Every few months, a new AI model arrives that is a little smarter than the last. What almost nobody mentions is how it got that way: by spending more money. A lot more. The quiet truth of modern AI is that progress has been bought, not invented — and the bill is becoming difficult to pay.
Here is the part that should give everyone pause. The cost of training the largest AI models is doubling roughly every eight months. Not every few years — every eight months. The leading labs already run single training jobs that cost on the order of a billion dollars, and their own executives openly predict ten- and hundred-billion-dollar models within a couple of years. By 2027, total industry spending on AI computation is projected to approach one trillion dollars.
A trillion dollars is hard to picture, so here is what it actually means. It means the ability to build frontier AI is becoming concentrated in the hands of a few organizations that can afford it — and no one else. It means the cost of running these models, every time you ask one a question, is climbing too. And it means the entire field is betting that the answer to "how do we make AI smarter" is simply "spend more," forever.
That is not a healthy way for a technology to grow. When progress depends on ever-larger checks, the people who can write those checks decide what gets built, who it serves, and what it costs to use.
There is another way to make AI better, and it gets far less attention because it is harder and less glamorous than buying more chips. You can make each part of the model work harder — extract more capability from every parameter, every watt, every dollar.
Think of it like an engine. One path to a faster car is a bigger engine that burns more fuel. The other is a better-designed engine that does more with the same fuel. The frontier labs are building bigger engines. The opportunity almost no one is funding is building better ones.
Progress that depends on ever-larger budgets ends up controlled by whoever has the largest budget.
This is the problem Arch exists to work on. We are not trying to out-spend anyone — that race is already lost to the labs with billions to burn. We are trying to prove that a smarter mechanism can deliver more capability per dollar, so that good AI does not have to mean expensive AI.
We do not have the answer yet, and we will not pretend we do. But we think the most important question in AI right now is not "how big can it get?" It is "how much can we do with less?" — because the alternative is a future where only a handful of trillion-dollar players get to decide what AI is for.