seven other people, discussing changes to a portal that only I could build. All of them were redundant. The meeting lasted two hours.
This is how most SaaS companies operate. This is how their internal structure is built. And the problem is that in order to catch the new gradients AI is creating, you need to be flexible. You need to tear down walls and build new ones. But that becomes nearly impossible when you have layers of management with no building skills. They aren’t walls that channel energy; they’re walls that block it.
And the best example goes right back to Lovable. A competing product called Base44 does exactly the same thing, builds websites for you with an LLM. It was created by one person. One guy who sold it to Wix for $80 million in six months. No team, no managers, no meetings about meetings.
There’s a recursive loop here that makes this worse. The moment a startup takes VC money, it needs to raise the complexity of its organizational structure to accommodate the scale that venture capital demands. That complexity means more people, which means more managers, which means more waste heat. It’s thermodynamics: inject energy into a system and the system builds channels to dissipate it. The money doesn’t make the product better. It makes the organization more complex. And that complexity becomes the very thing that prevents the company from adapting when the ground shifts.
Meanwhile, AI keeps getting better, and one or two people with the right tools are outperforming entire teams that need coordination, consensus, and Confluence pages. The gap between what a lean operator can do and what a funded organization can do is flipping. It used to be that money bought capability. Now money buys mass, and mass is a liability.
This isn’t new. The exact same thing played out in the late nineties. Internet infrastructure compressed fast; fiber optics, servers, browsers all hit stable primitives with clean interfaces. The reservoir filled. And then thousands of companies rushed the application layer trying to be the turbine between all that infrastructure and actual users. Pets.com, Webvan, Kozmo. None of them had real channels. They were just energy spilling everywhere.
What happened next wasn’t a bubble popping. It was the system cleaning house. Amazon survived. Google survived. Eventually Facebook. They found specific channels and built them tight enough that energy flowed through with minimal spill. Everyone else was waste heat.
My prediction: the AI application layer will consolidate brutally within 18 months. Not because funding dries up. Not because the technology disappoints. Because this is what systems do. They eliminate inefficient pathways as they find their flow patterns. The infrastructure will survive. The chips will be produced. The data centers will be built. But the companies capturing that energy won’t be the ones we know today.
The Reddit commenter is right about the absurdity but he’s wrong about the conclusion. The demand is real and the money isn’t fake. But the turbines we built to capture it are the wrong shape: the pressure will find new channels as it always does.
If you’re building something in AI right now and you can’t identify the specific gradient you’re channeling, not “AI for whatever” but a real, narrow problem where you’re the tightest path between supply and demand, then you’re not the turbine.
You’re the spill.
If you want to understand the physics behind this, start with my manifesto.