I keep seeing the same reaction whenever a major AI company reports a huge loss: “See? AI isn’t sustainable.”
Maybe the current spending levels aren’t. But that doesn’t mean the technology itself isn’t.
Anthropic’s November 2026 filings make the point pretty well. It reported $33.2 billion in revenue for the first nine months of the year, yet disclosed about $895 billion in infrastructure-related spending commitments. That’s a staggering number. It’s also important to understand that those commitments stretch across many years, not one giant bill due tomorrow.
There’s a real financial problem to solve here. I just don’t think “AI is dead” is the conclusion.
We’ve seen expensive infrastructure races before
Think about the early cell phone industry. Companies had to build towers, networks, switching systems and coverage before they had enough paying subscribers to justify the expense. Some companies struggled or failed. Cellular service didn’t disappear.
The same thing happened with parts of the internet boom. Plenty of businesses failed while the underlying technology became part of everyday life.
AI has some similarities, but I think the closer long-term comparison may be the American power industry.
You have companies that generate electricity, companies that move it through infrastructure, and businesses that package and deliver services to customers. Most businesses don’t build a power plant just to keep their lights on.
AI may separate into similar layers: data-center and compute providers, companies developing models, and businesses delivering useful AI-powered tools. Some companies will operate in more than one layer, but they won’t all have to.
That’s where the economics could get interesting.
Bigger isn’t automatically better
Right now, the headlines focus on enormous data centers, increasingly capable models, and how many billions companies are spending.
But customers aren’t buying benchmark scores. They’re buying results.
If a smaller model can handle a customer’s task at a fraction of the cost, why use a much more expensive model every time? If a local AI can perform routine work on a computer someone already owns, why send every request to a cloud data center?
Local computing has limits, and cloud infrastructure isn’t going away. But a system that chooses the appropriate model for each task could make AI substantially cheaper to operate.
I suspect the companies that figure out how to deliver reliable intelligence efficiently will have a durable advantage over those relying on spending alone.
This isn’t just another business competition
It’s also a global technology race.
It reminds me of the Space Race mixed with the cellular buildout and the Industrial Revolution. The stakes include national security, economic influence, jobs, energy, and who controls the infrastructure businesses depend on.
China is competing aggressively while also regulating AI services and agents. The United States and Europe are wrestling with their own questions about safety, privacy, competition and infrastructure.
None of those governments can afford to ignore the technology. But that doesn’t mean they’ll all regulate it the same way.
Regulation may become the biggest wild card
Politicians have to make decisions about technologies that can change faster than legislation gets written. They rely on advisers, companies, advocacy groups and voters who don’t always agree on the risks or the solutions.
In the United States, the November 2026 midterms have set up a Democratic-controlled Congress alongside a Republican president beginning in January 2027. That could mean federal gridlock on major AI legislation, but it doesn’t mean state regulations, agency actions or congressional oversight stop.
I expect more pressure over the next few years on data-center construction, energy use, privacy, autonomous systems and accountability.
Some guardrails are necessary. But poorly designed rules could also make it harder for a small developer to compete while the largest corporations simply absorb the compliance costs.
That is a genuine concern.
So is AI sustainable?
Parts of the AI economy are already profitable. Nvidia is a clear example of a profitable company supplying the infrastructure. That doesn’t prove every frontier-model developer will make a profit, any more than profitable electrical-equipment manufacturers guaranteed every power company would survive.
There will probably be failures, mergers, changing business models and more political fights over where all of this is headed.
But I wouldn’t confuse the financial challenges of today’s biggest AI companies with proof that AI has no economic future.
I’m building with AI because I see what it can do when it’s put to work on actual problems. I’m also very aware that running a system efficiently matters just as much as making it capable.
The winners won’t necessarily be the companies with the biggest models or the loudest announcements.
They’ll be the ones that can deliver useful work reliably, affordably and at scale.
That’s the part of this race I’ll be watching.