Neuralink and the real state of brain-computer interfaces

01 WHY THE SEALED SKULL IS STILL A HARD PROBLEM

The human brain runs on electricity. Tens of billions of neurons fire in patterns that encode every word you read, every plan you make, and every twitch of every finger. The catch is packaging: that activity happens inside a skull that is an excellent insulator, wrapped in fluid and membranes that smear signals before they reach the scalp. An EEG cap can detect the summed rumble of millions of neurons, but it cannot cleanly separate the few hundred that encode the intent to move a cursor from the background noise of everything else.

Implanted brain-computer interfaces solve the signal problem by moving the sensor inside. Instead of listening from across the skull, an electrode array sits on or in the cortex, close enough to resolve the firing of individual neurons or small groups of them. The trade is surgical: the better the signal, the more invasive the procedure, and the more of the safety burden shifts from signal processing to neurosurgery and to the long-term behavior of an implant living in tissue.

The field is not new. Researchers demonstrated cursor control with an implanted array in a human participant as early as 2006 in the BrainGate project, and clinical pilot studies have run continuously since. What changed by 2026 is that well-funded companies, not just university labs, are running human trials with new hardware, purpose-built surgical tools, and ambitions well past a dozen participants. The useful question is no longer whether implanted interfaces work, but what the actual numbers are.

02 HOW IMPLANTED ARRAYS READ NEURAL SIGNALS

An electrode placed near neurons measures small voltage fluctuations in the surrounding extracellular fluid. When a nearby neuron fires, it produces a spike, a voltage transient lasting roughly a millisecond. Recording hardware samples each channel thousands of times per second, and decoding software sorts those spikes into sources and maps their firing rates onto intent. The decoder is the real magic trick: a statistical model, calibrated while the participant watches or imagines movements, that translates firing patterns into a cursor position or a keystroke.

Three hardware philosophies dominate human work. The Utah Array, the basis of the BrainGate studies, is a grid of one hundred rigid silicon needles about the size of a small pill, placed on the motor cortex. Synchron's Stentrode takes the opposite approach: sixteen electrodes mounted on a self-expanding stent are delivered through the blood vessels to rest inside the superior sagittal sinus, next to the motor cortex, with no opening of the skull at all. Neuralink's N1 implant sits between those extremes on the invasiveness spectrum: one thousand and twenty-four electrodes distributed across sixty-four flexible threads, each thinner than a human hair, inserted by a surgical robot, with all electronics and the battery sealed beneath the skull and data sent wirelessly.

Those channel counts frame the field's core trade-off. More electrodes mean more information, but also more surgery, more surface area for the body to react against, and more decoding complexity. Figure 1 compares the publicly reported electrode counts of the three systems.

03 WHAT FIRST PATIENTS CAN ACTUALLY DO

The headline capability is cursor control. Participants in the Neuralink PRIME study have used the implant to point, click, and drag in a standard graphical interface, and the first recipient, a man paralyzed below the shoulders after a diving injury, publicly demonstrated playing chess and strategy games, browsing the web, and moderating a livestream entirely by thought. Earlier BrainGate participants achieved similar control with the Utah Array, including typing on virtual keyboards at speeds that a 2015 study measured at up to roughly eight words per minute with error correction.

Text entry rates have climbed since. Cursor-based selection has pushed into the double digits of words per minute, and laboratory systems that decode attempted speech directly from speech-production cortex have reported rates approaching conversational pace in small trials, though those research systems use different electrode coverage than the commercial implants. For the implanted systems actually in people in 2026, the honest description is a cursor and keyboard that work at a fraction of able-bodied typing speed, but at a speed that restores email, messaging, gaming, and independent computer use to people who had lost all of it.

The caveats matter. Decoders require calibration, performance drifts from day to day, and every published result comes from study cohorts measured in single digits. These are restoration milestones, not consumer products, and the distance between a compelling demonstration and a clinical tool is measured in years of longitudinal data.

04 SAFETY, REGULATORY AND CLINICAL MILESTONES

Neuralink received breakthrough device designation from the US Food and Drug Administration, and after an earlier application was rejected, cleared its investigational device exemption for a first-in-human study in 2023. The first implant followed in January 2024, the study later expanded to additional sites, and a Canadian trial was approved in late 2024. Synchron's timeline is longer: its first human implant was performed in Australia in 2019, its US trial has enrolled participants since 2021, and recipients have publicly sent texts and tweets using only the implant.

The safety record so far is thin but not alarming. In the first Neuralink participant, some of the implant's threads retracted from the tissue in the weeks after surgery, which reduced the usable signal; the team compensated by re-tuning the decoding algorithm, and the participant retained and later improved functionality. No serious unanticipated adverse events have been publicly reported in either company's cohorts, though independent researchers caution that company-reported results in small studies are not the same as the long-term safety data regulators will eventually require.

The regulatory bar for a permanent implant is high by design: devices must demonstrate safety over years, not weeks. Every company in the field now faces the same slow grind of enrolling patients, reporting adverse events, and documenting that performance holds as the implant ages in the body.

05 WHAT REMAINS UNSOLVED

The fundamental gap is bandwidth. A thousand electrodes is a landmark, but the motor cortex alone contains millions of neurons, and the control people imagine, a dexterous robotic hand or fluid speech, needs orders of magnitude more channels plus decoders that keep up. Every current system reads a tiny sample of neural activity and infers intent from it, which is why a cursor remains the flagship application rather than something more ambitious.

Longevity is the second open problem. The brain treats foreign objects as problems to be contained, and glial tissue can gradually wall off electrodes, damping their signal. Flexible thread designs and less rigid arrays aim to reduce that immune response, but nobody has multi-year data for the newest hardware, because the newest hardware has existed for only a couple of years. Power is solved well enough for now, with charging through the skull, but every design decision, thread count, sampling rate, wireless bandwidth, spends from the same battery and heat budget.

The third gap is translation, not technology. Robotic insertion surgery, implant manufacturing, and decoder support do not scale like consumer electronics, and nobody yet knows who pays: trial participants are not customers, and no payer has set a price for a restored ability to type. Cognitive augmentation for healthy users, the most quoted long-term vision, remains speculative and largely untested, and it carries a separate set of ethical questions.

06 IMPLICATIONS FOR MEDICINE AND BEYOND

The nearest-term beneficiaries are people with ALS, high spinal cord injuries, brainstem stroke, and other conditions where the cortex is intact but the connection to muscles is broken. For this group, reliable thought-driven text entry is not a convenience; it is the difference between dependence and participation. Trial results that sound incremental in a press release, another five points of typing accuracy or another hour of stable decoding, compound into qualitatively different lives for the people who receive them.

If safety and durability data hold, implanted interfaces follow the path of other implantable electronics, pacemakers and cochlear implants, from experimental procedure to standard of care for a defined patient group. That path runs through surgical capacity, since each Neuralink procedure uses a custom robot and trained team, and through reimbursement, because a therapy nobody pays for helps nobody.

The spillover value is already real. Flexible electrode manufacturing, implantable wireless electronics, robotic microsurgery, and neural signal decoding are general-purpose technologies. Improvements aimed at interfaces are migrating into adjacent fields, from deep brain stimulation to prosthetics control, whether or not mass-market neural implants ever arrive.

07 THE LONG ARC FROM LAB CURIOSITY TO CLINICAL TOOL

The through-line runs from the first human EEG recordings in the 1920s, through early laboratory interfaces in animals in the 1990s, to the 2006 BrainGate result that let a paralyzed participant move a cursor, to Synchron's endovascular implant in 2019, to the first wireless high-channel-count implant in a human in 2024. Each step traded some safety margin for signal quality, and each hardware generation raised the channel count available to a patient.

Money has followed the milestone count. Neuralink's reported funding rounds grew from two hundred five million dollars in 2021 to three hundred twenty-three million in 2023 to a reported six hundred fifty million in 2025, on top of earlier rounds and investment. Figure 2 shows the reported round sizes; they measure investor patience more than clinical progress, but they pay for the surgical robots, manufacturing, and trials that university grants never could.

The realistic near future is more patients, more sites, and longer follow-up rather than a consumer launch. The signals worth watching are unglamorous: how many participants carry working implants at three years, how much decoding accuracy degrades over time, whether adverse events stay rare and manageable, and whether the second generation of hardware can be implanted with less robotic theater. If those trends hold, brain-computer interfaces will do for severe paralysis roughly what cochlear implants did for profound deafness, and on a similar timescale.