A billion neurons in silicon: neuromorphic computing grows up.

The brain runs general intelligence on roughly twenty watts. A modern AI datacenter needs a power plant. That gap — five or six orders of magnitude — is the entire thesis of neuromorphic computing: process information the way cortex does, as sparse asynchronous events rather than dense synchronous matrix math, and the energy bill collapses.

The field's scale milestone is Intel's Hala Point, announced in April 2024 and deployed at Sandia National Laboratories. It packs 1,152 Loihi 2 processors into a chassis the size of a microwave oven, simulating 1.15 billion spiking neurons — a scale Intel compares to an owl's brain — while peaking at 20 petaops and drawing a maximum of 2,600 watts. On sparse workloads, Intel reports efficiency exceeding 15 trillion operations per second per watt, territory GPUs cannot reach on always-on tasks.

The commercial end of the field is maturing as well. BrainChip, whose Akida processors target microwatt-class inference at the sensor, raised $25 million in December 2025 to push its Akida 2 and AKD1500 silicon into medical wearables, hearables, and defense sensors — real products, real power budgets, not lab demos.

Why does this field matter so much? Because neural interfaces and frontier models both run into the same wall. An implant cannot dissipate GPU heat inside a skull; event-driven silicon is how always-on neural decoding becomes implantable. And the economics of frontier-scale inference increasingly hinge on energy, not FLOPs. The deeper bet behind neuromorphic engineering is that the findings of neuroscience — sparse codes, event-driven attention, local plasticity — are not just biology trivia. They're a chip specification.