AI Can’t Cost This Much (Predictions 2026, #8)

Some of you gave disagreed with my last prediction, that Anthropic would file for an IPO, stating, accurately, that OpenAI has a far more pressing need for fresh capital, given its commitments to various partnerships totaling  more than $1.4 trillion and counting. That’s a good point, but I don’t think OpenAI will ever really spend that money, and my next prediction explains why: I think the costs involved with delivering AI will come down significantly in 2026.

I’m not either an economist nor a supply chain expert, so what I’m about to write is informed more by historical rhyming than quantitative analysis. But when I see eye-watering numbers about the cost of data centers, compute, and chips, I start to wonder if innovation has been factored into the calculations. When trillions of dollars are projected to be spent, trillions that would require trillions more in revenue (and profit) to justify, a lot of butterflies start to flap their wings.

Here’s a historical analogy. Imagine it’s 1960, and the computer industry analysts of the day (there weren’t many of them!) accurately projected today’s demand for compute power. If they then used the costs associated with the “big iron” of the 1960s as the base for funding that demand, they’d quickly find themselves running out of zeroes on their calculators. With the most sophisticated computers from IBM costing roughly $23 million in today’s dollars and having roughly 90 billion times less compute power than a typical data center node today, well, those analysts would have no choice but to conclude that the future of computing would costs orders of magnitude more than the entire global GDP. They’d be laughed off the pages of the world’s newspapers (there were many of them back then!), and then likely fired.

Of course, in 1960 no one had heard of Moore’s Law, because it hadn’t yet been postulated.

With that hypothetical in mind, let’s return to the present day. However many trillions will be spent on AI over the course of the next decade, one thing will remain constant: There are huge opportunities to be exploited in finding more efficient and less expensive ways to deliver technology to both consumers and businesses. DeepSeek was just the starting gun to that race, and nearly every major research lab is utilizing lightweight, open models to both cut costs and speed development. Just as Google revolutionized the Web by creating homemade, cheap, and replaceable servers, putting the entire Internet in RAM in the process, so will today’s tech industry find a way to do far more with far less. There’s just too much money to be made – or at least not lost – for this principle to not kick in.

Put succinctly, I think 2026 will be a year of innovation when it comes to the cost of compute, as well in how much compute is actually needed to perform the magic we’ve come to expect from AI applications. It’s happened over and over again in this industry, and I think pricing the future based on the cost of the present is a losing bet.

—

This is the third in a series of post I’ll be doing on predictions for 2026. The first two are here and here. When I get to #1, I’ll post a roundup like I usually do. 

You can follow whatever I’m doing next by signing up for my site newsletter here. Thanks for reading.