A Third Door: Thoughts on the AI Era

Nobody knows what is going to happen with AI.

I want to say that first, because almost everyone talking about it sounds certain. I am not. There are a few scenarios for how this plays out, and I can't rule any of them out. But I do have a bet, so let me put it up front. It is neither of the two doors everyone argues about: that AI is a fad, or that it ends human work. It is a third one.

I think AI is the next big thing. I also think we are in a bubble. I think the bubble pops, a lot of people lose a lot of money, and what is left behind is a mountain of cheap compute: dark GPUs and dark data centers, the same way the dot-com crash left behind dark fiber. And I think the people who win are the ones who build on top of it afterward. Not the end of the software engineer. The largest rise of application and product engineers the world has ever seen.

Last year I wrote that AI is a rocket ship for the mind. I still believe that. What I don't believe is the version of the story where it ends everything.

Adoption Without Profit

Start with the money, because that is where the story gets shaky.

Individual use is everywhere. Gallup found that 52% of US workers now use AI in their role. But the Census Bureau, which asks businesses whether they use AI in their operations, puts it at about 24%. That is up from about 17% last December, but still less than half the share of workers.

And the returns are thin. In PwC's survey of more than 4,000 CEOs, 56% said AI had neither lowered their costs nor raised their revenue. Only 12% got both. OpenAI reportedly ran an operating loss of $12.3 billion in the second quarter of this year on $6.7 billion in revenue. In the third quarter, revenue reportedly jumped to about $10 billion. The loss, on an adjusted basis, was $10.9 billion. Still larger than the revenue.

Profitability hasn't arrived for most of the people who have adopted it. If the tools are this good, why isn't the money showing up?

The Bubble

Michael Burry looks at all this and sees the dot-com era, with Nvidia in the role of Cisco: the company selling equipment to everyone in the race. His description of what went wrong back then is the one I can't shake: "catastrophically overbuilt supply and nowhere near enough demand." He also argues the big cloud companies are flattering their earnings by pretending chips last longer than they do, to the tune of about $176 billion in understated depreciation between 2026 and 2028.

The money moves in a circle. Nvidia invests in the companies that buy its chips. At the end of the 1990s, Lucent, Nortel and Cisco did a version of this, lending billions to customers so they could buy equipment. Nvidia says its situation is different, and says its customers pay on time. But the shape is familiar.

Meanwhile the largest tech companies are spending nearly all the cash they generate. Capital spending took 80% of operating cash flow at Alphabet, Amazon and Microsoft over the past year, up from 42% in 2023. Across the five biggest spenders, new debt now equals about a third of capital spending, up from 9% two years ago. Some writers already compare the data center debt to 2008. I wouldn't go that far. But real estate looked safe too, right up until it didn't.

This isn't only a short seller's view. Sam Altman has said investors are overexcited about AI, and Jeff Bezos has called it an industrial bubble. The people building it agree on the word. They just think they will be the ones left standing.

Dark Fiber, Dark GPUs

Here is the part I find exciting.

During the dot-com boom, telecom companies buried fiber optic cable everywhere. Far more than anyone needed. In 2000, less than 7% of it was lit. The rest was "dark fiber," sitting in the ground unused. Many of the companies that laid it went bankrupt.

The cable stayed. Kai Wu at Sparkline Capital summarizes what happened next: with 85% of the cable unused, the cost of bandwidth fell by 90%, and that "ultimately helped fuel the rise of firms like Netflix and Facebook." Tim O'Reilly points out that YouTube was built on "cheap bandwidth that someone else had paid for." The investors lost. Everyone who came after them won.

So what if this is the era of the dark GPU? The dark data center? David Sacks used the same phrase. He said a glut of "dark GPUs" would be a disaster for everyone, though he doesn't expect one. A disaster for the people who paid for them, sure. For everyone else, there is a very real chance that inference and compute become dirt cheap, the same way the internet did. It is already heading that way. The price of getting the same quality of answer from a model has been falling about 10x a year.

The Deep Caveat

There are holes in this analogy, and I don't want to skip them.

Fiber lasts for decades. GPUs don't. One estimate, from a single unnamed Google architect, puts the life of a heavily used data center GPU at one to three years. Nvidia now releases a new generation every year, and Jensen Huang joked that once the new chips ship, you couldn't give the old ones away. A warehouse of three-year-old chips is not the same gift as a cable in the ground.

Fiber also costs nothing to leave in the ground. A GPU needs electricity and cooling every hour it runs. Compute will never get as close to free as bandwidth did. The floor is the power bill.

And there is no glut yet. Rental prices for older chips went up this year, and in August CoreWeave said it had just signed a contract running into 2029 for A100s, a chip design from 2020. That looks like it contradicts the weak demand I just described. I don't think it does. The chips are busy because capital is paying for them: labs spending investor money, cloud companies spending borrowed money. The customers at the end of the chain mostly aren't earning a return yet. Demand funded by investors lasts as long as the investors do. If those customers start making real money before the funding runs out, there is no bust, and I am wrong.

So if the bust comes, where does the cheap compute come from? Not from old chips. The idle hardware would be the newest hardware, bought at the peak by companies that can no longer pay for it. Their lenders would rather rent it out for anything above the cost of power than let it sit. That is distressed capacity priced near what it costs to run, inside buildings that are already paid for. Those buildings, with their power connections and cooling, are the part that lasts: Microsoft now assumes its data center buildings last 25 years, up from 15. As one analysis put it, "In AI, the slow-wearing part is the powered shell: the land, the grid connection, the cooling and the building."

The software side survives too. Part of that 10x a year price drop comes from better models and better engineering, and nobody can repossess that. Smaller open models running on hardware someone else already wrote off would do most of what people pay a premium for today.

The Application Engineer

Now the bet itself. Just like after the dot-com crash, the software engineer, or whatever they call the job in the future, is not going anywhere. After a long period of turmoil, application and product engineers will build on top of the cheap layer.

What would they build? Think about YouTube again. Video on the internet was possible in 1999. It was slow for most people to watch, and too expensive to give away. Cheap bandwidth didn't invent anything. It changed what was worth doing. My guesses for cheap inference are the same kind of thing.

One is software that checks everything instead of sampling. Today a program calls a model when a user asks for something, because every call costs money. When a call costs almost nothing, a program can read every contract, every transaction and every support ticket, all the time, in the background.

Another is software for small audiences. There are thousands of communities and niches too small to justify a product team. When building and running a product gets cheap enough, two people can serve them. Altman has talked about a betting pool for the first one-person billion-dollar company. I think the more common outcome is thousands of small companies that could never have existed before.

The third is buying reliability with compute, and this is where the engineer comes in. METR measures how long a task AI agents can complete. In their May report, the best public models could finish tasks that take a human about 12 hours, at a 50% success rate. Ask for 80% reliability and that drops to about an hour and a half. Anthropic's own usage data shows businesses using Claude through the API saw tasks succeed about half the time, with the rate falling as tasks got longer. A coin flip on a 12-hour task is impressive. It is not something you hand your business to. Cheap compute helps: you can run the task ten times and check the answers against each other, though agreement is not proof, since models tend to repeat the same mistakes. But someone has to design that system, decide what counts as right, and own the result. That is the job. Steering inference to make realistic decisions and solve actual problems.

Where Scarcity Shows Up

The turmoil comes first, and it has started. Indeed's data shows software job postings are still about 19% below where they were before the pandemic, though they have been rising since early 2025, and employment for developers aged 22 to 25 is down nearly 20% since 2024. Junior engineers are being squeezed right now. Students have noticed: new computer science majors at research universities fell 13% in the latest count, the first clear sign of contraction.

That squeeze is what makes the bet pay off. Fewer juniors hired today and fewer students enrolling means fewer experienced engineers a few years from now. If the dark GPU era arrives around then, demand goes up just as supply has thinned out. Everyone keeps asking where scarcity will exist in an AI world. It shows up here first: in the people who know how to build with it.

The obvious objection is that AI fills the gap. If the models can do a junior's work today, they will do a senior's work by then. But the reliability numbers say the hard part is the long task, which is exactly where judgment and accountability live. The market seems to agree: in Indeed's July analysis, senior roles made up 71% of the recent rise in software postings. If the models become reliable on week-long work with nobody watching, I'm wrong about this. Until then, experience is the scarce thing, and we are producing less of it.

My own experience points the same way. I've spent the last year and a half building applications with these tools. They made me faster. They did not make it easy. It still took an incredible amount of effort, and I haven't seen AI replicate what I built on its own.

Who Is Actually Being Automated?

A while ago I was at a networking event, telling people about Finjani, an app I built to connect Arabs across the Western world. One of the people I talked to was an investor. I wasn't pitching him. He is Arab himself, so I was trying to get him to sign up.

He treated it as a pitch anyway. Every answer got the same response: "What else?" His one real criticism was that retention would be bad. When I told him my retention and growth numbers, he seemed taken aback. He also seemed to assume I hadn't already run the idea through four different models, at least a hundred times, and then done my own thinking on top of the answers.

Look at what each of us brought to that conversation. His objection was the generic one. I brought the product, the numbers, and the judgment about which of those hundred answers were worth keeping. A lot of investors seem to think the AI era made the engineer irrelevant. Have they considered that generic judgment is the thing being automated, and that it might be theirs?

There is a warning in that for everyone, me included. The danger isn't that nobody will be thinking anymore. It is that everyone believes they are thinking independently while leaning on the same few models, and so on the same underlying assumptions. If that happens, the people who don't blindly trust the machine become considerably more valuable.

The Endgame Scenario

That is my bet. The people running the industry are making a very different one. Listen to them and you would think we are at the finish line. Knowledge work is about to be fully automated. The singularity is around the corner. There will be nothing left to do.

Altman wrote last year that "we are past the event horizon; the takeoff has started." Dario Amodei wrote in 2024 that the equivalent of a whole country of geniuses, running in a data center, could arrive as early as 2026, though he said it could also take much longer. This January he put it around 2027, and wrote that half of entry-level white-collar jobs could be displaced within one to five years.

Bill Gates is the one that stays with me. On The Tonight Show, Jimmy Fallon asked him whether we will still need humans. His answer: "Not for most things." As Harvard Magazine put it after one of his talks, the first PCs amplified human capabilities, while AI has the potential to replace them.

If It's Unprecedented, Why So Confident?

Here is what I can't get past. These people describe what is happening as unprecedented. And if it is unprecedented, how are they so confident about how it ends?

Forecasts work when you have something to compare against. Arvind Narayanan and Sayash Kapoor made this point about predictions of AI catastrophe: there is "no reference class," so all a forecaster can do is "make up a number." The same logic applies to the confident economic predictions. You can't call something a first in human history and also tell me the year it happens.

That cuts against me too. My dot-com analogy is also a forecast. The difference is what is being forecast. The end of work has never happened. Investors overbuilding infrastructure for a new technology has happened again and again: railways, then fiber. The technology is new. The financing is not. And I'm not giving you a year.

The track record of the confident predictions doesn't help them. In 2016 Geoffrey Hinton said people should stop training radiologists. By last year the Mayo Clinic's radiology staff had grown 55%. In 2019 Elon Musk promised "over a million robotaxis on the road" by the next year. Even Altman admitted this summer that "we've all been too ambitious on timelines."

There is also the obvious part. These are people raising enormous amounts of money, and short timelines are good for fundraising. When an industry goes mainstream, a lot of people end up needing the hype to be true.

The skeptics have been wrong too. Coding really did change, faster than most people expected, me included. Forecasters have consistently underestimated how quickly these models improve on benchmarks. My doubt isn't about whether the technology is getting better. It is about the jump from "the models are getting better" to "the economy as we know it is over."

Linus Torvalds said it in a way I keep coming back to. He thinks AI is really interesting, and he also called the industry around it "90% marketing and 10% reality." That is where I sit. A very big deal, as a tool.

What If They're Right?

Still, take the endgame seriously for a moment. Suppose everyone has access to real intelligence and the jobs go away.

Then who buys the products? There is an old story from a Ford plant in the 1950s. A company official, showing off new automated machines, asked the union leader Walter Reuther how he would collect union dues from them. Reuther's reply, as the earliest account has it: how are you going to get them to buy Fords?

That is why the endgame, as described, undercuts itself. The companies spending hundreds of billions on this need customers with paychecks to earn it back. If the jobs really vanish, so does the revenue that justifies the spending. The people at the top know this, which is why the same people predicting mass job loss are also proposing things like Elon Musk's "universal high income" or Amodei's idea of taxing AI revenue. The world may well enter a different era, where relationships and taste are what stay valuable. But a change that large gets negotiated, slowly. It doesn't arrive on a product roadmap.

And so far the data doesn't show it. Stanford's researchers, who track this closely, wrote in August that they "do not see widespread, economy-wide job displacement associated with AI," even as hiring of young workers in exposed jobs gets squeezed.

Now suppose the best version happens. AI, robotics and abundant energy largely eliminate disease, poverty, hunger and involuntary labor. Even then, I don't think there is nothing left to do. For the first time, humanity wouldn't have to spend most of its effort surviving, and the questions change. Can we build self-sustaining cities on Mars, or reach another star? Can we extend healthy human life, or understand consciousness? What is dark matter? We have barely explored our own oceans, and solving a problem often creates questions nobody knew existed.

Solving material scarcity also doesn't solve the human condition. In a world of abundance, people still have to work out relationships, identity, meaning, and how they want to live. So maybe intelligence isn't the endgame. Maybe it is what finally lets us start exploring. That is the tool argument again. The future I find exciting is not one where machines take over everything. It is one where machines make previously impossible things achievable for people.

I Don't Know

The honest answer is that I just don't know.

But something is telling me this isn't the end of everything. A large part of it is hype. The money has gotten ahead of the technology, and a lot of people are going to lose a lot of it. What is left afterward will be a very effective tool, cheap enough that a new generation of builders can do things with it nobody is predicting today.

That is not a small outcome. People thought the internet was a fad once. Then the bubble burst, and they thought they had been proven right. The internet turned out to be bigger than even the hype said, just later, and for different people than the ones who paid for it.