The artificial intelligence industry’s obsession with scaling is headed for a cliff
A new study MIT suggests that the largest and most intensive AI models may soon be less efficient than smaller models. Charting scaling laws against continuous improvements in model efficiency, the researchers found that performance leaps from giant models become more difficult, while productivity gains could make models running on more modest hardware increasingly capable over the next decade.
“In the next five to 10 years, things will probably start to slow down,” says Neal Thompson, a computer scientist and MIT professor who contributed to the study.
Leaps in efficiency, like the one seen in DeepSeek’s ultra-low-cost model in January, have already served as a reality check for the AI industry, which is used to burning through massive amounts of computing.
As things stand, a frontier model from a company like OpenAI is already much better than one trained with a fraction of the calculations from a university lab. While the MIT team’s prediction may not be true if, for example, new training methods such as reinforcement learning yield surprising new results, they suggest that big AI companies will be less dominant in the future.
Hans Gundlach, the research scientist at MIT who led the analysis, became interested in the subject because of the intractable nature of advanced models. Together with Thompson and Jason Lynch, another research scientist at MIT, he mapped the future performance of frontier models compared to models built with average computational tools. Gundlach says the predicted trend is particularly evident for the inference models now in vogue, which rely more on additional computations during inference.
Thompson says the results show the value of refining the algorithm as well as scaling up the computation. “If you’re spending a lot of money training these models, you should definitely spend some of it on developing more efficient algorithms, because that can be very important,” he adds.
The study is particularly interesting given today’s AI infrastructure boom (or should we say “bubble”?) — which shows little sign of slowing down.
OpenAI and other US tech companies have signed $100 billion deals to build AI infrastructure in the US. “The world needs a lot more computing,” OpenAI president Greg Brockman declared this week as he announced a collaboration between OpenAI and Broadcom for custom AI chips.
An increasing number of experts are questioning the authenticity of these transactions. Almost 60% of the cost of building a data center is spent on GPUs, which depreciate quickly. The partnership between the main actors also seems circular and non-transparent.
