Neurotechnology · Brain-Computer Interfaces · AI · Medical Devices · United States · North America · Claim CheckRemarks of 6 October 2026, record to October 2026
The AI labs and the brain
Decades Rather Than Years
On 6 October 2026 the Stanford neuroscientist Andrew Huberman told a podcast that the real arms race in artificial intelligence is not about models at all: it is about reading and writing the human brain without implants, and OpenAI, Meta and Anthropic will all end up as biotech companies. The three are named as one bloc. They are standing in three different places. One had already written the largest cheque in a non-invasive brain-interface round nine months earlier — and that venture’s own estimate of how long the work takes is decades rather than years. One spent four years on non-invasive brain reading, stopped, and shipped a wristband instead. For the third, no public neurotechnology activity could be found at all.
What Happened
Three companies named as one bloc, standing in three different places What was said, what each company has actually done, and how well each of those is established.
- 20.9Words a minute from Meta’s shipped wrist interface, in its own Nature paper
- 100Words a minute its brain interface aimed at, before the programme ended in 2021
- $250mMerge Labs’ seed round, OpenAI’s largest single cheque — a second report says $252m
- DecadesThat venture’s own estimate of the timescale — “decades rather than years”
- 0FDA clearances on record for the two consumer brands checked in the agency’s database
| Standing | What | Where it is established |
|---|---|---|
| Confirmed | Huberman said the goal is “to read and write from your brain non-invasively” and that this is “what the real arms race of AI is about”. | Episode 494 of Invest Like the Best, released 6 Oct 2026; the publisher’s own description states the same thesis. |
| Confirmed | OpenAI wrote the largest single cheque in Merge Labs’ seed round. The venture is explicitly non-invasive: neurons reached with molecules instead of electrodes, and ultrasound to carry the signal. | Reported by two outlets on 15 and 16 Jan 2026, both quoting OpenAI’s own post and the company’s own statement. |
| Confirmed | Meta ran a non-invasive brain-reading programme for four years, ended it on 14 July 2021, open-sourced the software, and moved to the wrist. The wristband shipped on 18 September 2025. | The company’s own posts of 14 Jul 2021 and 18 Sep 2025, and its paper of 23 Jul 2025. |
| Confirmed | One company holds nine FDA authorisations for a prescription nerve stimulator and sells a consumer version that its own page says the FDA has not evaluated. | The FDA’s own device databases, queried 9 Oct 2026, and the manufacturer’s product page. |
| Confirmed, but not verifiable here | Merge Labs concedes the work may take “decades rather than years”. | One report of 16 Jan 2026 attributing it to the company. No company document saying so was read here. |
| Unconfirmed | That the named companies will become biotech companies, and that a device will one day turn a person’s motivation up and then off again. | A prediction. It is one person’s position, carried accurately; it is not a finding about anybody. |
| Not public | Whether the third named company has any neurotechnology activity. Nothing public was found — which is not the same as nothing existing. | Searching turned up no investment, product, programme or statement. A private one would be invisible here. |
What the record supports
That an AI company has funded a non-invasive brain interface, that a second abandoned one, and that the devices on sale today are sold as wellness products. All three are documented and dated.
What it does not
That any of it amounts to a race. One venture says decades; one company left the field for commercial reasons and said so; and for the third there is nothing public to put on either side of the ledger.
Timeline
Nine years, and the direction of travel is not the one the claim describes Everything below is dated by a document. The one thing that cannot be dated is an absence — see the last row of the table above.
- 2017Facebook Reality Labs starts a brain-computer interface programme with “an ambitious long-term goal: to develop a silent, non-invasive speech interface that would let people type just by imagining the words they want to say”.
- 2019Facebook buys CTRL-labs, a neural-interface startup working at the wrist.
- Feb 2021Flow Neuroscience acquires the rights to Halo Neuroscience’s brain-stimulation technology. Halo’s own domain now forwards to a social account whose last post is dated September 2020.
- 14 Jul 2021Facebook ends the brain programme. It open-sources its software, shares its head-mounted optical prototypes with researchers, and states the reason: wrist-based electromyography has “a nearer-term path to market”.
- May 2023Nature Neuroscience publishes the strongest non-invasive language decoder to date. It needs an MRI scanner, and the authors report that subject cooperation is required both to train and to apply it.
- 23 Jul 2025Meta’s wrist work publishes in Nature: one model that works across people with no per-person tuning, handwriting at 20.9 words a minute. It reads muscle signals, not brain activity.
- 18 Sep 2025It ships as a consumer product, bundled with display glasses, built on research with “nearly 200,000 research participants”.
- 11 Nov 2025UNESCO’s General Conference adopts the first global instrument on the ethics of neurotechnology. It treats neural data as especially sensitive, and it binds nobody.
- 15 Jan 2026Merge Labs leaves stealth, and OpenAI writes the largest single cheque in a seed round reported at $250m to $252m, at an $850m valuation. Its chief executive is one of the venture’s co-founders.
- 18 Mar 2026A $495 brain-stimulation headset is reported, with its founder saying it is positioned as a non-medical device so that it needs no regulatory clearance, and that no trials have been published.
- 6 Oct 2026The interview is released. “That’s what the real arms race of AI is about.”
- 7 Oct 2026The remarks are reported, in English and in Chinese, published at the same instant. None of the three named companies answers a request for comment.
The Argument
One bet placed, one bet withdrawn, and one blank The prediction and the record, company by company.
The claim is that reading and writing the brain without implants is what the AI companies are really building toward, and that this will turn them into biotech companies. As a description of one of the three it is already true, and was true before it was said: in January 2026 OpenAI wrote the largest single cheque in the seed round of a venture whose stated method is to reach neurons “using molecules instead of electrodes” and to carry signals “using deep-reaching modalities like ultrasound”. That is a biotech in the literal sense — the interface is chemistry, not wiring. It is also the strongest evidence for the claim, and it carries its own correction: the same venture concedes the work may take decades rather than years.
As a description of the second it points the wrong way down a road already travelled. Meta’s predecessor spent four years trying to read the brain non-invasively, with near-infrared light measuring blood oxygenation through the skull, aiming at a silent interface typing 100 words a minute. On 14 July 2021 it published the result: the software open-sourced, the prototypes handed to researchers, and the effort redirected to the wrist, because electromyography had “a nearer-term path to market”. What shipped four years later reads muscle signals — in the company’s own words, “the hand and finger movements you’ve already decided to make” — and writes 20.9 words a minute. Decisions caught on their way out, after the brain has finished with them.
Two things follow, and neither is an argument against the technology. The first is that the retreat was commercial, not scientific: Meta said in the same breath that it still believed in the long-term potential. The second is that a reader should not take the wristband for what it is not. Reading motor commands at the muscle and reading intention in the cortex are different problems, and the distance between them is most of the claim.
For the third company, nothing public was found at all — no investment, no product, no programme, no statement. That is a statement about what is visible, not about what exists; a private holding or an unannounced effort would leave no trace here. But the claim names three companies as one bloc, and on the public record one of them has nothing to put in either column.
- DecadesThe venture OpenAI funded, on how long its own work takes
- 20.9Words a minute, at the wrist — against 100 aimed for at the brain
- 0Public neurotechnology activities found for the third named company
| Non-invasive brain interface | What is on the public record | |
|---|---|---|
| OpenAI | Funded, Jan 2026 | Largest single cheque in a $250m–$252m seed round at an $850m valuation; molecules rather than electrodes, ultrasound rather than implants; its chief executive is a co-founder of the venture. |
| Meta | Ended, Jul 2021 | Four years of optical brain reading, open-sourced and handed over; the effort moved to wrist electromyography, which shipped in Sep 2025. The company says it still believes in the long-term potential. |
| Anthropic | None found | No investment, product, research programme or statement could be found. This records what is visible; it is not a finding that none exists. |
What Others Add
What you can buy, what the literature settles, and what the rules say The devices were checked one at a time against the regulator’s own database and the sellers’ own pages.
| Device | What it does | Regulatory standing |
|---|---|---|
| gammaCore | Stimulates the vagus nerve at the side of the neck. Prescription only. | Authorised. A 2017 De Novo plus eight later clearances, Class II, specifically to treat cluster headache and reduce attack frequency. |
| Truvaga | The same company’s consumer device, for the same nerve. | Unevaluated, and says so. Its own page: a general wellness product “not been evaluated by the US Food and Drug Administration”, not intended to diagnose, treat, cure or prevent anything. |
| Pulsetto | A neck-worn vagus-nerve wearable for winding down. | No clearance on record. The regulator’s database returns nothing under the name. |
| Neurable | Brain sensing built into ordinary headphones; tracks focus and fatigue. It reads, and does not stimulate. | No clearance on record, which is consistent with a product that measures rather than treats. |
| LIFTiD, Mave | Headsets passing a small direct current through the scalp, 20 minutes a day, for focus and mood. | Sold as non-medical. One maker says plainly that this positioning is what lets it sell without clearance, and that it has published no trials. |
| Halo Sport | A brain-stimulation headset for athletes, still named in round-ups of what is on sale. | Not on sale. The technology went to another company in February 2021; the maker’s domain now forwards to a social account last used in September 2020. |
Reading
It works, and it needs you to help
- The strongest published non-invasive decoder reconstructs continuous language, not just a short list of words.
- It runs on an MRI scanner. There is no headband version of this result.
- The authors tested the obvious worry directly: cooperation is required both to train and to apply it.
Writing
A decade of human studies, and no agreed dose
- Focused ultrasound reaches deep structures through the skull. A 2024 review counts 42 healthy-human studies and 22 clinical ones, with a good safety record.
- A 2025 paper names the obstacle as choosing the parameters, and its contribution is to propose a framework for thinking about dose — so there is not yet one.
- A second 2025 review is blunter: there is still controversy over which parameters excite and which inhibit. The direction of the effect is unsettled, before any question of personalising it.
The rules
A global instrument that binds nobody
- UNESCO’s General Conference adopted the first global text on the ethics of neurotechnology, covering the whole life cycle of a device.
- It treats neural data as especially sensitive and says a person may decline or withdraw at any time.
- It is a Recommendation, not a convention. It creates no binding obligation; what happens next is up to each member state.
| Also worth knowing | Detail |
|---|---|
| The same nerve, two worlds | What puts a device in the regulated world is the claim the seller makes, not the current it passes. Say it treats a disease and it must be shown to; say it supports wellness and it need not be. |
| Unevaluated is not disproved | A general wellness product has not been assessed. That is a statement about what is known, not a finding that the device does nothing or does harm. |
| Muscle is not brain | A wrist band reading electromyography is reading motor commands that have already left the brain. It is a neural interface; it is not a brain interface, and the shipped product does not claim to be one. |
| What is not carried here | One well-known user’s account of a nerve stimulator rests on a single second-hand report of a documentary, and a seventh device could not be verified at all. Neither is on this page. |
Conclusion
The interesting claim is not the race; it is that the rules decide themselves What to take from this, and what to hold loosely.
Settled
One AI company has funded a non-invasive brain interface and said why. Another ran one for four years, stopped, and shipped a wristband that reads muscles. A global ethics instrument exists and binds nobody. And a prescription nerve stimulator and an unregulated consumer one, from the same company, act on the same nerve.
Open
Whether any of this is a race. The timescale offered by the one venture inside it is decades; the parameter that decides whether stimulation excites or inhibits is still argued over in the literature; and the third company named has nothing public on either side.
Hold this loosely
“Nothing public was found” is not “nothing exists”, and a prediction is not falsified by being early or late. What a reader can do is keep the two columns apart: what a company has done, and what someone expects it to do.
- Read the regulatory line on the box, because it is the most informative sentence on it. “General wellness product, not evaluated” tells you nobody has checked. A clearance tells you what, specifically, the device was shown to do — and for whom.
- Check whether a device reads or writes before anything else. A headphone that measures your attention and a headset that passes current through your scalp are not the same kind of object, and only one of them is doing something to you.
- Treat a name in a list of what is on sale as a claim to check. One of the brands still named in such lists stopped trading in 2021; its own web address now leads to an account last used the year before that.
- When a timescale and a headline disagree, the timescale is usually the better guide. The venture doing the work says decades. That is not pessimism; it is the estimate of the people who have to deliver it.
Take-home and sources
- the interview as reported by Business Insider in English and in Chinese, and the publisher’s own description of episode 494 of Invest Like the Best
- TechCrunch and The Register on OpenAI’s participation in Merge Labs’ seed round, 15 and 16 Jan 2026
- Facebook Reality Labs’ own post ending its brain-computer interface programme, 14 Jul 2021
- the sEMG wristband paper in Nature, 23 Jul 2025, and Meta’s product announcement of 18 Sep 2025
- the US Food and Drug Administration’s 510(k) and device-classification databases, queried 9 Oct 2026
- product pages for Truvaga, LIFTiD and Neurable
- TechCrunch on Mave Health, 18 Mar 2026
- Xtalks on Flow Neuroscience’s acquisition of Halo Neuroscience, 17 Feb 2021, and Halo’s own dormant account
- Tang and colleagues in Nature Neuroscience, May 2023
- reviews of transcranial ultrasonic stimulation in Biomedical Engineering Letters (2024), Brain Stimulation (2025) and Bioengineering (2025)
- and the UNESCO Recommendation on the Ethics of Neurotechnology, adopted 11 Nov 2025. All read 9 Oct 2026.