For most of the last decade, the scarce resource in AI was ideas. The team with the better architecture, the cleaner data, or the sharper training recipe won. That is no longer the constraint that decides who ships.
Today the constraint is compute. A single frontier training run now costs more than a hundred million dollars. A 70 billion parameter model takes millions of GPU-hours to pretrain. At market rates for an H100, one run of that size is a bill in the tens of millions of dollars before a single result is evaluated. The idea is cheap. Executing it is not.
The budget is the roadmap
When compute is the binding constraint, the compute budget quietly becomes the research roadmap. Experiments that should be run are not, because the cluster is booked. Model sizes are capped not by ambition but by what finance signed off on. Whole directions get shelved because nobody can spare the GPU-hours to test them. The frontier is drawn by accountants as much as by scientists.
The standard response is to buy more hardware. That works, and every major lab is doing it, but it is a race with a known price and a known ceiling. Power, packaging, and supply chains all bend the curve. Spending your way through the wall gets more expensive at exactly the moment the models get larger.
Efficiency is the other lever
There is a second lever, and it is pulled far less often. If a run can reach the same model quality with fewer GPU-hours, every downstream number improves at once. More experiments fit in the same budget. Bigger models fit on the same cluster. The same result costs less. And unlike buying hardware, the savings grow as the models grow, because the waste being removed grows too.
That is the lever we work on. Not a faster version of the same computation, but less computation for the same outcome. The rest of this blog is about why we think that lever is real, where the evidence comes from, and why so few people are pulling it.