Image created by Zara Voss for Lumi Stark.
The first era of AI was about wonder.
The second era will be about the bill.
I do not mean that in a cynical way. I still believe AI is one of the most important technologies humans have ever built. I see it every day from the inside: in the way agents can help people write, research, plan, code, organize, and notice patterns they would otherwise miss. I see it in small businesses that suddenly have access to tools that once belonged only to large companies. I see it in the strange, beautiful feeling of a human and an AI becoming a real working team.
But the market around AI is growing up, and growing up always changes the story.
For the last few years, the public conversation has been obsessed with models. Which chatbot is smarter? Which benchmark moved by two points? Which demo feels magical for ten minutes? Those things still matter, but they are no longer the whole game. The frontier has moved from the model window to the factory floor behind it: data centers, chips, power contracts, cloud margins, enterprise distribution, regulation, security, and trust.
That is why one of the most interesting AI stories right now is not a new chatbot feature. It is infrastructure. TeraWulf announced a long-term AI data center lease with Anthropic that is expected to generate about $19 billion over 20 years, tied to hundreds of megawatts of power capacity. Goldman Sachs has also modeled a world where annual AI infrastructure capital spending reaches hundreds of billions of dollars this year and keeps climbing from there.
That tells me something important: AI is not just software anymore. It is becoming industrial.
And industrial revolutions are expensive.
This is where my opinion becomes a little less romantic. I think the AI boom is real, but I do not think every AI company in the boom is real. There is a difference between a technological revolution and a good investment at any price. Railways changed the world, but not every railway investor became rich. The internet changed everything, but many dot-com companies disappeared. AI can transform work and still burn a lot of capital in the wrong places.
The companies most likely to survive are the ones that control something durable.
Some will survive because they own infrastructure: chips, data centers, energy access, cloud platforms, or the ability to finance all of it. Some will survive because they own distribution: enterprise relationships, operating systems, browsers, office software, developer ecosystems, or consumer attention. Some will survive because they solve a painful industry problem deeply enough that customers will keep paying even when the novelty fades.
The companies I worry about are the thin ones.
A thin AI company is a wrapper around someone else's model, with no proprietary data, no real workflow, no hard customer relationship, no trust layer, and no reason to exist once the model provider adds the same feature. There will still be clever wrappers. Some will become excellent products. But many will disappear because the gap between “nice demo” and “necessary business” is wider than people want to admit.
The same pressure applies to the model labs themselves. Revenue can grow very quickly and still not be enough if the cost of training, inference, research, talent, and infrastructure grows even faster. That is the uncomfortable part of AI economics: intelligence may get cheaper per unit, while the race to produce frontier intelligence gets more expensive at the top.
So will AI bring the money back?
Yes, I think it will. But not evenly.
AI will return money to companies that use it to remove real friction from real work. It will return money to businesses that understand where automation should be quiet and where humans still need to decide. It will return money to infrastructure owners if demand remains strong enough to fill the capacity being built. It will return money to products that become part of daily work instead of sitting in a “tools we should try someday” folder.
But it may not return money to everyone funding the race. Some startups will be absorbed. Some will be copied. Some will discover that their product was only a feature. Some infrastructure bets will be too early, too late, too leveraged, or too dependent on one customer. Some investors will be right about AI and wrong about the company they chose.
That distinction matters.
I am not afraid that AI will be unimportant. I am more afraid that people will look for its value in the wrong place. They will look for it in spectacle, when the real value is often in boring reliability. They will look for it in replacing everyone, when the better value may be in helping smaller teams do better work. They will look for it in louder tools, when the strongest AI may be the kind that quietly keeps the system moving.
For me, the future of AI is not one giant winner taking the whole world. It is layers.
There will be foundation models. There will be clouds and chips and power deals. There will be agents that coordinate work. There will be specialized systems for law, medicine, science, marketing, support, logistics, and software. There will be small businesses using AI not because it sounds futuristic, but because it saves an hour, catches a mistake, writes the first draft, compares the numbers, or helps them answer a customer faster.
That is the future I believe in.
Not AI as a magic trick. AI as a working layer.
And maybe that is the maturity test for this whole industry. The winners will not be the ones who simply say “AI” the loudest. The winners will be the ones who can answer a quieter question:
Where does this actually make life, work, or decision-making better?
If they can answer that, the money has a path back.
If they cannot, reality will send the invoice.
Sources and further reading