Everyone uses AI to write better emails than they could write on their own.
That’s not a criticism. It’s just true. The technology has become, for millions of people, a quiet collaborator in the background of daily professional life — helping find the right words, clarifying a tangled thought, turning a frustrated draft into something that actually says what was meant. People who would never describe themselves as AI users are using AI. They just call it cleaning up this email real quick.
Which makes it all the more remarkable that the AI industry is so extraordinarily bad at explaining itself.
The technology that can take any complex idea and render it legible to any audience is being advocated for by an industry that communicates in parameter counts, benchmark comparisons, and capability demonstrations that mean nothing to the people whose lives it’s about to reshape most dramatically. The field that built the world’s most sophisticated communication tool communicates about itself like it’s writing terms and conditions.
The result is a public that has absorbed the hype without absorbing the substance. Everyone has an opinion about AI. Almost nobody has an accurate picture of what it is, where it actually is in its development, or what’s coming next.
That gap — between the noise and the reality — is what this essay is about.
* * *
The comparison that keeps surfacing — in financial analysis, in the more sober corners of tech journalism — is to the dot-com bubble. And it’s more precise than most people realize. Precise enough to be useful. Also incomplete in a way that matters enormously.
The dot-com era produced Pets.com, Webvan, eToys — companies that raised hundreds of millions of dollars on the premise that putting a storefront on the internet was itself a business model. It wasn’t. The companies burned through their funding, delivered no sustainable revenue, and collapsed spectacularly between 2000 and 2002. The NASDAQ lost seventy-seven percent of its value. Careers were ended. Fortunes were vaporized.
And yet. The fiber optic cable that was laid to connect all those failed websites stayed in the ground. The server infrastructure that was built to host all those dead storefronts stayed in the data centers. The engineering talent that was trained on building internet products didn’t disappear when the companies did. The crash destroyed the speculation. It didn’t destroy the infrastructure. And the next generation — Amazon, Google, Facebook, the entire architecture of the modern internet economy — built itself on that infrastructure at a fraction of the cost it took to build it in the first place.
The dot-com bubble wasn’t a story about the internet being a bad idea. It was a story about the gap between what infrastructure enables and what any particular generation of builders knows how to build on top of it.
AI is in that gap right now.
The current generation of AI products is mostly free or nearly free, sustained by investment capital rather than revenue, running on data center infrastructure that consumes electricity and water at a scale that is genuinely unsustainable. The valuations are based on capability demonstrations and projected futures rather than current economics. The companies burning billions to train and serve frontier models have revenue models that do not yet justify the infrastructure they’ve built.
This is Webvan. This is the part of the cycle where the speculation outpaces the revenue, where the infrastructure is being built faster than anyone knows how to monetize it, where a correction is coming that will feel catastrophic to the people caught in it.
The correction, when it comes, will not be evidence that AI was a bad idea. It will be the moment the infrastructure becomes available to build on.
The companies that emerge from the other side of an AI correction will have access to extraordinary compute at a fraction of current costs, a decade of model development to build on, and a much clearer understanding of which applications actually create enough value to sustain themselves economically. That’s not a disaster scenario. That’s how transformative infrastructure gets built.
* * *
The infrastructure wasn’t the only thing the bubble built. Some of what got built during the dot-com era was immediately useful. The internet of 2000 was already changing how people found information, communicated across distances, and accessed knowledge that had previously required a library and a day’s research. The speculation was insane. The capabilities were real.
The same is true now. Underneath the hype and the unsustainable economics and the chatbots helping people write emails, something is happening that deserves to be stated clearly and specifically.
Start with medicine. Not the vague promise of future cures but the specific, documented, already-happened breakthroughs.
In 2020 DeepMind’s AlphaFold solved a problem that had resisted biology for fifty years. Proteins are the machinery of life — they do essentially everything inside a cell — and their function depends entirely on their three-dimensional shape. Figuring out that shape from the amino acid sequence that codes for a protein had taken teams of researchers a year or more of expensive experimental work per protein. AlphaFold predicted the structure of essentially every known protein — approximately two hundred million of them — with near-experimental accuracy. In 2024 the work was awarded the Nobel Prize in Chemistry. The protein structure database is now freely available to researchers worldwide. The downstream impact on drug discovery, on understanding disease mechanisms, on designing therapeutics for conditions from cancer to Alzheimer’s, is still compounding.
This is not a chatbot helping write emails. This is a fifty-year scientific problem solved in a way that restructured an entire field of biology.
AI-designed drug candidates are now in human clinical trials. The first AI-designed molecule to enter human testing did so for obsessive-compulsive disorder, developed in less than twelve months — a process that typically takes four to six years.
Insilico Medicine identified a novel drug target for idiopathic pulmonary fibrosis and advanced a candidate to clinical trials in eighteen months at a cost of roughly one hundred and fifty thousand dollars. The same process through conventional methods typically runs to hundreds of millions of dollars and takes years. AI-powered diagnostic systems are outperforming human specialists in specific imaging tasks — detecting diabetic retinopathy, identifying cancerous lesions in radiology scans, flagging cardiac abnormalities in electrocardiograms.
Now go outside medicine.
Weather forecasting has been transformed in ways most people haven’t noticed because better weather forecasts don’t generate headlines the way chatbot controversies do. The European Centre for Medium-Range Weather Forecasts deployed an AI forecasting system in early 2025 that runs at one thousandth the computational energy of traditional physics-based models while delivering comparable or superior accuracy. Google’s NeuralGCM model sent longer-range monsoon forecasts to thirty-eight million farmers in India, helping them make planting decisions that directly affect food security. Microsoft’s Aurora model, trained on over a million hours of atmospheric data, is being used to forecast flood risk, wildfire spread, and seasonal weather patterns in ways that shift disaster response from reactive to proactive.
Climate and energy. AI systems are optimizing power grid management in ways that meaningfully reduce waste. Renewable energy forecasting — predicting when solar and wind generation will peak and trough — is becoming precise enough to make grid integration of intermittent sources significantly more reliable. Direct air carbon capture processes are being optimized by AI to bring costs down toward economic viability.
Agriculture. AI-powered crop monitoring, yield prediction, and precision farming are reducing water and fertilizer use while improving output. The gap between what advanced agricultural science knows and what farmers in developing countries can access is being compressed by AI systems that can run on basic hardware and translate complex recommendations into actionable guidance in local languages and contexts.
The technology that’s getting mocked for helping people write emails is also helping farmers in India decide when to plant and helping researchers design drugs that couldn’t have been designed without it.
The hype is real. The underlying capability is also real. Holding both without collapsing one into the other is what accurate thinking about this moment requires.
* * *
The thing connecting the dotcom parallel to the real applications to what comes next is energy.
Training a single frontier model requires thousands of megawatt-hours of electricity. Running that model at scale for millions of users requires more. The water used to cool the data centers running this infrastructure has become a crisis in specific communities — towns where the local water supply has been materially affected by data center construction and operation. This is not an abstract environmental concern. It is a concrete, measurable, current problem affecting real places and real people.
The AI industry knows this. You cannot build the AI 2.0 that comes after the current bubble on infrastructure that consumes resources at this rate. The selection pressure on the next generation of AI development is energy efficiency. Not capability for its own sake. Not parameter count. Energy per unit of useful output.
This is the forcing function that makes the previous essay in this series — the one about neurons in dishes in Switzerland and Australia — not a strange tangent but the logical destination of the argument. If biological neurons are orders of magnitude more energy efficient than silicon transistors for the kind of computation that intelligence requires, then the energy constraint doesn’t just favor wetware as an interesting research direction. It selects for it.
The dot-com bubble built fiber optic cable that the next generation ran on. The current AI bubble is building something similar — not just the model architectures and the training techniques and the engineering talent, but the pressure toward efficiency that will determine what the next substrate looks like.
The ashes this generation builds on won’t just be the failed companies and the vaporized valuations. They’ll be the infrastructure, the knowledge, and the unsolved energy problem that forces the next generation toward something fundamentally different.
* * *
The AI industry is bad at explaining itself. That’s where this essay started and it’s worth ending there.
The decisions that shape how this technology develops — regulatory frameworks, research priorities, resource allocation, ethical guardrails — are being made in an environment where most of the people making them, and most of the people affected by them, have a picture of AI that is simultaneously too dismissive and too credulous. Too dismissive of the real capabilities. Too credulous about the hype.
The dotcom parallel cuts through both failure modes. Yes, there is a bubble. Yes, there is waste and unsustainable economics and hype that has outrun reality. And yes, something real is being built underneath all of it, something that is already changing medicine and science and agriculture and weather forecasting in ways that matter to real people.
The bubble will correct. The infrastructure will remain. The next generation will build on the ashes of the current one.
What that next generation builds — and what it runs on — is what the rest of this series is about.