Neuralink Brain Chip Explained: What the Hardware Actually Does Inside the Human Brain
Most people focus on the headlines. Engineers focus on the hardware. Here’s how Neuralink’s brain-computer interface works, from ultra-thin electrodes and custom chips to wireless communication and real-time neural signal processing.
Table of Contents
Let me start somewhere in the middle, because that’s where the real story is.
Roughly 85% of Noland Arbaugh’s electrode threads pulled back from his brain tissue in the weeks after his implant went in. Neuralink itself disclosed only that “a number” of threads had “retracted”; the roughly 85% estimate came from Arbaugh, in an interview with The Wall Street Journal. Either way, it wasn’t small. Most of the threads. And yet — the cursor still worked. Neuralink disclosed the issue publicly, adjusted its decoding algorithm to compensate, and performance recovered. That one detail — hardware degrading, software adapting, and the system still functioning — tells you more about what this tech actually is than any launch event video ever could.
Now let me back up.
January 28, 2024. Arbaugh — 29 years old, quadriplegic, hadn’t moved his hands in eight years after a diving accident — sat in front of a screen and moved a cursor. With his thoughts. No hands. No joystick. Nothing. Just intent, picked up by a chip the size of a coin beneath his skull, fired wirelessly to a receiver nearby.
That moment wasn’t magic. It was hardware. This is the Neuralink brain chip explained from the inside — not the press release version, but the actual hardware version.
I’ve been in PCB and electronics manufacturing for over 20 years. I’ve worked on complex board-level systems with some tight specs. And I’ll say this plainly: what Neuralink pulled off at the chip and system level is one of the harder things I’ve seen done. Not because the concept is new — BCIs have been around for decades. Because the physics is brutal and the substrate, a living human brain, actively fights back.
If you want to know what’s actually happening inside that skull when the chip runs — keep reading.
The Big Picture: What Is a BCI, Really?
Here’s what most explainers get wrong right out of the gate. A brain-computer interface (BCI) doesn’t “read minds.” It reads voltage.
Your neurons talk by firing tiny electrochemical spikes called action potentials. That’s it. A BCI’s job is to catch those spikes, figure out what they mean, and turn them into a digital command fast enough to feel real. No thoughts. No feelings. Just electrical events, detected and decoded.
I think people hear “brain chip” and picture something like a USB drive plugged into a skull. It’s nothing like that. It’s a signal-capture problem operating at tolerances that would make most fab engineers sweat — in a substrate that moves, bleeds, and tries to wall off anything it doesn’t recognize as self.
Neuralink’s system — branded Telepathy for its first consumer use — ranks among the highest channel-count intracortical BCI systems deployed in humans to date. Not the only player (Blackrock Neurotech and Synchron are both operating clinically), but by raw electrode count, it’s near the top.
Three hardware pieces run this:
- The N1 Implant — the actual chip that lives inside your skull, handles all the signal capture and processing, and is the piece nobody adequately explains
- The R1 Robot
- The N1 User App and external receiver, which is the wireless decode layer that turns compressed neural data into something a laptop can act on
Each one is its own hard engineering problem. Don’t let the clean product photos fool you.
The Neuralink N1 Implant: What’s Actually Inside Your Skull
Size and Placement
Let’s start with a number most coverage buries: 23mm × 8mm. That’s the N1 Implant’s size per PRIME Study docs (NCT06429735). Roughly the size of a large coin. Sits flush with the skull. No external bump, no antenna, nothing you’d see or feel from the outside.
Here’s something people consistently get backward — the hermetic seal on this thing protects the electronics from the body, not the other way around. Cerebrospinal fluid, ion gradients, enzymatic activity — all corrosive to unprotected silicon over time. The enclosure resists those conditions and keeps the chip alive in an environment that would eat unshielded electronics fast. If you want to understand how far implantable electronics have come to make this possible, I covered the broader hardware engineering story in Implantable Medical Devices: 7 Breakthrough Truths That Will Change How You Think About Healthcare.
Power? Small internal battery. Charged wirelessly through the scalp via inductive charger. No cables, no skin breach. (Honestly, that part is elegant. One of the cleaner design choices in the whole system.)
The 1,024 Electrodes: Here’s Where It Gets Real
I’ve noticed that coverage of this technology tends to skim the electrode architecture and move straight to the “paralyzed man controls computer” headline. Understandable. But you miss the whole engineering story if you do that.
The N1 carries 1,024 total electrodes, per the PRIME Study brochure. Thread config has evolved — Arbaugh’s original January 2024 implant used 64 threads, 16 electrodes each. Multiple outlets including TechTimes (May 2026), RobotCentral (January 2026), and Neurapod all report later patients got 128 thinner threads with 8 electrodes each, same 1,024 total. No peer-reviewed spec for that change yet, but it lines up across multiple sources and directly addresses the mechanical problem that caused Arbaugh’s thread retraction.
For scale: deep brain stimulation devices for Parkinson’s — a different device class built for stimulation, not recording — typically run around 10 electrodes. Neuralink’s recording array is substantially more. That said, I know what you’re thinking — channel count isn’t the whole story. You’re right. Electrode geometry, impedance, and placement depth all affect actual signal quality independently. Channel count is the headline. Signal quality is the actual measure.
Why does count matter then? It’s pure sampling math. More electrodes, more neurons sampled at once. More neurons, richer signal. Richer signal, better decode of intent. My read — and this is an engineering inference, not a documented threshold — is that 1,024 electrodes gives you enough redundancy to capture the motor cortex’s population-level activity without any single thread’s dropout collapsing the whole read. Arbaugh’s case is actually a decent stress test of that idea: even after losing the vast majority of his threads, the system kept functioning, just with software picking up the slack.
The N1’s design target, per Neuralink’s 2019 white paper (Musk & Neuralink, J Med Internet Res, PMC6914248), is placing electrodes within 60 microns of target neurons — not a hard universal law, but the specific proximity goal for this system, at which reliable action potential detection works. Miss it, and the signal drowns.
The N1 SoC: The Part Everyone Skips (And It Drives Me Crazy)
The chip itself is Neuralink’s N1 System-on-Chip (SoC). In my experience, this is the component that gets maybe one paragraph in mainstream coverage — if that. And honestly? I was going to skip a full section on it myself when I first started putting this together, because I figured readers would glaze over. Then I looked at the actual numbers from the primary source and changed my mind immediately.
If you only read one section of this piece, make it this one. The electrode count is the flashy part. The SoC is where the actual engineering lives, and the fact that nobody talks about it is just — look, it’s a pet peeve. I’ve sat through too many tech briefings where the chip architecture gets one slide, and the product renders get fifteen.
The SoC takes raw analog neural signals, amplifies them, digitizes them, and processes them before anything goes over the air. That step is not optional. Raw broadband neural data is enormous. Push it all wirelessly, and you kill the battery in minutes and flood the link.
Look, I know quoting directly from a white paper sounds incredibly dry, but bear with me — these are verbatim numbers from Neuralink’s 2019 prototype-era publication (DOI: 10.2196/16194, PMC6914248), and they’re worth reading slowly: “The on-chip ADC samples at 19.3 kHz with 10-bit resolution. Each analog pixel consumes 5.2 µW, and the whole ASIC consumes approximately 6 mW, including the clock drivers.”
Take a second with that. And one honest caveat before we go further: these numbers describe the architecture in that 2019 paper — the same prototype-era system that, as I’ll get into later, used a 12-ASIC, 3,072-electrode benchtop design distinct from the single-chip, 1,024-electrode N1 that’s actually implanted in patients today. Nobody has publicly confirmed these exact figures for the current clinical chip. They may be identical. They may not. What’s certain is the architecture the industry learned from — not necessarily every number carried forward unchanged.
19.3 kHz is 19,300 voltage reads per second per channel. Across every active electrode at once. That’s a flood of raw data — far too much to push wirelessly without wrecking the power budget. So the chip runs an on-chip spike detection algorithm first.
Per Neuralink’s 2019 launch event (engineer DJ Seo’s presentation — company-stated, not peer-reviewed, so keep that in mind), the algorithm hits more than 200× data compression in around 900 nanoseconds. On the other hand, maybe I’m too quick to take those exact figures at face value since they haven’t been independently verified in the peer-reviewed literature. Fair caveat.
But here’s the thing — given the channel count, the sampling rate, the wireless bandwidth constraints, and the power budget all discussed above, substantial on-chip compression is necessary, whatever the exact ratio turns out to be. You cannot run a wireless intracranial implant on a real battery while streaming uncompressed broadband neural data at this scale. The compression isn’t a feature. It’s what makes the device physically possible.
The ~6 mW figure is the other constraint that doesn’t get enough airtime — and it’s worth being precise about what it actually covers. That’s the ASIC’s power draw specifically: amplification, digitization, spike detection, compression, bidirectional stimulation. It is not a confirmed total-device power figure. The full implant also includes the battery, the wireless charging receiver, and RF transmission electronics — none of which are covered by that 6 mW number in any source I’ve found. Neuralink hasn’t published a total N1 power budget publicly.
What is well-established, as a general principle across implantable neural devices, is that intracranial implants operate under a genuinely tight thermal safety margin — tissue heating near the brain is a serious, well-documented biomedical design constraint, generally discussed in the neural-engineering literature in terms of fractions of a degree Celsius of allowable temperature rise. That’s a field-wide constraint, not a disclosed N1-specific spec — Neuralink hasn’t published its exact thermal budget for the N1. But it’s the reason power draw at every level of this system, ASIC included, isn’t a design nicety. It’s a hard patient safety consideration. It’s also worth noting that the brain itself runs on roughly 20 watts total — which is exactly why researchers are studying brain-inspired chip architectures as a model for low-power AI. I covered that angle in depth here: AI Is About to Hit an Energy Wall: How Brain-Inspired Chips Could Help Computing Get Past It.
One more thing worth stating: the N1’s nominal electrode count is 1,024. The functional count — electrodes actually recording useably — is lower and degrades over time as glial scarring kicks in and threads retract. Arbaugh lost roughly 85% of recording threads**, by his own estimate, in the weeks following surgery. Nominal versus functional electrode count is not a minor asterisk. It’s central to any real read of performance.
The R1 Robot: This Cannot Be Done by Hand
Jump straight to the constraint: thread diameters of 4 to 6 µm, per the 2019 white paper (PMC6914248). That’s 1/10th to 1/25th the width of a human hair. Here’s my read as someone who’s spent two decades around high-precision manufacturing: at that scale, hitting the speed, repeatability, and real-time vascular avoidance the procedure needs isn’t something a human hand can reliably deliver, case after case. Not difficult. Not inadvisable. Not something I’d bet on at scale, based on the numbers involved — no cited clinical study has actually ruled out manual insertion, but the geometry makes the conclusion pretty obvious.
The R1 Robot uses an ultrafine needle — much thinner than a human hair — to grasp, insert, and release each electrode thread, placing threads into a moving brain while avoiding blood vessels (per Neuralink’s 2022 progress update, as reported by New Atlas). The newer version of the robot Neuralink unveiled in 2026 adds a five-axis system with cameras and OCT scanners for real-time navigation (Interesting Engineering, April 2026). Nick a vessel and you get bleeding, inflammation, signal loss. Not acceptable.
In my experience working in high-tolerance manufacturing, the first instinct is to benchmark this against PCB fab tolerances — sub-50 µm HDI work, laser microvias, that world. And I get why. I’ve covered how laminate selection and signal integrity tolerances play out in real AI hardware builds in Choosing the Right High-Speed Laminate for AI Hardware: Megtron 6 vs. Megtron 8 vs. Tachyon 100G — the comparison here is intentionally high-level.
But I think that comparison undersells what’s actually hard here. PCB fabrication runs in controlled, rigid, stable material where you set every parameter. The R1 works in living tissue that pulses with every heartbeat, varies in stiffness between patients and regions, and has vasculature you cannot touch.
The tolerance number is similar. The environment is incomparable. Biological variance is the real engineering challenge — the tolerance is just the minimum bar to clear.
The procedure: craniotomy, general anesthesia, R1 handles the thread insertion phase under neurosurgeon supervision. Multiple threads per minute. Motor cortex targeting, per the PRIME Study brochure — the region that controls movement intent. That’s the source of Arbaugh’s cursor signals.
The Problem Nobody Has Cracked
There is a biological response to every foreign object placed in a brain. Every single one. Doesn’t matter how thin the threads are, how biocompatible the coating is, or how careful the robot is. The brain’s immune system wakes up.
Reactive astrocytes and microglia — the brain’s cleanup cells — detect the threads and start wrapping the electrode tips. That wrapping raises impedance at the electrode-tissue contact point and kills signal quality over time. This is glial scarring.
It’s widely regarded, across the neural-engineering field, as one of the central long-term challenges for every implantable BCI — not just Neuralink’s. Polikov, Tresco, and Reichert mapped the full cellular cascade in their foundational 2005 Journal of Neuroscience Methods paper. The field has been engineering around it since. Nobody has solved it. This biological boundary — where silicon meets living neural tissue — is also the central tension in a broader trend I wrote about here: The Next Computer Isn’t Made of Silicon: How Living Cells and Circuit Boards Could Power the Future of Computing.
Now, here’s where it gets weird — and also kind of impressive, depending on how you look at it.
Back to Arbaugh: roughly 85% of his threads retracted — “retracted” is Neuralink’s word, the 85% estimate is Arbaugh’s — in the weeks following surgery. What that meant electrode-by-electrode isn’t fully documented; not every retracted thread necessarily went completely silent, and the public record doesn’t map retraction one-to-one onto electrode failure. What is documented is that Neuralink adjusted its decoding algorithm — credit where it’s due, the cursor kept working — and performance recovered. The hardware side had genuinely degraded, and however you characterize the mechanism precisely, it’s a total mess to deal with from a long-term reliability standpoint, even though the short-term outcome, thanks to the software fix, was fine.
I think that’s worth sitting with. On the other hand, a system that takes a significant hardware hit and still functions through adaptive software is arguably a better real-world outcome than something brittle that fails hard when one thread goes down. I genuinely go back and forth on how to frame it, and I’ve rewritten that sentence about four times while putting this together.
Thinner threads reduce the mechanical disruption that starts the glial cascade. Less stiffness, less tissue damage, less immune response. Helps. Doesn’t fix it. Nothing does yet.
Wireless Transmission and the Security Gap
Once the SoC compresses the data, it fires wirelessly to an external receiver, which feeds the N1 User App, which decodes patterns into commands — cursor movement, scrolling, clicks. For future CONVOY trial participants (robotic arm feasibility study, approved in late 2024), it’ll mean controlling a physical arm.
By the way — this next section is the one I debated cutting entirely because I wasn’t sure it belonged in a hardware explainer. Then I remembered that the whole point of writing about hardware is to be honest about what we don’t know, not just what we do.
Here’s the thing I want to flag directly for any engineer reading this, because it’s not getting enough coverage.
As of June 2026, Neuralink has published nothing on the N1’s wireless security architecture. No SBOM. No threat model. No disclosed encryption or auth protocol.
Under Section 524B of the FD&C Act and FDA’s updated cybersecurity guidance (revised February 2026), manufacturers submitting a cyber device through 510(k), PMA, PDP, De Novo, or HDE pathways now need a Software Bill of Materials, a Secure Product Development Framework, and documented threat modeling as part of that submission.
Here’s the nuance worth getting right: the PRIME Study currently runs under an Investigational Device Exemption, or IDE — a clinical-investigation pathway, not one of the marketing-authorization pathways 524B’s documentation requirements are tied to. So Neuralink isn’t necessarily required to produce that full cybersecurity package right now, during the trial phase. That requirement becomes directly relevant if and when Neuralink files a marketing submission — a PMA, most likely, given the device class — down the road. An undocumented wireless implant isn’t an active submission violation today. But it’s a real forward-looking regulatory consideration for whatever comes after PRIME.
Fair enough, but I want to be careful here — absence of public documentation doesn’t mean the work hasn’t been done internally. Halperin et al. (2008, IEEE Symposium on Security and Privacy) showed real exploitable gaps in cardiac implantable devices before the field took this seriously. That paper is 18 years old.
Don’t sweat the 2008 date — the point is that wireless implant security is an established engineering discipline with documented precedent, and the public record for the N1 has none of it. Absence of public documentation is not evidence of either a problem or safety — it’s just an absence.
If you want to understand how deep the undocumented hardware security problem runs across medical and embedded devices, I’ve covered two related angles on this site: the cryptographic threat to embedded hardware in Your Embedded Hardware Has a Quantum Problem Nobody Is Talking About, and the RF attack surface problem in automotive systems in Self-Driving Cars Have a Hardware Security Problem Nobody Wants to Talk About. The N1’s undocumented wireless link is a different device class — but the same underlying problem.
Where Things Stand Right Now
Multiple outlets have reported additional PRIME Study participants across US, UK, Canada, and UAE. Neuralink’s own public materials say “multiple participants” — no count, no peer-reviewed update. Media “20+” figures are out there but unverified at the primary source level. Don’t sweat the exact number — what matters is the trial is running and expanding geographically.
May 2026: Neuralink announced the next-gen robot is aimed at brain regions beyond the motor cortex. Parkinson’s, refractory epilepsy, treatment-resistant depression all named as future targets. No published spec on scope yet.
The roadmap beyond that — Neuralink’s 2019 vision described a four-chip architecture, potentially thousands of electrode channels working together. Not confirmed for human deployment. Timeline not available.
The gap between the 2019 prototype (3,072 electrodes, 12 ASICs, 96 threads) and the deployed N1 (1,024 electrodes, one chip) is the gap between what you can demo and what you can actually make reliable enough to put in a person. That gap is the engineering project. It’s a toss-up whether the prototype ambition or the shipped hardware tells you more about where this field really is — I’d argue the shipped hardware every time.
What This Hardware Actually Proves
Noland Arbaugh moved a cursor.
Simple sentence. Not simple hardware. What Neuralink’s published architecture showed — and what the 2019 specs back up — is that you can put over a thousand electrodes in a living human brain, sample neural signals at rates as high as 19.3 kHz in the architecture that established this approach, compress that data more than 200-fold on-chip in under a microsecond, and send it wirelessly to an app that turns thought into action.
I was looking back over this data earlier while writing this section, and it hit me that what makes the Arbaugh moment remarkable isn’t any single spec — it’s that all of them held at the same time.
The electrode placement was precise enough; the ASIC had to operate within a tightly constrained intracranial thermal envelope — Neuralink hasn’t publicly disclosed the N1’s exact thermal budget or measured margin — the compression algorithm was fast enough that the wireless link didn’t choke, and somehow the decoding software was flexible enough to keep working even when roughly 85% of the electrode threads had already retracted.
That’s not one engineering win. That’s five or six of them stacking simultaneously, in a human skull, on the first real try.
Not science fiction. A clinical trial with a registration number.
The headlines chase the vision. Engineers — and honestly, anyone who actually wants to understand what’s happening — should follow the chip.
Correction (October 1, 2026): An earlier version of this article attributed the roughly 85% thread-retraction figure to Neuralink and suggested it may have occurred up to three months after surgery. Neuralink disclosed only that “a number” of threads retracted in the weeks after surgery; the ~85% estimate came from Noland Arbaugh. This version also corrects the CONVOY study’s approval date to late 2024, clarifies which surgical-robot capabilities belong to Neuralink’s newer 2026 robot, and updates the sources.
About the Author
Imran Valiani | Sales Director, PCB Electronics Manufacturing
20+ years working with major Bay Area and global tech clients. Founder of Silicon to Software, where I write about the hardware layer — PCB fab, AI gear, autonomous systems, and cyber — the stuff most tech writers have never touched. Literally.
Follow: X @SiToSoftware | LinkedIn
For more analysis of embedded hardware, electronics, and system-level engineering constraints, explore the Embedded Systems & Electronics Engineering Resource Hub.
This article was developed with AI assistance and edited, fact-checked, and reviewed by the author. See my full AI disclosure.
Sources
- Musk, E. & Neuralink — “An Integrated Brain-Machine Interface Platform With Thousands of Channels.” Journal of Medical Internet Research, October 2019. DOI: 10.2196/16194. PubMed Central: PMC6914248
- Neuralink PRIME Study Brochure — Precise Robotically IMplanted Brain-Computer InterfacE. Clinical trial registration: NCT06429735. neuralink.com/pdfs/PRIME-Study-Brochure.pdf
- ClinicalTrials.gov — PRIME Study registration, NCT06429735. clinicaltrials.gov/study/NCT06429735
- Polikov, V.S., Tresco, P.A. & Reichert, W.M. — “Response of brain tissue to chronically implanted neural electrodes.” Journal of Neuroscience Methods, 2005. Foundational peer-reviewed study on glial scarring in implantable BCI systems.
- Halperin, D. et al. — “Pacemakers and Implantable Cardiac Defibrillators: Software Radio Attacks and Zero-Power Defenses.” IEEE Symposium on Security and Privacy, 2008. Peer-reviewed precedent for wireless implantable device security vulnerabilities.
- U.S. Food and Drug Administration — Cybersecurity in Medical Devices: Quality System Considerations and Content of Premarket Submissions. Guidance document, updated February 2026. fda.gov/regulatory-information/search-fda-guidance-documents
- U.S. Food and Drug Administration — Section 524B, FD&C Act: Cybersecurity Requirements for Cyber Devices. FAQ and portal: fda.gov/medical-devices/digital-health-center-excellence/cybersecurity
- New Atlas — “Neuralink gives a wide-ranging update on its brain chip progress,” December 2, 2022. newatlas.com
- Interesting Engineering — “Neuralink unveils surgical robot to fully automate brain chip implants,” April 30, 2026. interestingengineering.com
- TechTimes — “Neuralink BCI After 28 Months: Arbaugh Tells Robotics Summit What Published Research Cannot,” May 28, 2026. techtimes.com
- Neuralink — “PRIME Study Progress Update — User Experience,” May 8, 2024 (thread-retraction disclosure). neuralink.com
- NBC News — “Implant from Elon Musk’s Neuralink suffers setback as threads retract from patient’s brain,” May 2024. nbcnews.com
- Popular Science — “85% of Neuralink implant wires are already detached, says patient” (reporting Arbaugh’s interview with The Wall Street Journal), May 21, 2024. popsci.com
- Neuralink — “CONVOY Study Launch,” November 2024. neuralink.com
- Neuralink — Official technology and trial pages. neuralink.com/trials/device-control
