The Future Of AI Will Be Decided In The Physical World

Daisuke Okanohara is CEO of Matlantis Coporation, and Co-founder and CTO of Preferred Networks, Inc.

​Every conversation about the future of AI circles back to the same assumptions: bigger models, more data, more compute. We treat progress as a contest of intelligence. Yet the constraint that will decide how far AI goes over the next decade has little to do with intelligence and a great deal to do with physics.​

AI is running into two physical walls that no amount of clever software can talk its way past.

The First Wall: Energy

In a modern AI system, far more energy goes into moving data around than into performing the calculations themselves. The calculations inside a chip are remarkably cheap. The real expense lies in shuttling enormous volumes of data between memory and processor, billions of times every second. By some engineering estimates, that movement can consume ten to a hundred times more energy than the computation it serves.​

For two decades, processing power has raced ahead while the speed of feeding data to it has crawled, and today’s large language models are limited less by how fast they can think than by how fast they can be fed.​

The bill lands on the balance sheet and on the power grid. Running the AI we expect to deploy across the economy could require tens of gigawatts of additional electricity. The ceiling on AI’s growth is becoming an energy ceiling, and that energy is spent in a place few people think to look: the distance data must travel.

The Second Wall: Materials

The second wall sits one level deeper, in the materials our hardware is built from.​

Better chips, better batteries, better energy systems: nearly every hardware advance the AI era depends on eventually comes down to discovering or improving a material. Materials discovery has always been agonizingly slow. The number of possible combinations is effectively limitless, while the number humanity has tested remains vanishingly small.​