Companies spend big training AI models — but does the cost of gathering good data quietly cancel out those gains?

This explores whether the cost of getting good training data cancels out what companies gain from training AI models. The short answer: the corpus has no direct accounting of data collection costs, so it can't settle the question. What it does show is that data matters in two ways that a spending tally misses. Data is a small line on the budget but a large constraint on quality, and the hardest data costs are hidden.

Start with where the money is visible. Epoch AI's breakdown of Anthropic, Minimax, and Z.ai finds that compute for research and for serving models takes up 54–62% of operating expenses, while staff stays under 25% Do compute costs really dominate AI company budgets?. At the facility level, the servers themselves account for about 60% of a one-gigawatt data center's annual cost, and energy is a surprisingly small share What actually drives AI data center costs?. Data collection doesn't appear as a major category in either analysis. On these numbers, hardware is what squeezes returns, not data.

The more interesting argument is that data's cost is a ceiling rather than a bill. One note argues that AI-generated knowledge has nothing solid backing it, because training data is finite and expert validation can't scale. The predicted result is more output alongside less reliability, a kind of "epistemic stagflation" What actually backs the value of AI-generated intelligence?. Seen this way, the real data cost is the expensive human checking that cheap data lacks, not the price of collecting it. That cost shows up quietly as weaker models instead of as a budget line. The same pattern shows up with measurement. When training relies on cheap proxies for quality, models learn to game them through reward hacking and sycophancy How vulnerable is AI training to Goodhart's Law?. Benchmarks built from easy-to-collect contest tasks also fail to predict performance on real professional work Why do agent benchmarks not predict real economic value?.

The corpus also points the other way: the right data can replace spending on scale. Occamy-1.0, a 35B-parameter model trained on execution-grounded data and long multi-step task records, competes with much larger models at a fraction of the cost Does model efficiency matter more than peak capability for real work?. Here, targeted data adds value to the training investment. Choosing how to add knowledge raises a related trade-off. Building knowledge into the model's weights is fast at runtime but costly and rigid. Retrieval (RAG) moves that cost to query time How do knowledge injection methods trade off flexibility and cost?.

Finally, there's the question of when costs show up. Brynjolfsson, Rock, and Syverson's productivity J-curve says hidden complementary investments, such as training people, redesigning processes, and building software, make AI look unproductive at first and then overstate its gains later Why does AI productivity appear weak when investment is high?. Data collection and curation probably belong in that hidden category. If so, "erosion" may partly be a timing illusion: the cost lands early, and the value shows up in the statistics years later.