For years, the major AI labs followed one formula: build bigger models, train them on more data and give them more computing power. The formula delivered huge gains and set off a global race for scale. That race shaped how companies set AI budgets and how investors valued the most advanced models. Now the formula’s costs are rising faster than its benefits.
Ilya Sutskever, a cofounder of OpenAI, said in November 2025 that scale alone is no longer enough and that future progress depends on new research ideas. Scale has been the leading labs’ main edge. Their models still score well on tests, but in real work, reliability is falling further behind those scores. Business leaders see it in accuracy that drifts, results that vary from task to task and productivity gains that arrive slower than expected.
The data that fed the formula is running short. An analysis by Epoch AI, a research group, shows that the supply of high-quality human-written text suitable for training is close to used up. That puts a hard ceiling on the old approach. Developers are turning to text written by AI models, known as synthetic data. MIT researchers found it can help with narrow tasks, but models that learn from their own output risk becoming less reliable, with quality problems that compound over time.
Computing power is tightening too. The OECD points to two growing bottlenecks in training the most advanced models: energy, and how fast chips can move data in and out of memory. Hardware keeps improving, but not as fast as demand. Power is now a major cost, and each new generation of models needs more money, time and infrastructure.
Safety checks are slowing the rollout of new models as well. METR, a group that tests AI models for safety, finds that larger models often gain unexpected abilities in sudden jumps, which makes them harder to prove safe. The extra testing stretches timelines and adds cost.
Meanwhile, the returns from size are flattening in everyday work. Smaller models that do a narrow job go into use faster, cost less and carry less risk, and their impact is easier to predict. Applied research from MIT finds that with extra training on targeted, high-quality data, those smaller models can match or beat large general-purpose models in specific business settings.
