The era of achieving intelligence through brute computational power is hitting a dead end. Andrew Ho, an OpenAI alumnus, has launched a startup focused on generating specialized data after realizing the obvious: Scaling Laws no longer provide a qualitative leap. According to Ho, the industry is entering a phase where labs will need to spend over $100 billion—not on Nvidia chips, but on targeted data collection. He argues that the skills critical to the economy are simply absent from the massive troves of junk internet text. The results are telling: even advanced models like GPT-5.6 Sol show a dismal 30% success rate in complex scientific analyses such as bioinformatics.
Ho's skepticism is shared by researchers at Cambridge and Google DeepMind. As researcher Adam Hunt notes, while coding and mathematics progress thanks to clear reward signals, language quality and basic logic are stagnating. We are hitting a ceiling: models are "suffocating" on their own hallucinatory content, unable to move beyond the probabilistic repetition of the internet.
Main takeaways
Investment focus is shifting from GPU procurement to the creation of proprietary datasets. Traditional web scraping has exhausted its utility for solving applied business problems. A shortage of "contextual" data is hindering AI adoption in knowledge-intensive industries.
Business value is rapidly migrating from general-purpose chatbots to closed datasets in biomedicine and materials science. Ho’s project aims to encode "contextual" workflows and counterfactual paths—the things human experts do intuitively that cannot simply be scraped from the web.
In our view, this is the logical conclusion of the arms race: competitive advantage is no longer determined by the size of a server cluster, but by the possession of verified data that allows models to operate in high-stakes industries.
Recommendations for business
Audit internal repositories for unique, high-value data. Digitize expert workflows that are not available in the public domain. Build proprietary training sets before specialized providers monopolize your niche and set prohibitive pricing.