A few weeks back I wrote about the uncertain future of AI pricing — the idea that we’re all enjoying a productivity glitch right now, and none of us really know when the bill comes due.

Then I wrote about how most of what small businesses call “AI” is actually just automation wearing a badge it didn’t earn.

Both posts kept circling the same question without quite landing on it:

Who actually wins here? And who gets left with nothing?

I think Warren Buffett answered this a long time ago. Just not about AI specifically.

Buffett has this idea he keeps coming back to. Some technologies make businesses more profitable. Others just benefit the consumer — and the companies that built them end up with very little to show for it.

His example is airlines. Extraordinary technology. Changed the world. And collectively, the industry has produced almost no profit for investors across its entire existence. All that progress went to the passenger. Cheaper seats, more routes. Great for society. Terrible for shareholders.

He called it simply: some technologies flow to the business, some flow straight to the consumer, skipping the investor entirely.

Right now, AI looks a lot like the latter.

Token costs have dropped something like 99% in two years. Every time one model gains ground, the others match it or undercut it. That’s not a sustainable business — that’s a commodity market forming in real time.

But not everything in AI is the same bet. There are really three places your money could land, and they’re not equal.

Layer 1: The Model Companies

OpenAI, Anthropic and the rest — are in the most brutal part of the race. Enormous capital, fierce competition, open-source alternatives constantly eating at their position. This is where the consumer wins and the investor might not. And if nobody can hold a price up, what does the business actually look like in five years?

Layer 2: The Pickaxe Sellers

Nvidia isn’t competing to give you AI for free — they’re selling the equipment that everyone building AI desperately needs. That’s closer to what Buffett calls a moat. But how long does it hold before custom chips from Google, Meta and Amazon start closing the gap?

Layer 3: The Quietly Positioned

This is the group most people aren’t talking about. Microsoft didn’t build GPT-4, but they own the relationship with hundreds of millions of Office users. When they weave AI into those products, they’re deepening a moat they spent decades building. Same with Salesforce, Adobe, the whole enterprise software world. AI becomes another reason to stay, not a new thing to compete over. The question is whether that’s genius positioning or just luck — and whether smaller players can find a version of the same move.

The dot-com era in the late 1990s played out the same way. Most companies that built the early web didn’t survive it. The ones that did were either infrastructure you couldn’t avoid, or platforms — Google, Amazon — that locked in a behaviour so completely that leaving became unthinkable.

We’re probably early enough that the AI equivalent of those two companies hasn’t fully separated from the pack yet.

But Buffett’s framework suggests they won’t necessarily be the companies making the most noise right now.

They’ll be the ones that were already hard to dislodge — and just got harder.