Notes from the Trough: a developer's guide to post-hype AI

#ai #opinion #futurism

I want to make a prediction that sounds pessimistic and is actually the opposite: the most exciting decade of AI starts after the hype dies, and the hype is dying right now.
This isn't a hot take anymore, it's the consensus reading of the data. By spring 2026 the analysts who track these things put generative AI squarely in what Gartner calls the trough of disillusionment — the part of the cycle where reality finishes mugging the marketing. The widely-cited number is that around 95% of enterprise GenAI pilots never made it to production. McKinsey's surveys put the share of AI projects actually delivering core profitability under 40%. Meanwhile the foundation labs hit eye-watering valuations without a clear story on when the unit economics close.

Generative AI is in the trough. AI agents are sitting exactly where GenAI sat two years ago — at the peak, right before the drop.
So the "is it a bubble" question has, frankly, gotten boring. Of course there's a bubble. There's always a bubble. The interesting question — the one worth being a futurist about — is what survives the trough, and what the post-hype stage actually looks like for the people who build things.
Let me try to answer that with specifics instead of vibes.
The analogy everyone reaches for, and why it's half wrong
The comforting story is fiber optics and the dot-com crash. The internet bubble burst, fortunes evaporated, and yet the infrastructure that got overbuilt — the dark fiber, the data centers, the protocols — turned out to be the foundation of the next twenty years. The bubble was real and the technology was also real. Both things at once.
That story is true and it's also a trap, because it smuggles in a happy ending. Here's the part people skip: the telecom fiber boom destroyed something like two trillion dollars of equity value on its way to being useful. The infrastructure survived. The capital that funded it mostly did not. The companies that built the future were, overwhelmingly, not the companies that profited from it.
Now look at the AI numbers with that lens. The five largest US cloud players are guiding toward something like $635–690 billion in capital expenditure in 2026, roughly double 2024, with about three-quarters of it going to AI infrastructure. Goldman's math suggests that to maintain their historical returns on that spend, these companies would need to generate over a trillion dollars in annual profit — more than double current consensus. Amazon's free cash flow is projected to turn negative this year. Morgan Stanley expects hyperscaler debt issuance to blow past $400 billion. Three of the four big hyperscalers lost market value after their most recent earnings calls, specifically on capex anxiety.
That gap — between what's being spent and what's being earned — doesn't get closed by a better prompt. It gets closed by a correction. So here's prediction number one, and it's the unglamorous one:

The infrastructure is real. The financing structure is not. Expect the first serious write-downs and at least one high-profile collapse in the "neocloud" GPU-as-a-service layer within roughly 18 months. The data centers will keep humming under new ownership at cents on the dollar. AI doesn't die in the trough. A specific generation of AI capital does.

This is the fiber lesson applied honestly. The picks-and-shovels survive. The people who bought the picks at the top do not.
The disappearance
Here's the prediction I'm most confident about, and it's the one that matters most if you write software for a living.
"AI" is going to stop being a category.
Think about what happened to the word "internet." In 1999, "internet company" meant something — it was a kind of company, a sector, a thing you put in a pitch deck to add a zero. By 2008 it was meaningless, not because the internet failed but because it won so completely that describing a company as an "internet company" became like describing it as an "electricity company." Every company used it. It dissolved into the substrate.
AI is on exactly that path, and the trough is the moment it accelerates. When everyone has the capability, the capability stops being a differentiator. The label peels off. My specific bet:

By around 2028, "AI startup" reads the way "internet startup" read in 2008 — quaint, and slightly suspicious. The serious builders stop leading with the model. They lead with the problem, and the model is plumbing. The companies that survive the trough are the ones that already think of AI as a component, not a personality.

You can already feel the gravity of this. The interesting work is quietly migrating away from the foundation models — which are commoditizing fast, with open-weight options closing the gap and gross margins at even the top labs reportedly sitting in the unglamorous 30–40% range — and toward the unsexy layer around them: integration, orchestration, verification, the boring glue that turns a probabilistic text generator into something you'd let near a payment system.
The constraint moves
Every technology has a binding constraint, and you can read the maturity of a field by watching which constraint it's currently fighting.
For the last three years the constraint was capability: can the model even do the thing. That fight is mostly over. The models can do an astonishing amount. The constraint has already moved, and it's moving to three places at once:
Power. This is not a metaphor. The headline bottleneck for the AI buildout is increasingly electricity and grid interconnection, not talent or algorithms. Data center siting is starting to be driven by where you can get a gigawatt, full stop. Whoever was telling you AI was an abstract software story was describing 2023. The 2027 story is transformers — the heavy iron kind, in substations.
Trust and verification. A model that's right 95% of the time is a miracle in a demo and a liability in production, because the 5% arrives silently and confidently. The next stage of value isn't a smarter model, it's a trustworthy system built around an imperfect one: the boundaries, the checks, the audit trails, the "refuse to proceed when the state is wrong" machinery. The bottleneck shifts from intelligence to accountable intelligence.
Distribution and data. When raw capability commoditizes, the moat reverts to the things it always was — who owns the channel, who owns the proprietary data, who already has the users. Capability was a temporary moat. It's evaporating, and the old moats are reasserting themselves underneath.
Agents: the next thing to deflate
A specific sub-prediction, because this is where today's hype is loudest. As of the most recent Gartner cycle, AI agents sit right at the peak of inflated expectations — exactly where generative AI sat two years before its fall. You know what that means. The peak is the part right before the drop.

The "autonomous agent that does your whole job" narrative deflates hard in 2026–2027. What survives is narrower and more useful: agents that operate inside tightly constrained, verifiable domains, with explicit guardrails on what they're allowed to do and hard checks on what they produced. The fantasy of the general autonomous worker recedes. The reality of the bounded, supervised, contract-bounded agent quietly ships and actually works.

The pattern repeats at every level: the grandiose version dies in the trough, the constrained version survives and compounds.
What would prove me wrong
A futurist who can't tell you how they'd be wrong is just a fortune teller. So here's the scenario that breaks everything I just wrote.
If capability scaling doesn't plateau — if the next generation of models crosses some threshold into reliable, self-correcting, genuinely autonomous reasoning, and especially if systems start meaningfully improving themselves — then the trough never completes. The curve breaks upward instead of leveling off, the economic math that looks insane today gets retroactively justified by a step-change in what the technology can do, and the cautious "AI becomes boring infrastructure" story I'm telling looks as quaint as people in 1995 saying the internet was a fad.
I don't think that's the base case for this decade. But I hold it loosely, and you should too. The honest position is that the economics are in a bubble while the technology may or may not be near its ceiling, and those are different questions that the discourse constantly mashes together.
The part that should make you optimistic
Here's why I opened by calling the collapse the best thing that could happen.
Hype is a tax on builders. When a field is at peak hype, attention and capital flow to the loudest demo, the boldest claim, the thing that sounds most like the future. Substance is a competitive disadvantage. The trough inverts that. The tourists leave. The capital gets disciplined. The grift stops paying. And suddenly the people who are left are the ones who actually wanted to build the thing — and the ground is littered with cheap infrastructure, mature tooling, and problems that everyone finally admits are real.
The trough of disillusionment isn't where technologies go to die. It's where they get serious. The plateau of productivity — the boring, durable, world-changing part — is on the other side of exactly this.
The hype is leaving. Don't mourn it. It was never on your side.

Sources and further reading, for anyone who wants the underlying numbers:
Gartner's 2025 Hype Cycle for AI (placement of GenAI in the trough and agents at the peak); the MIT/McKinsey figures on enterprise pilot and profitability failure rates; Goldman Sachs and Morgan Stanley analysis on hyperscaler capex versus profit; BloombergNEF on data-center buildout and neocloud financing; and the Belfer Center on AI's collision with the electric grid. Links: