Is AI getting smarter mainly because companies spend more on chips, or because they're finding cleverer methods?

This explores how much of AI's progress comes from spending more (more chips, more compute, bigger data centers) versus getting smarter about how the same resources are used. One caveat up front: the corpus has no clean percentage split of capability gains between the two. What it does have is evidence on each side, and together that evidence suggests the two are tightly intertwined rather than competing explanations.

The spending side is very large. Epoch AI's analysis of Anthropic, Minimax, and Z.ai finds that compute for research and for serving models makes up 54–62% of operating expenses. Staff stay under 25%, even with high salaries Do compute costs really dominate AI company budgets?. And the money mostly goes to hardware, not electricity. In Epoch's cost model, a one-gigawatt data center costs about $8.5B a year, of which roughly $5B is servers and only $0.6B is energy What actually drives AI data center costs?. So when people say AI is "compute-bound," what they really mean is that it's bound by how fast you can pay off expensive chips before they go out of date.

The efficiency side moves surprisingly fast too. Researchers studied open-weight models so they could strip out falling hardware prices and price wars between providers. They estimate that better methods alone make running a model about 3× cheaper each year, and independent energy measurements back this up How fast is AI inference actually becoming more efficient?. A concrete example: Occamy-1.0, a 35B-parameter model trained on real execution traces and long multi-step tasks, competes with much larger models at a fraction of the cost. That suggests targeted training can stand in for raw size, at least for multi-step work Does model efficiency matter more than peak capability for real work?. Sakana AI makes the same bet on purpose. Working with Japan's smaller compute budgets, its lab treats self-improvement as a problem of learning from fewer examples rather than throwing more compute at it, and it has solved hard optimization problems with about 150 samples Can sample efficiency replace compute scale in recursive self-improvement?.

You might not expect this next part: AI is starting to produce its own efficiency gains. In one setup, an outer AI loop read the code of an inner research loop, found its bottlenecks, and wrote new search methods while running. The result was a 5× improvement on a GPT pretraining task Can an AI system improve its own search methods automatically?. But measuring this kind of gain is hard. Benchmarks that hold the evaluation budget fixed show optimization getting better, yet they can't tell you whether real research gets cheaper per discovery Do fixed-budget efficiency gains translate to real research progress?. Some researchers argue that any capability claim should report what it cost to achieve, because benchmarks can make a model look better or worse than it really is Do automated benchmarks hide what frontier AI systems can really do?.

The takeaway: spending and efficiency don't compete, they multiply. Labs pour money into hardware because each year's better methods make that hardware worth more. A 3× yearly efficiency gain on top of growing compute budgets compounds quickly. Still, if you want a hard answer like "60% compute, 40% algorithms" for training-time capability, this collection doesn't have that breakdown yet.