The wiring wall
A passive microelectrode array like the Utah array needs one wire per electrode. That is fine at a hundred channels and impossible at a hundred thousand: the bond pads, the cabling, and the percutaneous connector all scale linearly and quickly dominate the device. Passive arrays hit a wiring wall well before they hit a scientific one.
Active CMOS probes and multiplexing
An active (CMOS) electrode array integrates amplifiers, switches, and an analog-to-digital converter onto the shank itself. A Neuropixels probe carries close to a thousand sites and reads out a few hundred at a time by time-division multiplexing them onto a shared digital output — the discontinuity that recently bent Stevenson's curve. The engine is a neural front-end ASIC doing amplification and digitization in situ.
The multiplexing budget. A shared converter running at f_{\text{ADC}} must service all C channels at per-channel sample rate f_s; you buy channel count by giving up per-channel rate, resolution, or by adding converters (which cost power).
Recording yield at scale
A site is only useful if it hears a neuron. Recording yield at scale — the fraction of channels that actually return a sortable unit — tends to fall as you pack sites closer, because the spatial reach of an electrode is finite and neighbours over-sample the same cells. This is the coverage-versus-density tradeoff: past a point, denser sampling of the same tissue adds sites without adding neurons.
Yield also decays over time as micromotion and the foreign-body response wall off electrodes. So the headline channel count of a chronic implant is an upper bound on its scientific yield, and the gap between the two widens exactly as you scale — a reason to report neurons resolved, not electrodes fabricated.
The thermal ceiling
Every amplifier, converter, and multiplexer on the implant dissipates power, and that power becomes heat in tissue that tolerates only about a degree or two of sustained warming before harm. This thermal budget is not an engineering inconvenience you can optimise away — it is a hard physical limit set by tissue heating.
The thermal cap. With per-channel power p_{\text{ch}} and a safe dissipation P_{\text{safe}} (set by keeping the tissue temperature rise to roughly a degree), the maximum channel count is fixed regardless of how clever the decoder is. Scaling C therefore requires driving p_{\text{ch}} down.