Neuralink

Neuralink is a privately held American company developing an implanted Brain–computer interface consisting of a coin-sized wireless processor and a set of thin polymer threads carrying electrodes, inserted into cortex by a purpose-built surgical robot. It was founded in 2016 by a group including Elon Musk, who has been its public face and principal funder. The company began implanting the device in humans in January 2024 under an FDA investigational device exemption, and in January 2026 it said 21 people were enrolled in its trials worldwide.¹ As of October 2026 its devices remain investigational and unapproved by any regulator.²

Neuralink's technical contribution lies largely in engineering and surgery rather than neuroscience: flexible electrode threads that move with the brain, automated insertion, high channel counts in a sealed wireless package, and a hermetic implant with no percutaneous connector. Its scientific results to date have been communicated mainly through livestreams, company blog posts and social media rather than peer-reviewed publication, the criticism academic BCI researchers most often level at it.

§Technology

The N1 implant houses a battery, wireless radio, and signal-processing electronics in a package that sits in a recess milled into the skull, flush with the surrounding bone. From it extend 64 polymer threads bearing 16 electrodes each, 1,024 electrodes in total.³ Each thread is narrower than a human hair and far more compliant than the rigid silicon shanks of a Utah array. The design is meant to reduce the micromotion-driven tissue damage that causes glial encapsulation and signal loss in stiff arrays.

Because the threads are too floppy to insert by hand, the R1 robot uses a needle and optical guidance to place each one individually, steering around surface vasculature. The company described the architecture in a 2019 white paper based on rodent work.⁴ The implant processes the electrode signals and sends them wirelessly to a Neuralink application on an external device such as a computer, which runs the decoder; the battery is recharged inductively through the scalp.³ The company says an hour of charging gives 4-8 hours of use, depending on the implant version.² Users control a cursor by attempting or imagining movement, and a periodically recalibrated decoder learns the mapping. This is the same basic paradigm of Neural decoding established in academic laboratories two decades earlier.⁵⁶

The hardware is changing. In January 2026 the company said it was moving from roughly 1,000 to 3,000 electrodes per implant, testing mechanical features to keep threads anchored, and investigating inserting threads through the dura mater rather than removing it.¹ In June 2026 it reported its first procedure in which the threads went straight through an intact dura into cortex.⁷

The robot is being redesigned as well. In a June 2025 company presentation a Neuralink engineer said the R1 that implanted the first seven participants needed 17 seconds per thread at best, and presented a next-generation robot that he said inserts each thread in 1.5 seconds, can place threads more than 50 mm below the brain surface, and is compatible with more than 99% of the population.⁸ In a recruiting talk published in December 2025, co-founder DJ Seo said the robot in clinical use had implanted 13 people, placing threads about 4 mm deep.⁹ An April 2026 company video showed the new design's eight cameras and an optical coherence tomography sensor that sees beneath the tissue surface; its engineers said it moves in five axes so it can reach more regions of the brain, and that it was "not in its final form".¹⁰ All of these figures are Neuralink's own.

As used in its trials so far, the device records but does not stimulate, which puts it in a different category from the implanted neurotechnologies already in routine clinical use, such as the Cochlear implant and Deep brain stimulation, which write signals in rather than reading them out. Stimulation is required for the visual programme, but the safety profile of chronic cortical stimulation is less well characterized than that of chronic recording.

1. 2016
   Company founded
   Registered in California by Elon Musk and a group of neuroscientists and engineers; publicly reported in early 2017.
2. 2019
   Threads and robot unveiled
   A launch event and accompanying white paper describe polymer electrode threads, robotic insertion, and rodent recordings.
3. 2020
   Pig demonstration
   A livestream shows live neural recordings from an implanted pig, demonstrating a sealed wireless device rather than any new decoding capability.
4. 2021
   Macaque cursor control
   A monkey plays a cursor-based video game using the implant, replicating a capability shown in primate laboratories years earlier.
5. 2023
   Human trial cleared
   The FDA grants an investigational device exemption in May; the PRIME study opens recruitment in September for people with quadriplegia.
6. 2024
   First human implants
   The first participant receives the device in January; the company later discloses that many threads retracted from the cortex. Blindsight receives an FDA breakthrough device designation; CAN-PRIME and the CONVOY robotic-arm study open.
7. 2025
   Speech and expansion
   Speech restoration receives a breakthrough device designation; studies open in the United Arab Emirates and Great Britain; the VOICE speech study starts in October. In June the company shows a faster next-generation surgical robot.
8. 2026
   Speech data and Blindsight wait
   The company reports 21 enrolled participants in January and early speech results at a conference in October; Musk says Blindsight awaits regulatory approval.

§Human trials

§Device control: PRIME and Telepathy

The PRIME study (Precise Robotically Implanted Brain-Computer Interface) runs under an FDA investigational device exemption granted in May 2023 and enrols adults with quadriplegia from cervical spinal cord injury or amyotrophic lateral sclerosis.¹¹ The first participant, Noland Arbaugh, received an implant at Barrow Neurological Institute in Phoenix in January 2024.³ Within weeks he was using it to move a cursor, play chess and video games, and operate a computer for long stretches daily; the company reported research sessions of up to eight hours on weekdays.¹²

In the weeks after that surgery a substantial fraction of the threads retracted from the cortex, reducing the number of channels recording usable signal. Neuralink disclosed this in a blog post several months later and reported recovering performance through changes to signal processing and decoding rather than reimplantation.¹² For the second participant the company said it reduced brain motion during surgery and narrowed the gap between the implant and the brain surface, and it reported no thread retraction in that participant.¹³ By January 2026 it said 18 of the 20 participants implanted after these changes showed higher signal quality, while performance still varied with anatomy and with disease stage.¹ Thread retraction is one instance of a general problem for every penetrating array, which is that the brain moves, pulses, and reacts to foreign material.

Neuralink calls the device-control application Telepathy. On the company's own description it is a mind-controlled mouse: participants move a cursor and issue left, right and middle clicks by attempting or imagining hand movements, mainly on a supplied MacBook, and keeping control may take up to a few hours of calibration a week. Keyboard, game-console, wheelchair and robotic-arm control are experimental features not available to every participant.² The company reports that several participants have reached and sometimes exceeded the 8-10 bits per second it gives for able-bodied mouse use, that a "ten finger keyboard for the mind" decoded from imagined finger movements has reached 40 words per minute, and that its trials have had no serious device-related adverse events.¹ All three figures are self-reported.

The same protocol now runs in other countries: CAN-PRIME at University Health Network in Toronto from November 2024,¹⁴¹⁵ UAE-PRIME at Cleveland Clinic Abu Dhabi from May 2025,¹⁶ and GB-PRIME at University College London Hospitals and Newcastle from July 2025.¹⁷ A companion study, CONVOY, opened in November 2024 to let PRIME participants control an investigational assistive robotic arm,¹⁸ moving the device from computer access toward the broader ambitions of Neuroprosthetics; in January 2026 the company described a participant using a robotic arm to feed himself.¹

§Speech restoration: VOICE

The FDA granted Neuralink a breakthrough device designation for speech restoration in May 2025.¹⁹ Its speech study, VOICE, began in October 2025 at UT Southwestern Medical Center under principal investigator Nader Pouratian, with an estimated enrolment of six adults with severe speech impairment and impaired arm function.²⁰ Instead of the hand area used for Telepathy, the threads go into speech motor cortex, and a decoder turns attempted or imagined speech into text or into synthesized speech in a re-creation of the participant's own former voice.²¹ This is the paradigm demonstrated earlier by academic groups and covered under Speech neuroprosthesis.

The company says its median speech participant decodes about 50 words per minute, against roughly 150 for natural conversation.²¹ On 28 October 2026 it reported that one participant with ALS reached 75 words per minute at a 5.8% word error rate, on a 125,000-word vocabulary, in the first week of use. Those figures were presented as an abstract at the annual Northeast ALS Consortium meeting rather than in a full peer-reviewed paper.²² A day later Neuralink publicized a 72-hour engineering visit to a VOICE participant that produced prototypes including automatic detection of when the user is trying to speak, an integration of the company's existing speech app, Bloom, with iPhone calls, and a feature that starts a Grok computer-use agent when the user thinks of saying "Hey Neuralink".²³

§Blindsight

Blindsight is Neuralink's visual prosthesis: an implant meant to produce visual percepts by stimulating visual cortex directly, bypassing the eyes and optic nerve. The FDA gave it a breakthrough device designation in September 2024,²⁴ a status that speeds contact with the agency but is not a finding of safety or effectiveness. Musk wrote at the time that it would let people who had lost both eyes and optic nerves see and, "provided the visual cortex is intact", people blind from birth, while warning that vision would at first be low resolution, "like Atari graphics". A vision scientist interviewed by IEEE Spectrum that month pointed out that no preclinical evaluation of the device had been published.²⁵

As of October 2026 no Blindsight study had been registered on ClinicalTrials.gov or announced by the company. Neuralink's visual-prosthesis page asked people to join its patient registry to be told when U.S. trials begin,²⁶ and on 30 October 2026 Musk wrote that the device was "ready from an engineering standpoint" and "Just waiting on regulatory approval".²⁷ The same post said it would let "even those who are completely blind or have never seen" see, dropping the earlier proviso. By that proviso the approach depends on a working visual cortex, so it cannot help blindness caused by damage there. Cortical prostheses have also been tested mainly in people who lost their sight later in life, and whether one can usefully restore vision to someone born blind, who has never used the visual cortex as a sighted person does, remains an open question.²⁸ Experts in 2024 expected even the best cortical implant buildable with current technology to produce nothing like normal vision, at most an aid alongside a cane or guide dog, though one added that it might still transform blind people's lives.²⁵ Earlier attempts are covered under Cortical visual prosthesis, and the resolution limits that constrained them, described under Retinal implants and visual prostheses, apply with equal force here.

Company reporting, not trial reporting

Participant counts, performance figures, and adverse-event descriptions for Neuralink's studies have come predominantly from company statements, livestreams and social media. In October 2025 Bloomberg reported that a paper on the first three PRIME participants had been submitted to the New England Journal of Medicine.²⁹ As of October 2026 no peer-reviewed clinical report from any Neuralink trial has been published, and the speech results exist only as a conference abstract, so durability, complication rates, and decoder performance cannot yet be assessed independently.

§Funding and structure

Neuralink is funded by venture capital and by Musk personally. It announced a $280 million Series D led by Founders Fund in August 2023³⁰ and a $650 million Series E in June 2025, with investors including ARK Invest, Founders Fund, G42, Sequoia Capital and Thrive Capital.³¹ The company did not publish a valuation; Semafor reported that the Series E valued it at about $9 billion before the new money.³² That figure does not rest on product revenue, since Neuralink has no approved product. The capital advantage is real. Implant development, robot engineering, animal work, and multi-country regulatory submissions cost more than academic consortia such as BrainGate can spend. At the end of 2025 Musk said the company would start "high-volume production" of its devices and move to an automated surgical procedure in 2026,³³ a forecast that depends on regulators as much as on engineering.

The founding team has largely dispersed. Several co-founders left within the first few years, and one went on to co-found Precision Neuroscience, which pursues a thin-film surface approach closer to Electrocorticography interfaces than to penetrating threads. In December 2025 the company hired David McMullen, who had previously worked at DARPA, the NIH and the FDA, to lead its medical affairs.³⁴ Competitors now span the invasiveness spectrum, including Synchron with its endovascular Stentrode, Paradromics with a high-channel penetrating array, and Blackrock Neurotech.

§Reception

Assessments divide along a consistent line. Engineers credit the company with solving practical problems the field had left unsolved: a fully implanted wireless device with no skull-mounted pedestal, an insertion robot that makes hundreds of precise placements, and a manufacturing process for polymer electrode arrays. Those problems stand between a laboratory demonstration and a product.

Neuroscientists are more sceptical of the framing. The cursor control demonstrated in 2024 was qualitatively similar to results published from academic implants more than a decade earlier, and the product name "Telepathy" describes controlling a computer with a decoder rather than thought transmission, an overstatement of the kind examined under Brain-to-brain interfaces. In October 2026 Musk described Telepathy as enabling "high-speed communication with computers or humans simply by thinking" and predicted that its output bandwidth "will vastly exceed the fastest human speech or typing in a few years".³⁵ Any communication with other people runs through a computer, speech is handled by the separate VOICE study, and the highest speech rate the company has reported is about half the conversational rate it cites itself.²²²¹ Claims that the device will one day enable symbiosis with artificial intelligence, discussed as Human–AI merger, are untested and rest on a bandwidth argument that the low-dimensional structure of cortical activity does not obviously support.

The company has also faced sustained criticism over animal research. Reuters reported in 2022 that federal investigators were examining animal-welfare practices following internal complaints.³⁶ In 2023 the Department of Transportation opened an inquiry into the packaging and transport of implants removed from monkeys' brains, after the Physicians Committee for Responsible Medicine, an advocacy group, obtained records of primate experiments Neuralink had run with the University of California, Davis.³⁷ Neuralink has denied wrongdoing and published its own account of its animal-care programme.

A third strand of criticism concerns data. A device that streams cortical activity to a phone creates a record of neural signals held by a private company, which raises the questions taken up under Mental privacy. A cursor decoder does not read thoughts, but the underlying recordings support inferences the user never intended to make available.

§Outlook

The questions that will determine whether Neuralink matters are practical ones: whether threads stay put over five years, what fraction of participants develop infection or require explant, whether a decoder can run for weeks without a technician, and whether the device works in a home, unsupervised, on a bad day. Every trial Neuralink has registered so far is described as a feasibility study.²⁰¹⁶ Elsewhere, China's medical-products regulator approved Neuracle's NEO in March 2026, an implant placed outside the dura that drives a pneumatic glove, in what was widely reported as the first commercial approval of an implanted BCI.³⁸ Neuralink's penetrating threads have no such approval, and the company has not announced the larger pivotal trial a marketing application would normally need.

A separate question is what happens if the company stops. Implanted devices create obligations that outlast product lines, an issue the emerging Neurorights frameworks have begun to name but have not yet resolved.

§See also

• Brain–computer interface
• Speech neuroprosthesis
• Cortical visual prosthesis
• Stentrode
• Utah array
• Neural decoding
• Mental privacy
• Synchron

§References