Ex-OpenAI researcher: AI labs will burn $100B on training data because scaling is hitting a wall

Andrew Ho spent eight months inside OpenAI. He walked away convinced the industry is chasing the wrong rabbit. The problem isn’t that models need more parameters or compute — it’s that they’re starving for data that actually captures how humans do economically valuable work.

“Most work is highly contextual and not easily encoded into a gradable environment,” Ho writes. He’s not talking about writing sonnets or passing bar exams. He means the messy, tacit knowledge embedded in bioinformatics pipelines, wet lab experiments, and the millions of other domain-specific tasks where current AI fumbles. Even OpenAI’s GPT-5.6 Sol hits roughly a 30 percent success rate on complex scientific analyses, a figure Ho knows firsthand from his time there.

His new startup is a direct bet against the “scale is all you need” religion. The first products target specialized datasets for bioinformatics and routine lab work — think researchers snapping photos of experiments and expecting reliable AI evaluation. Chemistry, materials science, and healthcare are next. Ho pegs the coming spend on targeted data collection at over $100 billion, a number that sounds wild until you look at the economics. Frontier labs like OpenAI and Anthropic, he notes, are chronically unprofitable. They’re forced to sink growing fortunes into each new model generation just to fend off cheaper competitors like Qwen and Kimi.

Ho’s skepticism has company. Cambridge researcher Adam Hunt describes his own trajectory from LLM optimist to pessimist. The latest models, he argues, aren’t becoming more versatile — they’re specializing. Programming and complex math climb while language quality and basic reasoning stagnate or degrade. Google DeepMind’s Tom Zahavy offers a structural diagnosis: language models excel at deduction and induction but can’t perform creative abduction, the leap to invent explanations for which no linguistic precedent exists. His proposed fix involves action-controllable world models that enable counterfactual experimentation. Hunt himself puts his confidence in the scaling-alone thesis at only about 40 percent. The billion-dollar question — whether stitching specialized models together can approximate the general intelligence that pure scale couldn’t deliver — remains completely open.