The key constraint in building AI infrastructure has shifted. Previously, the challenge was securing floor space, networking, and then GPUs. Now, the first and most critical step is securing a power contract, as energy accounts for 50% of the cost of serving a token. The goal has become monetizing every available watt.
Each layer of abstraction in modern computers—from digital 1s and 0s up to neural networks—is inherently "lossy" and creates inefficiency. A more efficient approach bypasses these layers by directly connecting the physics of a semiconductor to the structure of a neural network, mimicking how brains compute without linear algebra.
A "dynamical computer" performs calculations as an emergent property of its physical system, much like metronomes synchronizing on a plank. This paradigm merges compute and memory into one element, eliminating the massive energy cost of moving data that defines traditional von Neumann architecture.
Counter-intuitively, selectively removing connections in a complex computing system (sparsity) can lead to superior results. This approach not only makes the system more energy-efficient and scalable (avoiding n-squared scaling), but it also improves its overall performance by making it more trainable—a "holy grail" in system design.
A radical improvement in compute efficiency won't just lower costs; it will trigger Jevons' paradox, where consumption increases by more than the price drops. Making AI compute 1000x cheaper will unlock currently unimaginable applications, creating a market far larger than linear projections and potentially the largest in human history.
The primary organizational challenge in deep tech is managing the vast intellectual and cultural span between abstract thinkers (like dynamical systems theorists) and practical builders (chip engineers) who typically "don't talk to each other." Facilitating communication across these disparate disciplines is as critical as the scientific breakthroughs themselves.