I’ve been scrolling the same feed for months now. Goldman Sachs on AI and GDP. A protest outside a data center. A LinkedIn post about a job title that didn’t exist two years ago and still doesn’t make sense. Someone’s entry-level role, gone. Someone else, overqualified and waiting. Every week, a different angle on the same underlying thing. And somewhere underneath all of it, the same quiet feeling: this can’t hold.
The numbers that don’t quite add up
AI is, at the moment, a remarkably expensive thing that hasn’t paid anyone back yet. A bottomless pit that just takes, takes, takes and hasn’t really given anything back. The money going in keeps growing because the technology, which keeps getting more complex, demands it, and because no one who’s already spent billions can afford to be the one who blinked first. If they want to stay in the race, they have to keep investing, no matter the cost. According to Stanford’s 2025 AI Index, corporate AI investment reached billions in 2024, more than thirteenfold what it was a decade ago. The big four: Meta, Alphabet, Microsoft, and Amazon are projected to spend nearly $700 billion on infrastructure in 2026 alone. And data center construction is moving so fast that, for the first time on record, the American market is investing more in building data centers than in office buildings.
But none of those billions are coming back yet. A Harvard economist calculated that AI-related investment accounted for almost all of US GDP growth in the first half of 2025. Which sounds amazing, if you don’t look at it too closely. Because when you do, you will see that all of this is just investing in buildings, chips, and power grid upgrades. It’s the infrastructure that supports AI, not AI actually doing anything useful yet. And this last one is actually the one investors were betting on. And as JPMorgan warned, there’s a real boom-bust risk if these do not materialize. So, with the risk of sounding like a broken record, the investment is real, the return is still theoretical.
And the people building it know it. There is something almost poetic about the fact that a senior NVIDIA executive, the company that profits most from the AI infrastructure buildout, recently said that for his team, the cost of compute far exceeds the cost of the people using it. A 2024 MIT study found the same pattern more broadly: AI automation only makes financial sense in about 23% of the roles examined. So for the vast majority of roles, the human was still cheaper. The companies laying people off to cut costs may have skipped a step in the math.
In The Illusion of Endless Growth, I wrote about this hype cycle and what a correction might look like. The dot-com parallel, the deflation of expectations, the idea that what survives might actually be stronger. I still think that’s true. But I wrote it from inside the logic of markets and cycles, and I didn’t spend enough time on the people who fall into the gap between the old system and whatever comes next. That gap is worth naming.
The infrastructure that doesn’t fit
The main thing that was never mentioned in the previous post is that the physical face of the AI investment boom is data centers, and they’re running into their limits of the physical world. Energy demand from AI infrastructure has become significant enough to shape national energy policies, causing utility price spikes and pushing electricity bills higher for ordinary consumers. Communities near proposed sites have been protesting. The permitting and grid capacity bottlenecks are real constraints, not regulatory inconveniences.
The environmental dimension of this is something I explored in The Villain We Needed. The fact that AI’s footprint is real, under-regulated, and exists inside a much broader industrial ecosystem that receives far less scrutiny. The point here is different: the infrastructure required to keep scaling AI is hitting limits that aren’t ideological. They’re physical. Space, power, water, and community tolerance. Something has to give, and it might not wait for the market to ask permission.
The job market that doesn’t add up
Here is one more thing that I find hardest to process, and I say this as someone currently living inside it. Unemployment rates across much of Europe remain low by historical standards. The official numbers suggest a functioning job market. Telling every job searcher that we have the power now, not the companies anymore. And yet entry-level job postings have dropped roughly 35% since early 2023, according to data cited in one of my previous posts, Of Course It Didn’t Work. Mid-level roles are slower to fill and more competitive than they’ve been in years. The people who should be entering the workforce are finding fewer doors. The people already in it are finding the doors narrower.
The official story and the lived experience are pointing in different directions. And I think this gap is not sustainable in the long run. You cannot hollow out entry-level roles, the ones that build skills, that feed the pipeline, that pay taxes and rent, and contribute to the consumer economy, and expect everything to hold indefinitely. At some point, the abstraction catches up with the ground truth.
There’s also something almost darkly comic about the specific shape of the breakdown. Not the dramatic displacement of science fiction with robots at the factory gate, and humans made obsolete, but something quieter and stranger. Entire categories of work are quietly made harder to enter. New job titles appearing to manage the tools that were supposed to make jobs unnecessary. A LinkedIn ecosystem generating roles to clean up after the thing that eliminated the other roles. It looks like a system is improvising as we go.
What it feels like to live inside it
And when I say “Something’s got to give’’ I’m not trying to make a prediction. I don’t know what will give first, if it will happen at all. Whether it’s investor patience, energy infrastructure, regulatory pressure finally landing with teeth, or simply the quiet accumulation of a million individual decisions that together constitute a market correction. All of these are possible. Some combination is probably likely.
What I’m trying to say in this post is something more like a mood. The specific discomfort of watching multiple unsustainable things run in parallel, knowing that systems under this kind of strain don’t usually resolve themselves elegantly. There’s the financial strain; investment at scale, returns still theoretical, and as the Bank of England noted, valuations that look stretched when examined closely. There’s the physical strain; infrastructure running into the limits of power grids, planning permissions, and community patience. There’s the human strain; a job market that doesn’t match its own statistics, and a generation of workers trying to navigate a transition nobody expected.
And there’s something else, harder to quantify: the mental load of living in a moment when every scroll brings another data point pointing in the same negative direction, and yet “the machine” keeps running. That probably deserves its own post (the ways AI is already reshaping how people relate to information, to work, to themselves). But it felt wrong not to mention it here.
Something’s got to give
To me, “Something’s got to give” doesn’t have to mean collapse. It might mean correction, restructuring, a painful but necessary recalibration. The dot-com crash was devastating for many people and productive for the long-term development of the internet. Both things were true at once. That’s probably the most honest template for what’s coming here.
But I think it’s worth sitting with the “got to” part of that phrase. Not “something will eventually adjust” but “something cannot continue as it is“. The investment math requires returns that haven’t materialized in proportion. The infrastructure requires space and power that are increasingly contested. The job market is generating a gap between its official numbers and what people are actually experiencing. These aren’t independent anxieties. They’re signals from the same underlying tension.
I wrote in The Illusion of Endless Growth that maybe the bubble needs to burst so we could stop chasing magic and start building meaning again. I still believe that. I just have more clarity now about what the burst might cost, and who is most likely to pay it. Not the big companies leading the AI race. Not the investors who got in early. The people in the gap. The ones whose entry-level roles quietly disappeared, whose mid-level applications sit unanswered, whose lived experience is running ahead of the statistics that are supposed to describe it.
Something’s got to give. The question worth asking isn’t just what, it’s who absorbs it when it does.