From Billions to Trillions… will it ever stop?

Everyone knows it's expensive — that was never the problem

Nobody is hiding the cost of AI — it is the headline. Training a frontier model runs into the billions, and the industry's build-out is now discussed in the hundreds of billions, with talk of trillions. Even the biggest labs are stuck in a loop of ever-larger funding rounds just to stay in the race. But raising more money isn't the same as solving the problem. It's feeding a bigger meter.

Because the real cost isn't the training run — it's what comes after. Serving the model costs money on every single query, and the standard business model charges you per token. You don't buy the tool. You rent its output by the word, and the meter never stops.

The retreat

So companies are quietly pausing and scaling back. When the cost of each query doesn't clear the value it delivers, the math stops working — and leaning on a vendor's pricing, uptime, and terms turns into a business risk, not just a line item. The technology works. The economics often don't.

The bet the whole market is making

The cost of first-generation AI is starting to spill past the labs and onto everyone else. More and more, the money keeping it running is public money — raised through IPOs, and sitting in the index funds and pensions that track the market — poured into ever-larger datacenters. Raising money from the public at that scale, to fund a brute-force approach that shows no sign of trying to need less, is irresponsible. It would be one thing if the spending were buying a path to something durable. Almost none of it is: it goes toward building bigger, not toward needing less, and the plan for the climbing energy bill is to add more capacity.

We have already seen a preview. When DeepSeek did more with less — out of necessity, not as a strategy, because the expensive way was never open to them — it woke the established labs up. For a while, efficiency was suddenly the thing everyone cared about. Then the moment passed and they went back to building bigger. That reaction is the tell: needing less was never their business plan, and a brief scare didn't change it.

The people inside see it too. Researchers have been leaving the big labs in a steady stream, and when they explain why, it often comes down to the same thing: decisions driven by profit and scale rather than by where the science actually points. When the people who understand the work best keep walking out over exactly that, it is worth taking seriously.

Picture a telephone company sitting on a monopoly it knows will end — and going public anyway, taking the public's money with no plan for how it competes once the monopoly is gone and the rules change. That is the shape of the bet here: an advantage you can already see expiring, sold to the market as if it won't.

And those new rules don't need anyone to enforce them. Most people can't see them yet, but they're coming all the same — not because someone is pushing them, but because a cheaper way exists and the logic does the rest. Once the same work can be done for a fraction of the cost, the market moves to it. Nobody has to make that happen; it happens because it's cheaper.

That makes it a fragile bet, and not only for the companies making it. The value of a datacenter built for a brute-force approach rests on that approach staying the only one on offer. The moment something equivalent — or better — runs at a fraction of the cost, a great deal of that investment is stranded: capacity built for a method the market has already moved past. When money is raised at this scale, from the public, the responsible move is to show you're working to need less. Building bigger instead is a wager that today's way is the only way — and we don't think it is. When a wager that size goes wrong, the loss doesn't stay with the people who placed it; it lands on the funds, the pensions, and the ordinary people who never chose to play. Everyone loses, and it's hard to call that anything but irresponsible.

A different bet: second-generation AI

We're making a different bet. Call it second-generation AI. It comes down to three things:

• Ownable. You host it, you have it. No metered output, no per-word meter running while you think.
• Efficient, and greener for it. The cost math is a design goal, not something we bolt on afterward: a wide-shallow trunk, a lean serving path, and reasoning handed off to a light symbolic machine instead of burned into massive activations. The target isn't a small discount — it's orders of magnitude less compute for the same work, and a smaller energy footprint with it.
• Durable. Something you own outlasts pricing changes, funding rounds, and hype cycles. Intelligence you keep, not output you rent by the word until the day the vendor turns it off.

To be clear, we're not against paying for software. There will be subscriptions. What we refuse is the per-token meter — being billed by the word for the privilege of thinking out loud. A price you can predict, or a model you can self-host. Never a meter that punishes you for using the thing you're paying for.

Our proof: self-funded since 2023

We're not arguing this from the cheap seats. We've been at this since 2023 — early, when the field was brand new and almost no one was asking whether it had to cost this much — and every bit of it has been self-funded. No venture round, no outside capital, nothing. If serious AI genuinely required billions, we wouldn't be here. That we are is the first piece of evidence that the bill is a choice, not a law of nature.

Why we can claim this

None of this is a discount we're promising. Low cost falls out of the architecture itself. Because the reasoning runs in a cheap symbolic machine rather than in the model's activations, and because the model is built to be lean to serve, efficiency is a consequence of the design rather than a sales line.

And you can watch it happen. Our world-model agent solves escalating 3D mazes — finding its own abilities, planning, running a small affective brain — with no policy network and no large model burning compute on every move. It runs in real time, light enough to host on a small box or in a browser. See it reason in the open →

What it still costs to run

We're not going to pretend Cubby runs on nothing. It uses compute, and we'll always say so. What we're chasing is the size of that number — orders of magnitude less than the standard approach, because the efficiency is designed in from the start rather than bought by the acre. Sustainable, open, and manageable, and a great deal greener than renting intelligence from a data centre that never stops growing.