The Hidden Latency Tax of Wireless BCIs

The pitch is irresistible. Take the percutaneous connector — the screwed-into-the-skull pedestal that has been the open infection vector of every long-term cortical implant since the late 1990s — and replace it with a sealed, fully-implanted device that talks to the outside world over inductively-coupled radio. No more dressing changes. No more bacterial entry point. No more daily ritual of cleaning around an open wound.

Every major neurotech company in 2026 — Neuralink, Synchron, Paradromics, Precision, BrainGate's Blackrock-led wireless follow-on — is shipping or about to ship a fully implanted, wireless device. And they are right to. The infection risk on percutaneous pedestals is the single biggest reason chronic cortical implants did not enter mainstream clinical use in the 2010s. Going wireless is a genuine engineering achievement and a genuine clinical breakthrough.

It also introduces a tax that the marketing materials do not mention, and that the engineering teams are acutely aware of: latency.

What the Numbers Actually Are

A wired Utah array running into a connected Cerebus signal processor delivers neural data to the decoder with a round-trip latency, from spike to decoded output, of roughly 8–12 milliseconds. The recording electronics are fast, the data path is short, and the bottleneck is the spike-sorting algorithm rather than the wire itself.

A wireless implant, depending on architecture, sits somewhere in the range of 25–80 milliseconds round-trip. The variance is wide and the breakdown is informative:

• On-chip pre-processing. Wireless devices cannot ship raw broadband data off the implant — the radio bandwidth and the power budget would not allow it. So they do spike detection, often with on-chip threshold crossing or template matching, before transmission. This is fast (sub-millisecond) but it means the decoder operates on already-compressed information and cannot easily revise its spike-sorting if the noise floor shifts.
• Radio framing and ECC. The actual over-the-air transmission of a single packet is fast, but the framing, forward error correction, and retry-on-collision logic add several milliseconds of irreducible latency.
• External coil reception and host-side processing. The behind-the-ear or headworn coil receives, demodulates, and forwards to a host computer (usually a phone or tethered laptop) which runs the decoder. Each hop is small. They add up.
• Output device latency. A cursor on a screen has 8–16ms of display latency on top of everything else. A robot arm has 30–60ms of mechanical command latency. A muscle stimulator has its own delay.

When you add it all up, the user experience of a fully wireless system is closer to 60–100ms from intent to observed action. That is, in human-factors terms, perceptible.

Why Perceptible Latency Matters Here Specifically

Most people who work in human-computer interaction know the rough numbers. Mouse latency below 20ms feels direct. Between 20 and 50ms feels responsive but unmistakably mediated. Above 100ms feels like remote control. Above 200ms feels like the system is broken.

BCIs operate in a regime where the brain's own internal feedback loops — the cerebellar correction loop, the spinal stretch reflex, the visual-motor learning loop — are tuned for sub-100ms intervals. When you give a healthy human a cursor and a mouse, the cortico-spinal command-to-twitch latency is around 60ms; the visual confirmation of the cursor moving comes back to V1 around 80–100ms after the muscle fires. That whole loop is fast enough that the brain treats the cursor as a body part, not as an external object.

When you decode motor intent directly off cortex and route it through wireless radio plus host-side processing, you can end up adding 50–80ms of external latency to a loop that the brain expects to close in 100ms. The cerebellum, which is doing the heavy lifting on motor adjustment, has a hard time learning to predict a cursor whose response time wanders by 20ms packet-to-packet.

Pittsburgh's group published a small, infuriating study in 2023 showing that wireless decoder performance, even when nominally matched to wired performance in offline analysis, was meaningfully worse in closed-loop tasks. The participant could still control the cursor. It just felt, in their words, "like reaching through a wet blanket."

The wire is not just a data channel. It is a participant in the user's body schema. Cutting it has costs the engineering papers do not usually quantify.

What the Companies Are Doing About It

The wireless latency tax is being attacked from three angles, and the solutions are technically interesting.

On-implant decoding. Synchron's most recent Stentrode generation runs a quantised motor decoder on the implant itself, transmitting decoded intent commands rather than spike data. This collapses the wireless link out of the closed loop entirely — the decoder output goes to the external device with single-digit millisecond latency. The tradeoff is that the decoder cannot be updated as often, and re-calibration requires more clinician involvement. Neuralink has hinted at a similar architecture on the N2.

Predictive decoding. Several groups (BrainGate, the Stanford Henderson lab, Paradromics' applied research team) are publishing on Kalman-filter and small-network predictive decoders that extrapolate the next 50–80ms of intended trajectory and pre-position the output. If the prediction is right, the perceived latency drops to near zero. If it is wrong, the cursor overshoots and snaps back, which users describe as "the cursor lying to me." Tuning the prediction aggressiveness is an art.

Body-side coil placement. Moving the receiving coil from behind the ear to the temple, or to a high-bandwidth chest pack with a sub-cutaneous lead, shortens the RF path and improves SNR enough to reduce retry rates. The cosmetic and surgical implications are non-trivial, and the patient acceptance data is, predictably, mixed.

What This Means for Whose Implant You Get

The pragmatic clinical answer for the next several years is that which wireless implant you get is going to matter a great deal more than the marketing materials suggest. A patient whose primary use case is sentence composition on a tablet — typical of an ALS-stage cohort — can tolerate 80ms of round-trip latency just fine. Typing is not a tight motor loop. A patient whose primary use case is bilateral robotic arm control for object manipulation, on the other hand, will notice 80ms immediately and will perform measurably worse for tasks involving timing.

The implant manufacturers know this. The trial protocols reflect it. The next generation of FDA submissions is going to start including end-to-end closed-loop latency as a primary performance metric, alongside the spike yield and decoder accuracy numbers that have dominated the literature so far. That is, quietly, a substantial maturation of the field.

The Wire Was Honest

A small thought, before closing. The percutaneous pedestal was a bad design in almost every clinical respect — infection, comfort, social acceptability, MRI compatibility — but it was honest. The data path was a wire. The latency was deterministic. The failure modes were obvious. The engineering tradeoffs were on the table.

Wireless implants are, in every respect that matters to patients, better. The latency tax is real but tractable, and the trajectory of the engineering is fast enough that the gap to wired performance will probably close within five years.

But the era when a neural implant was, fundamentally, just a wire is ending. What replaces it is a stack of compromises — radio physics, embedded compute, predictive models, host-side software — that has to be reasoned about as a system. The clinicians prescribing these devices, and the patients receiving them, are going to need a vocabulary for that complexity that the field has not yet developed.

We will get there. Just slower than the press releases imply.