The Brain Chip That Streams Thoughts in Real Time: What Columbia & Stanford Just Built
Now I have everything I need. Let me write this article.TITLE: The Brain Chip That Streams Thoughts in Real Time: What Columbia & Stanford Just Built SUBTITLE: A paper-thin silicon implant smaller than a fingernail just rewrote what brain-computer interfaces can do — and it may make the current state of the art look like dial-up.
Picture a chip the thickness of a human hair, roughly the size of your fingernail, sliding silently into the gap between your skull and your brain. No bulky canister. No tangle of wires threading through your cortex. Just a single wafer of silicon that listens to 65,536 electrodes simultaneously and fires neural data over a wireless link at 100 megabits per second — about 100 times faster than anything else currently on the market. That is the Biological Interface System to Cortex, better known as BISC, and it landed in Nature Electronics on December 8, 2025 with the kind of engineering ambition that tends to make the rest of the BCI field reconsider everything.
This is not a Neuralink press release, a Musk tweet, or a venture-funded demo. BISC came out of a research collaboration between Columbia University, Stanford University, NewYork-Presbyterian Hospital, and the University of Pennsylvania — funded by DARPA’s Neural Engineering System Design program — and its paper is dense with the kind of real electrophysiology data that neuroscientists actually care about. Senior author Ken Shepard, Lau Family Professor of Electrical Engineering at Columbia, led the chip engineering side. Senior author Andreas Tolias at Stanford anchored the neuroscience. The bridge between them produced something that looks less like a medical device and more like what the semiconductor industry does to room-sized computers when it decides they should fit in your pocket.
What BISC actually is
The quickest way to understand BISC is to start with what a typical brain implant looks like. Most current systems are built around a canister — a module stuffed with electronics that gets sealed into the skull or mounted on the chest, then connected to electrodes in the brain through a spaghetti of wires. The surgeries are long, the device is bulky, and the signal chain is lossy. Shepard describes the alternative bluntly: BISC slides into the subdural space — the thin gap between the outer membrane of the brain (the dura mater) and the cortical surface — “like a piece of wet tissue paper.”
Here’s what that chip actually packs in its 3 cubic millimeters:
65,536 titanium nitride electrodes arranged in a 256 × 256 grid
1,024 simultaneous recording channels, selectable from any position on the array
16,384 stimulation channels for sending signals back into the brain, not just reading out
On-chip analog front-end, signal processing, analog-to-digital conversion, and a radio frequency transceiver — all on one CMOS substrate
Wireless power via inductive link, so no battery sits inside the skull 🔋
The external piece is a wearable relay station, a headstage that beams power in and receives neural data out over a custom ultrawideband radio link. That link is the headline specification: 100 Mbps throughput, which the researchers confirm is at least 100 times higher than any competing wireless BCI currently available. The relay station then bridges to standard Wi-Fi, which means the brain, in a very literal sense, is on a network. 🌐
“Semiconductor technology has made this possible,” Shepard said in the Columbia press release, “allowing the computing power of room-sized computers to now fit in your pocket. We are now doing the same for medical implantables, allowing complex electronics to exist in the body while taking up almost no space.”
The engineering that actually makes it work
The honest answer to “why hasn’t anyone done this before?” is that merging electrodes and computation onto one substrate at this density is extremely hard. Most high-channel BCIs split the problem in two — electrodes here, processing there — and connect them with wires. That works, but the wire count scales badly, the connectors degrade, and you end up adding surgical complexity with every extra channel you want.
BISC’s breakthrough is monolithic integration 🔬: the electrodes and all supporting circuitry live on a single CMOS die, thinned to 50 micrometers. That last point matters enormously. At 50 μm, the chip is mechanically flexible — it can curve to follow the brain’s surface rather than pressing flat against it like a rigid board. The electrode pitch is approximately 28 micrometers, fine enough to capture retinotopic maps of the visual cortex at genuinely high resolution, which the team demonstrated in chronic recordings over multiple weeks.
A few technical details that deserve attention rather than burial:
The analog-to-digital conversion and multiplexing happen on-chip, which is why you can stream 1,024 channels wirelessly without drowning the radio in raw analog noise
The implant operates within strict thermal limits for subdural placement — the burst-based ultrawideband radio keeps power dissipation low enough that the chip doesn’t heat brain tissue
Stimulation and recording can happen on the same device, which matters for closed-loop therapies where you read activity, process it, and respond with targeted electrical pulses in real time
Chronic stability was verified across sessions spanning weeks in cortical recordings, with receptive fields shifting less than 0.05 degrees in eccentricity on average 📐
That chronic stability point is not academic. One of the persistent failure modes in BCI implants is signal degradation over time — electrodes scar, insulation breaks down, data quality slips. The fact that BISC maintains stable recordings over extended periods in animal models is what makes the human trial conversation even plausible.
Are you tracking what this means for the bandwidth problem that has hobbled BCI research for a decade? If you’re not, the short version is: more electrodes, read in parallel, transmitted wirelessly, with enough data throughput to run sophisticated AI decoders in real time. That combination has not existed in a subdural package before.
What this could actually treat
BISC’s clinical ambitions are specific, not vague. The paper and associated announcements identify five concrete conditions the system may address: 🏥
Epilepsy — high-resolution cortical mapping for seizure prediction and targeted stimulation to interrupt seizure onset
Spinal cord injury — decoding motor intention at sufficient bandwidth to drive prosthetic limbs or exoskeletons with the kind of nuance that current systems can’t capture
ALS — speech restoration using the chip’s ability to decode intended speech from motor cortex activity before it reaches the throat
Stroke — mapping perilesional activity and supporting neural rehabilitation through stimulation
Blindness — encoding visual input into cortical stimulation patterns precise enough to restore useful vision
That last application is where Stanford’s neuroscience programs were particularly involved. The spatial resolution of 65,000 electrodes across the visual cortex opens up a different class of visual prosthetic than anything the Utah Array — the clinical workhorse since 2004, which typically offers a few hundred channels — could attempt. And BISC can stimulate as well as record, which is the fundamental requirement for writing visual information back into the brain. 👁️
The potential for ALS and paralysis applications builds on momentum already in the field. Stanford researchers separately demonstrated earlier this year that a BCI could decode unspoken sentences with up to 74% accuracy, using a device implanted in a person who had lost the ability to speak. BISC’s bandwidth and resolution should, in principle, push that accuracy considerably higher. “By combining ultra-high resolution neural recording with fully wireless operation,” Dr. Zeng said in the Columbia release, “and pairing that with advanced decoding and stimulation algorithms, we are moving toward a future” of restored function.
How BISC compares to Neuralink — and why the comparison is complicated
Neuralink is the name everyone knows, and that’s mostly because Elon Musk has a media presence that is itself a kind of neural implant for the news cycle. Neuralink’s Telepathy device, which has now been implanted in a small number of people with paralysis, uses penetrating electrodes — thin threads that go directly into the cortex, where neurons fire loudest and the signal quality is highest. The tradeoff is that “the brain doesn’t really like having needles put into it,” as Synchron founder Tom Oxley memorably put it in a 2022 TED talk.
BISC’s approach is different in a way that matters clinically. It sits on the surface of the brain, not inside it. This makes it less invasive than Neuralink in terms of tissue disruption, while still achieving dramatically more electrodes and bandwidth than previous surface-level devices. The honest comparison looks something like this:
Neuralink: penetrating electrodes, higher single-neuron signal quality, ~1,000 channels, existing human data ⚡
BISC: surface electrodes, 65,536 electrodes, 1,024 channels, 100 Mbps wireless, no human data yet
Synchron (Stentrode): delivered through blood vessels, lowest invasiveness, lowest resolution
Precision Neuroscience: surface arrays similar in philosophy to BISC, currently in early human trials
Paradromics: highest raw channel count ambitions, penetrating, entering clinical trials late 2025
The Wikipedia overview of BCI approaches makes the core tension clear: invasiveness scales with signal quality, and every design is a bet on where that tradeoff should land. BISC bets that surface recording at extreme density, combined with powerful AI decoders, can match penetrating electrodes without the associated tissue damage. It’s a defensible position, though that claim needs human trial data before it’s settled science. 🧬
What nobody disputes is the bandwidth gap. 100 Mbps wireless throughput is genuinely unprecedented for an implantable BCI, and it opens up AI decoding architectures that are simply impossible on current systems — you can now feed a deep neural network the real-time activity of thousands of electrodes and let it find patterns that no human engineer designed the model to detect.
Kampto Neurotech and the road to a human skull
Research papers are great. Surgeries on actual people are the benchmark.
To move BISC from animal models to human patients, the Columbia and Stanford teams launched Kampto Neurotech, a startup founded by Columbia electrical engineering alumnus Dr. Nanyu Zeng, one of the project’s lead engineers. Kampto’s immediate focus is producing research-ready versions of the chip for preclinical work while simultaneously raising the capital needed to pursue a first-in-human trial. The company has already licensed the BISC technology from Columbia under US Patent 11617890 (issued April 4, 2023). 🚀
The path from here to a human trial is not trivial. BISC needs to clear FDA safety reviews, demonstrate long-term biocompatibility in chronic animal studies, and survive the kind of electromagnetic interference testing that tends to expose flaws in clever wireless designs. Those aren’t reasons to be dismissive — they’re the normal gauntlet that separates a promising chip from a clinical device. Neuralink went through years of that process before its 2024 first-in-human implant.
What Kampto has going for it is the institutional backing and engineering pedigree that some BCI startups lack. DARPA-funded research, three major academic medical centers, and a team that published the full architecture in a peer-reviewed Nature Electronics paper with a public GitHub repository for the data — that’s the kind of transparent science track record that the FDA and the investment community actually respond to. The chip’s performance has to hold up in humans, obviously, but the provenance here is legitimate.
What do you think the real bottleneck is — the engineering still left to do, the regulatory timeline, or the question of whether patients will actually consent to something this new? That’s probably where the most interesting debate in this space is happening right now, and it’s worth forming an opinion before the first human trial announcement turns the conversation into noise.