How big is the firehose?
Start from the raw broadband rate. To catch extracellular spikes you sample each channel at tens of kilohertz with ten or more bits. The total rate is simply the product of channels, sample rate, and bit depth.
Raw data rate. At C=10^6 channels, f_s \approx 30\,\text{kHz}, b \approx 10\,\text{bits}, this is on the order of hundreds of gigabits per second — orders of magnitude beyond any implantable wireless link.
That number is decisive: it means whole-brain-scale recording can never simply stream raw voltage off the head. Something must reduce the data before it leaves the implant, which is why on-implant computation has become inseparable from high-bandwidth interfacing.
Compress before you transmit
The oldest and most effective compression is to keep only what you decode from. Threshold crossings — detecting spikes on chip and transmitting only their times and a small feature — cut the rate by two to three orders of magnitude versus broadband. On-implant compression can go further with compressed sensing, exploiting the sparsity of spiking to reconstruct from far fewer measurements.
Compressed sensing. A signal that is k-sparse in n dimensions can be recovered from about m random measurements — logarithmic, not linear, in n. Because spiking is sparse, this lets the implant digitize and transmit far fewer numbers than the naive Nyquist rate demands.
The wireless bottleneck
Even after compression, getting bits off the head is hard. A fully implantable system must transmit through skin without a percutaneous connector, and wireless telemetry is limited by power, not by information theory: every transmitted bit costs energy, and that energy competes for the same thermal budget as the amplifiers. This is the wireless data-rate bottleneck.
The link is power-bound. Achievable data rate is the radio power budget divided by the energy per transmitted bit E_{\text{bit}}. Since P_{\text{radio}} draws on the same milliwatts as P_{\text{safe}}, the data wall and the thermal wall are one coupled constraint, not two independent ones.
This coupling is the quiet reason high-bandwidth BCI has become a co-design problem. You cannot maximise channels, minimise heat, and stream broadband all at once; you must decide, on the implant, what is worth the joules to send. The winning architectures spend their power on computation that shrinks data rather than on radios that ship it.
Co-design at the edge
The logical endpoint is to move the decoder itself onto the implant, so that what crosses the skin is not neural data at all but a low-rate control or state signal. That pushes the field toward ultra-low-power, in-situ inference — the subject of the neuromorphic co-processor track — and turns the data wall from a transmission problem into a computation one.