Let me start with a number: somewhere north of four trillion dollars.
That is my best estimate of what the world has spent, combined, on the AI revolution so far. Infrastructure, compute, talent acquisition, startup funding, enterprise licenses, the whole circus. Four trillion dollars, and counting. To put that in perspective, that's roughly the GDP of Germany. We have spent the economic output of Europe's largest economy on a technology that, as of right now, we don't really have all that much to show for.
And before the replies start pouring in telling me I'm wrong, let me be clear about what I mean. I am not saying nothing has happened. I am not saying the technology is useless. I am saying that the return on four trillion dollars should be visible by now, and it isn't. Not at the scale that was promised. Not at the scale that would justify the spend. And there is a building mountain of research that backs this up, which we'll get into.
But first, I want to talk about a pattern I've been observing, because I think it tells you everything you need to know about where the industry actually is versus where it says it is.
When this all kicked off in earnest, the conversation was about AGI. Artificial General Intelligence. A machine that could think like a human, across all domains, flexibly and autonomously. That was the north star. For a brief, almost delirious moment in late 2023, some people were even talking about superintelligence, a machine smarter than all of humanity combined. I remember those conversations. They had a particular fever to them.
Then the conversation shifted to agents. Autonomous AI systems that could take actions in the real world, browse the web, write and execute code, manage workflows. Still ambitious, but notice the step down? We went from "a new form of intelligence" to "software that can do tasks." That's a significant recalibration dressed up as a product announcement.
And now, as I write this in early 2026, the conversation has shifted again. Now we're talking about tooling for agents. Infrastructure. Frameworks. The plumbing. We've gone from building God to building the pipes for slightly smarter software bots. Each step in this sequence is a quiet concession. Each one is the industry adjusting its expectations downward while trying very hard to make it look like progress.
AGI. Superintelligence. Agents. Tooling for agents.
Read that sequence again and tell me which direction it's heading.
This follows a pattern that I think is directly connected to the financial model that modern AI is built on. The western frontier model labs have, by and large, stopped meaningfully innovating on their core models. What we're seeing instead is incremental improvements, better packaging, and a whole lot of effort spent on making the thing we already have more useful, which to be fair is valuable work, but it is not what the four trillion dollars was supposed to buy.
The financial model is the key to understanding all of this. Training frontier models costs billions. Running them costs billions more. The revenue these companies generate, while impressive in absolute terms, does not come anywhere close to covering the capital expenditure. This is a classic pattern in technology hype cycles: you spend first and hope the returns come later, but "later" keeps getting pushed back, and at some point the math either works or it doesn't.
Now, about the Chinese.
I hear the whisper in the background, and it's worth addressing directly. The Chinese frontier model developers still do not have a single model that beats the American labs cross-domain. This is a fact. But what they are doing is, in my opinion, more interesting and more important than chasing the bleeding edge. They are finding ways to make these models dramatically more efficient, less resource-intensive, and then sharing them with the world as open source and open weights.
This is a fundamentally different philosophy. While the western labs are spending billions to push benchmarks up by a few points, the Chinese labs are asking a different question: how do we make what already works accessible to everyone? And they're succeeding at it.
So, point Chinese frontier model developers. 100%.
Because efficiency and accessibility is where the actual long-term value of this technology lives, not in another marginal benchmark improvement that costs another billion dollars to achieve.
The mounting evidence, and I encourage you to look into this yourself, is starting to tell a story that is difficult for the industry to spin. Enterprise adoption is slower than projected. Productivity gains are harder to measure than expected. The killer app, the thing that makes this all make sense the way the smartphone made mobile internet make sense, hasn't materialized. What we have instead is a collection of useful tools that are genuinely helpful in certain contexts but that have not, as of yet, justified the scale of investment that has gone into them.
Four trillion dollars. And the expectation staircase keeps going down.
In the next entry, I want to get into why. Not the business reasons, but the technical ones. Because there is a number that I think explains more about the current state of AI than any earnings call or product launch ever could.
That number is 85.