Everybody has an opinion on AI

Everyone has a take on whether AI is a bubble, and very few have read enough to say which part of it is.

Everyone has an opinion on AI. Not just the engineers and the investors. Everyone. Your uncle who just discovered ChatGPT. The startup CEO who swears his devs are literally 10x more productive. The PhD who calls it a gimmick if not a scam. The guy on Twitter calling it one Ponzi scheme after another while his portfolio is up 40% on NVIDIA. Everyone has a position. Very few have done the reading.

The bubble that worries me most is the one in AI opinions. We have created a secondary market of hot takes that trades at higher volume and with less underlying value than anything on the NASDAQ.

There are days when the term AI makes me physically recoil, especially when I'm elbow-deep building a markdown knowledge base with hierarchical folder resolution and a live reference-detection layer that flags wiki-links on the fly, and my girlfriend walks in from her fifth meeting of the day where they dropped AI more times than Tim Cook unveiling the latest iteration of Apple's Thoughtfully Reimagined Personal Intelligence Experience. At least that confirms I'm not some recluse obsessing over antique spoons all day, convinced that antique spoons are the pinnacle of human achievement, blogging about antique spoons to an audience of seven other antique spoon enthusiasts. People are talking about this. A lot. Most of them are wrong, but they're talking.

Balloonomania, 1783

The history of the first balloons is absurd enough. The phenomenon was called Balloonomania. After the Montgolfier brothers launched the first public balloon in June 1783, a wave of speculative hysteria swept across Europe. A crowd reported at 400,000 people gathered in Paris to watch Jacques Charles take off from the Tuileries that December. Balloon-themed fashion, furniture, and merchandise flooded the market. The Montgolfiers believed they'd discovered a new gas (they called it Montgolfier gas) that was making their balloons rise1. They hadn't. It was just hot air. Literally just heated atmosphere. They didn't even understand the basic physics of the thing they'd invented, and investors were already throwing money at them.

1Montgolfier brothers, Wikipedia, which also gives the June 1783 date of the first public demonstration at Annonay.

The template.

The tech was real: hot air balloons flew. The investment was real, and enormous sums changed hands. What failed was the assumption that because it works at all, it will work at any scale you want to fund. The same template has run for 240 years.

The railways of the 1840s crashed brutally and then carried a century of transportation. Radio crashed after the stock market frenzy of the 1920s and became mass media. The dot-com implosion of 2000 was followed by the digital economy that runs the world today. Each time the technology was real.

The template is ancient. What's changed is the speed at which we run through it, and the dollar amounts involved. The Montgolfiers' bubble was measured in thousands of livres. In the railway mania of the 1840s, actual railway investment peaked at about £44 million in 1847, some 7% of Britain's GDP in a single year1. The dot-com crash wiped out $5 trillion in market value. The AI money is larger again: in 2025, AI-related startups captured 50% of all global venture capital. Two hundred billion dollars poured into a single sector2. Half of all the venture money on earth, betting on one thing.

1Andrew Odlyzko, Collective hallucinations and inefficient markets: The British Railway Mania of the 1840s, University of Minnesota, 2010, table 1 (GDP of £604 million that year). 2Gené Teare, 6 charts that show the big AI funding trends of 2025, Crunchbase News, December 2025, which counts $202.3 billion and puts foundation models alone at $80 billion, up from $31 billion the year before (a 158% increase in one year).

The Montgolfiers at least had an excuse. They didn't know it was just hot air.

Everyone is right and nobody is listening

The debate splits into three camps, and each camp has the exact same problem: they can see their corner perfectly and assume it's the entire building.

Camp one: it's a scam

The loudest voices are variations on a theme. AI is one Ponzi scheme after another, a gimmick if not a scam. One commenter said the quiet part out loud, that investors realized that they could get far more return from hype than from any kind of legitimate business.

These people are looking at the stock market and they're not wrong about what they see. Companies with no revenue getting billion-dollar valuations. A quantum computing company raising $600 million and doubling its valuation to $10 billion, on the strength of technology that is, to put it diplomatically, not yet commercially useful. From a pure financial lens, parts of this market are detached from reality in ways that should make any sane person nervous.

But calling AI itself a scam because the stock market is overvalued is like calling electricity a scam because Enron existed.

Camp two: you don't understand the tech

The counter-camp is equally forceful. A startup CEO reports his developers going from 300 to 2,500 lines of code per day. A software engineer says AI makes possible what used to be impossible. Another snaps that AI is helping me make a lot more money than before, so I'm not sure where the cynicism is coming from. Someone else says they'd pay $2,000 a month if they had to.

These people are looking at their own experience and they're not wrong either. The tools work. I know because I use them. Every day. For this exact kind of work, in fact.

But it works for me is a personal argument, and the question here is macroeconomic. Plenty of things work for individual users while being economically unsustainable at the industry level. The product and the business model are different things, and this camp rarely engages with the difference.

Camp three: bubble, schmubble

The third group (smaller, and in my experience closest to right) shrugged and said: of course it's a bubble. So what?

I have no doubt the AI bubble will burst, probably soon. But it's the same scenario as when the PC and the internet stepped into the limelight. Bubble, pop, assimilation, productivity.

The dot-com bubble burst in 2000 and killed Pets.com, Webvan, and a thousand startups whose entire business plan was the internet, but for dogs. The internet came through it, and so did Amazon, Google and eBay. The crash judged the business models and left the technology standing.

This take generates no engagement and won't make you sound smart at dinner parties; it lacks the punch of it's all a scam and the tribal energy of you just don't get it. It only has the unfortunate quality of being historically accurate every single time.

My portfolio says one thing, my inbox says another

My situation is, I suspect, the situation of a lot of engineers right now. My positions in tech companies with heavy AI exposure are up over 150%. My daily toolkit includes AI-assisted coding, research, and writing. My productivity, by any honest measure, is meaningfully higher than it was two years ago. When I write code, I write it faster. When I research a topic, I reach the core faster. When I debug something, I find the bug faster. The tools are measurably useful. I wrote about why the underlying architecture works. It is real math doing real things.

My LinkedIn inbox, meanwhile, has gone quiet, and the reason is a contracting market. Entry-level tech postings in the US have dropped by an estimated 67%. In the UK, tech graduate roles fell 46% in 2024 with projections for another 53% drop by 20261. Senior roles are flat. The overall tech job market is in what analysts politely call a low-hire, low-fire equilibrium, which is corporate-speak for nobody's getting fired and nobody's getting hired, and if you're a fresh graduate, best of luck.

1Both figures as gathered in Rezi's report on entry-level jobs and AI, 2026, which takes them from an IntuitionLabs analysis of graduate hiring; the 67% compares 2024 with 2023 and comes from some analyses, not an official count.

The duality.

I am literally more productive than I've ever been, using tools that are eliminating the entry-level positions that people like me once held to become productive in the first place. My stocks go up because AI makes senior engineers more efficient. The junior hiring pipeline dries up for the exact same reason. The two facts are one fact, seen from different floors of the building.

And this is the part that neither the it's a scam camp nor the you don't get it camp wants to sit with. The technology works. The economics are brutal. The human cost is real and unevenly distributed. All three things are true simultaneously, and collapsing any of them into a slogan (AI is overhyped or AI is transformative) is the kind of intellectual shortcut that makes me tired.

I'm tired of LinkedIn posts from people who have never trained a model explaining that AI will replace all software engineers by Thursday, and of posts from people who have never shipped production code explaining that AI can't write production code. Tired of YouTube thumbnails with red arrows pointing at ChatGPT logos. The comment that stayed with me was quieter than the rest: someone worried that, burst or no burst, a lot of people would lose their jobs before this settles, and a lot of creative people would lose heart.

Half of all venture capital

In 2025 AI captured half of all venture capital deployed worldwide, $202 billion, up 75% from the year before. Foundation models alone (the raw engines, the stuff companies like OpenAI and Anthropic build) absorbed $80 billion, more than doubling from $31 billion in 2024.

When I say bubble, I'm making a mathematical observation about what happens when half the venture money on earth bets on one thing, and leaving aside whether AI is useful.

The technology is real; I use it and it works. A commenter who builds AI data centers confirmed that the hardware advances are genuine. Another pointed out that banks are already using machine learning for investment decisions in sub-second timeframes. One developer said the tools turned non-coders into coders and coders into code reviewers. Fair enough. All real.

The problem is that the people writing the checks don't understand what they're buying.

A CEO who can't explain what a transformer is allocates 30% of the IT budget to AI transformation1. A venture capitalist who has never seen a loss function funds a startup because the pitch deck says AI-native. A pension fund manager who doesn't know the difference between training and inference puts retirees' money into NVIDIA stock because it went up last quarter.

1The transformer is the neural network architecture behind every major AI system (GPT, Claude, Gemini, LLaMA); it processes text by computing attention scores between all words at once instead of reading them in sequence, and while the math fits on one line, the emergent behavior doesn't fit in anyone's head.

These people are making emotional bets on narrative far more than informed bets on technology, and narratives always correct.

What the correction looks like.

When the narrative corrects, the people who lose money will mostly be the investors who put billions into companies whose entire value proposition was AI as a prefix, the ones who paid for the balloon, while the engineers building useful AI tools in quiet companies with actual revenue keep working. The tech will survive and the hype layer won't, and the losses fall in the gap between them.

Someone in finance described the mechanism: vast and growing national debt means more and more public money is constantly funneled to the finance sector, which is where speculation lives, and they will put that money into something. The money doesn't care what the something is, as long as it sounds expensive enough to be plausible. It just needs a vehicle. AI is this decade's vehicle.

There's also a bubble inside the bubble that almost nobody is watching. Below the headline AI companies (the OpenAIs, the Anthropics, the model builders), there's an entire ecosystem of infrastructure startups: GPU cloud providers, vector database companies, MLOps platforms, AI deployment tools. The picks and shovels of the gold rush. My bet, which the picks-and-shovels crowd won't like, is that most of them are solving problems that the major cloud platforms will absorb in 18 months.

The shovel problem.

It is already happening. Weights & Biases, once the flagship MLOps platform, got acquired by CoreWeave in 2025, for about $1.7 billion1. The standalone vector database market (Pinecone, Weaviate, Milvus) watched their value proposition evaporate as PostgreSQL gained vector search through the pgvector extension and the database giants made their move. Snowflake acquired Crunchy Data for about $250 million. Databricks spent about $1 billion on Neon. The purpose-built vector DB became a feature of the general-purpose database. 2025 was a busy year for data infrastructure M&A. The consolidation is eating the picks and shovels layer alive, and most of the companies being digested haven't updated their pitch decks yet.

1The GPU cloud provider bought them outright, which turned the independent MLOps play into an infrastructure play, itself a bet on continued GPU demand.

The infrastructure bubble looks more fragile than the model bubble, because at least the models do something novel. The infrastructure companies are building things that bigger companies can replicate with a team and a quarter.

The dot-com pattern

Everyone in that comment section cited the dot-com bubble. Not one of them followed the analogy all the way through.

The NASDAQ lost 78% of its value between March 2000 and October 2002. Trillions evaporated. Companies with real revenue (Amazon, eBay, Priceline) lost most of their stock price anyway, because when panic hits, it doesn't check your balance sheet. Amazon went from $107 to under $6. Six dollars. A company with real sales and millions of customers, a few months from its first quarterly profit, trading at the price of a sandwich.

And then, quietly, without a single thought leader noticing, the actual revolution happened.

Cloud computing. SaaS. Social media. E-commerce at scale. The smartphone era. The entire digital economy that currently runs civilization was built after the bubble rather than during it. The bubble was the awkward, overfunded, over-hyped first date. The revolution was the two decades of compound utility that followed.

After the crash.

What the dot-com crash proved is that most early internet business models were fake, while the internet itself came through. The technology survived because it was useful and the speculation died because it was speculative. Confusing the two is the central error of the AI is a bubble crowd, and, paradoxically, also of the AI is not a bubble crowd. The bubble is real and so is the tech, and failing to hold both in your head at once is the bubble in opinions I started with.

If the pattern holds (and it has held for 240 years, since the balloons), what comes after the AI correction will be more consequential than anything happening now. The useful tools will survive. The grift layer will evaporate. The CEOs who allocated budget to AI transformation without understanding what they bought will quietly un-allocate it. And the engineers who built things that actually work will keep building, same as they were before anyone on LinkedIn had an opinion.

Someone wrote something I keep coming back to, that those who survive will be far more valuable and impactful than any company in history. That's either naive optimism or the most accurate prediction in the thread, and I don't know which.

And there's a wildcard that nobody at the investment conferences wants to discuss, open source. Optimized models running on consumer hardware need no data center and send nobody an API bill. A commenter predicted that open source optimized models for consumer hardware will kill big companies in 4-5 years. Maybe. Maybe not. But the fact that a viable open-source path exists means that the entire centralized infrastructure play (the hundred-billion-dollar bet on GPU clouds and proprietary APIs) has a timer on it. The picks and shovels become irrelevant when people start finding gold in their backyard.

Quantum is warming up

Speaking of the next thing, several commenters noticed that the hype machine is already warming up a replacement in quantum computing, on the theory that once the AI bubble pops, investors will need somewhere to park their optimism, and quantum computing is exotic enough to absorb it.

One prediction deserves a plaque: my guess is quantum computing investment will pick up as soon as they convince us they will deliver flying or self-driving vehicles.

A researcher who actually works with quantum pointed out something more substantive: quantum computers are good at specific algorithmic problems that involve very little data. Loading data into a quantum system is slow. The qubit counts are tiny. And the idea that quantum computing will solve AI's scaling problem is roughly as coherent as the idea that buying a faster blender will fix your diet. The tools do different things. But coherence has never been a requirement for investment theses.

Every time someone says there's a limit, they are proved invariably wrong because they always confuse the current trajectory with a physical law.

Maybe. Or maybe some limits are physical laws, and the people who point them out are doing physics. The history of technology is full of both: limits that turned out to be temporary engineering problems, and limits that turned out to be thermodynamics. The trouble is that investors can't tell the difference. And they don't need to, because by the time the answer is clear, they've already moved their money to the next balloon.

The it's a bubble people are right about the money. The it works people are right about the technology. The it'll be fine people are right about the pattern. None of them are talking to each other.

The disagreement reaches the top of the field, too. Yann LeCun, one of the three people who won the 2018 Turing Award for the breakthroughs behind deep learning, has spent the last two years publicly arguing that autoregressive LLMs are a dead end and that we need an entirely different architecture. Geoffrey Hinton, another of those three Turing Award winners, believes the models already understand more than we think, and that the real problem is control, building something we may not be able to steer. These two built the field. They mentored the people who run it. And they can't agree on whether the thing they created works.

If the founders of deep learning can't settle this over dinner, maybe nobody on LinkedIn is going to crack it in a carousel post.

The bubble will pop. Some companies will die. Some investors will lose badly. The financial correction will be painful and someone on LinkedIn will call it unexpected. The junior engineers who got squeezed out during the frenzy are the cost that nobody discusses at the conferences, and the one I'd most like the next cycle to get right.

And then, because this is how it always goes, the people who work with the technology will keep building. The useful tools will keep working. The hype layer will burn off like morning fog, and underneath it will be something transformative, built by people who kept shipping through the argument about whether it was a bubble.

My stocks might be down 40% by then. My tools will still work. And the next hot take will be about quantum.