Naveen Rao claims 4D chip solves AI energy crisis

AI is running out of power fast.

On All-In, Unconventional AI chief executive Naveen Rao warned that global AI workloads will exhaust data center power capacity within three years. Physical floor space and GPU availability previously constrained the sector. Grid power availability now forms a hard ceiling on AI expansion.

To illustrate the crunch, Rao pointed to Google's monthly processing volume of 3.2 quadrillion tokens. At 10 joules per token, Google consumes 12 gigawatts for AI services alone. That single workload eats nearly a third of the 40 gigawatts powering every U.S. data center combined.

Legacy silicon architecture bears primary responsibility for the bottleneck. Modern GPUs shuttle nearly 30 trillion bits in and out of memory every second, expending most of their energy merely moving data across the chip. Because Moore's Law has stalled, shrinking transistors no longer yields structural power savings.

Biology provides a stark contrast. The human brain operates on 20 watts, while a squirrel executes spatial navigation on eight milliwatts.

To mirror biological efficiency, Unconventional AI developed a non-von Neumann computer merging memory and processing into physical oscillating circuits. Termed 4D computing, the architecture combines 3D physical die stacking with time-based dynamical state trajectories. The system bypasses floating-point math and matrix multiplication, using semiconductor physics directly to process neural representations.

More than two years after taping out a prototype in June 2024, the company recorded image generation at 500 nanojoules per output - orders of magnitude below the millijoule range required by GPUs.

Rao aims to deliver a commercial data center rack system within two years. Software development teams will interface at the model layer through Python rather than writing low-level CUDA operations.

Physical limits are forcing hardware into uncharted territory.