The scaling hypothesis is the belief that intelligence is fundamentally a function of scale. Train a bigger model on more data with more compute, and you don't just get better autocomplete — you get qualitatively new capabilities: reasoning, planning, and eventually something that looks like artificial general intelligence. OpenAI, Anthropic, and Google have collectively raised over a hundred billion dollars on some version of this thesis.

The idea crystallized around 2020 when scaling laws showed smooth, predictable performance gains as compute increased. Then emergent capabilities showed up — multi-step reasoning, code generation — appearing abruptly at certain size thresholds. Predictable curves plus surprise capability jumps made the case compelling. It still drives the bulk of frontier model investment today.

The hypothesis is genuinely contested. If it holds, capabilities keep accelerating and the cost of waiting is enormous. If it stalls — and there are serious researchers who think current architectures are approaching a ceiling — then the current generation of foundation models is close to the best you'll get, and execution quality matters more than raw model access.

You don't need to pick a side to act well. The right move is systems that work either way: modular architectures that can swap in better models as they arrive, clean data pipelines that give any model better inputs, and evaluation frameworks that tell you when a new release actually moves the needle for your use case. Bet on flexibility, not prophecy. The hypothesis may be right. The companies that succeed won't be the ones who believed hardest — they'll be the ones who built for optionality.