AI's Airline Problem

While I often speak about how AI isn’t living up to the hype, I don’t think hallucinations are the industry's biggest issue. All the big players in the space: OpenAI, Microsoft, Anthropic, and Google are trying to reach profitability, and I don’t see a way they get there.

The numbers don’t lie; these companies are hemorrhaging cash on computing costs, and eventually they’ll need to raise prices. The problem is: I don’t think they can.

As an educator, I’m always looking for metaphors to help my learners connect concepts to something they’re familiar with, and last week, I hit on one I like. The space is starting to look like the Airline industry.

AI Models are a Commodity

When choosing an AI model today, the financial calculus is remarkably similar to booking a flight. Sure, I have preferences. Claude is the best coder, GPT5 tends to be a slightly better writer, and Gemini… exists. But the question I’ve been turning over in my mind is: “What if the prices were significantly different?”

For flights, I’ll pay a bit more for Delta because I generally get a better experience. But there’s a ceiling to that loyalty because my real reason for flying is to get to my destination. If Delta’s prices spike too high, I’m flying American Airlines, or maybe I’ll just drive instead.

AI models are in a similar situation. Yes, I prefer Claude for coding work. Suppose I’m on the $20/month plan, which is the same as the ChatGPT Pro plan. If Anthropic increases the price to $25/month, will I still pay it? Probably. Would I pay $35? Probably not. I’d switch to whatever model offers comparable quality at a better price point.

I think the moat here is much narrower than these companies want to admit.

Three Questions Nobody Can Answer

The commodity situation creates a standoff that should concern anyone in this space. I keep coming back to three fundamental questions:

1. Who raises prices first? This is going to turn into a game of chicken. Whoever moves first risks an immediate exodus to competitors who hold prices steady. Look back at what happened to Netflix in 2011. They raised prices, users fled to alternatives, and the company lost 800,000 subscribers in a single quarter while the stock price dropped 77%. In today’s “number must go up” corporate environment, this should be keeping executives awake at night.
   
   In particular, OpenAI should be terrified of this. Google and Microsoft have profits they can plough into continued losses. OpenAI will have to continue using debt and equity to keep burning cash at the current pace.
2. What is the actual premium for the “best model”? The market hasn’t established this yet because we’re still in the land-grab phase. Now, there are a lot of variables to consider, but my hypothesis is that the premium will be small, maybe a 10-20% premium for demonstrably superior performance on specific tasks. Beyond that, I don’t see business or especially consumers buying in.
3. At what price point does self-hosting become viable? This is the elephant in the room. Open source models are improving rapidly. Llama, Mistral, Deepseek, and others close the gap with frontier models in months, not years. When (not if) the frontier models increase costs to reflect reality, spinning up your own infrastructure starts to make economic sense for any organization or consumer with high volume.
   
   Nvidia recently launched its DGX Spark computer, which handles 200 billion-parameter models. The price? $4,000. Think about that for a moment. As a tech professional, if I find myself spending more than $300 per month on AI tools, the break-even point is about a year. This puts a hard cap on what AI SAAS can charge. Self-hosting is even more compelling when you consider privacy and security.

The Infrastructure Trap

We haven’t even gotten to the massive energy requirements to power these data centers everyone wants to toss up. Sam Altman’s plans will require the equivalent of 17 nuclear power plants to be built before 2030. That’s enough power to service 13 million homes. 9 Hoover dams!

But some people point to the fiber-optic networks that provided value after the Dot-Com Crash. Today’s massive investments will create lasting value even if individual companies fail.

It’s not the same, not even close.

Fiber laid in 1999 still carries data in 2025. It’s a physical infrastructure with decades of useful life. GPU clusters? They’re obsolete in 18 months. New chips deliver better performance per watt. New architectures change the economics entirely. Nobody’s going to want to run inference on 2025 hardware when 2030 chips are twice as efficient and half the cost.

The dotcom crash left behind infrastructure with enduring value because the underlying technology was relatively stable. AI compute is not the same. It’s a treadmill where you need to keep buying new equipment just to stay competitive.

Remember that premium I wrote about earlier? Good luck getting it when your competitors are measurably faster and more efficient.

The Incoming Bubble Pop

Here’s what I see coming: these companies are locked in a prisoner’s dilemma. They need to raise prices to approach profitability. But they can’t raise prices without losing market share to competitors who hold steady. And they can’t raise prices too high without making self-hosting economically attractive.

Airlines attacked this problem through consolidation and capacity discipline. The industry went through waves of bankruptcies until only a handful of carriers remained, and they learned to compete on service rather than price alone. Even then, their margins are terrible.

AI companies won’t have that option. There’s no natural limit to capacity because you can always spin up more GPUs. Open source models ensure there’s always a cheaper alternative waiting in the wings. The barriers to entry are high in absolute terms but low relative to the market opportunity. Sure, maybe they’ll find a way to eke out an existence, but it’s likely the early investors are going to get destroyed.

What Comes Next

We’re heading towards one of three scenarios, none of which is great for the current market leaders or their investors.

First, the companies could achieve breakthrough improvements that dramatically lower their costs. This is certainly possible, but so far, the evidence shows the gains are incremental. But maybe, with just another $1T in investment, it’ll work out, right, Sam?

Second, they could find ways to differentiate beyond model quality. This could mean better integration, enterprise features, etc. See Southwest Airlines’ “Bags Fly Free”. The problem is that there isn’t anything in today’s models that competitors can’t easily clone. And, as I warned startup investors over a year ago, they’re already integrating 3rd-party features into their platforms after someone else has done the product-market fit exercise.

Third, they continue burning cash and suppressing prices in an attempt to be the “last model standing”. The majority of players never achieve the unit economics they need, values compress, and consolidation follows.

I’m betting on scenario three. As I’ve written about before, I do think the technology has economic value. The business model has significant problems. The airlines figured out a long time ago that they were selling a commodity. AI companies are about to learn the same lesson.