TrendsHow is the field changing?

See where the field is moving.

How the BCI field is changing over time, from research output and patents to clinical translation, company formation and capital. Each signal is shown with its level, its direction and what it cannot tell you, so that growth is never mistaken for acceleration.

Data as of: 8 Oct 2026 · snapshot 16
Read from: 15,651 papers, 340 registered studies, 40 profiled companies, 20 funding rounds, 23 repositories
Trends are read on: Complete years only, 2010–2025. 2026 is shown, dashed, and never read as a direction.

01Six families of signal

The field in motion

Not how big the field is, but how it is changing. Each family answers one question, shows its level, its change and its trajectory, and says what it cannot know.

01Research output

Is the science still growing?

Yes, and faster than before. 3,734 scholarly works used the field’s terms in 2025, +41% on the year before. Growth over the latest three years ran at +23% a year, against +11% in the three years before that.

China now publishes the most: 1,120 works in 2025, 30% of the world count.

Level, 2025: 3,734 works
One-year change: +41%
Three-year growth: +23% a year; +11% in the three years before
Trajectory: AcceleratingConfidence high

All scholarly works (OpenAlex)
Biomedical papers (PubMed)

Publications on brain-computer interfaces a year, in OpenAlex and in PubMed (publications)
Year	All scholarly works (OpenAlex)	Biomedical papers (PubMed)
2010	584	357
2011	721	527
2012	771	529
2013	1,020	675
2014	1,032	771
2015	1,042	720
2016	1,128	749
2017	1,202	750
2018	1,358	832
2019	1,475	885
2020	1,744	1,014
2021	1,899	1,158
2022	2,007	1,163
2023	2,271	1,187
2024	2,657	1,374
2025	3,734	1,650
2026	4,266	1,284

Fig. 01 — Publications a year, two indexesThe dashed segment is 2026 to date. OpenAlex covers all disciplines; PubMed is the biomedical subset held as individual records.

Coverage, not growthThe 2026 line already stands above 2025, but only 73% of this year’s records carry an author country, against 97% last year, and OpenAlex lists 1,374 preprints for 2026 against 756 for all of 2025. That pattern points to a change in what the index covers. We do not read the open year as a trend.

Where it is published · scholarly works by region of author

Each panel has its own vertical scale, so compare shapes, not heights; the figure beside each name is its 2025 count. A work with authors in two regions counts in both.

By use case · PubMed papers, 2025

Category	Papers in 2025	Growth a year, latest three years	Trajectory
Digital autonomy	47	+25%	Accelerating
Speech / communication	192	+24%	Accelerating
Cognitive	417	+22%	Growing steadily
Neuromodulation	175	+20%	Accelerating
Rehabilitation	377	+16%	Slowing
Consumer	111	+14%	Slowing
Vision	32	+10%	Slowing
Motor control	240	+10%	Accelerating
Sensory restoration	91	+6%	Growing steadily
Hearing	30	−1%	Slowing

By modality · PubMed papers, 2025

Category	Papers in 2025	Growth a year, latest three years	Trajectory
Endovascular	20	+115%	Emerging
MEG / OPM	24	+39%	Emerging
ECoG / cortical surface	74	+25%	Accelerating
sEEG / depth	17	+24%	Slowing
Intracortical	160	+21%	Accelerating
Bidirectional / stimulation	185	+17%	Accelerating
Optical	101	+13%	Slowing
EEG	790	+10%	Slowing
Peripheral / neuromotor	133	+9%	Accelerating
Ultrasound	9	+4%	Slowing

Ranked by growth a year over the latest three complete years. Tags come from keyword rules over title, abstract and author keywords; a paper can carry several. 72% of papers could be given a modality and 62% a use case.

Who the studies are in · PubMed, 2025

Human studies 70% of papers: 1,157Accelerating
Animal studies 11% of papers: 178Accelerating
Reviews 18% of papers: 293Accelerating

Human or animal is read from the index terms PubMed assigns, with a text rule for records not yet indexed.

Institutions with the most works · 2021–2025

01Chinese Academy of Sciences409
02Tsinghua University223
03Tianjin University200
04Centre National de la Recherche Scientifique175
05Shanghai Jiao Tong University170
06Harvard University168
07Zhejiang University167
08University of Chinese Academy of Sciences166

Counted by OpenAlex from author affiliations. Volume, not influence: a count says nothing about what the work showed.

Method, coverage and limits· source read in full · last read 8 Oct 2026

What is counted: Articles, reviews and preprints across all disciplines whose title or abstract uses one of the field's terms, by year of publication.
Method: OpenAlex title-and-abstract search for eleven inclusion phrases, restricted to articles, reviews and preprints. Regions are counted from author affiliations; a work with authors in two regions counts once in each.
Source set: OpenAlex. Snapshot 16, source set 2026.10-b, taxonomy 2026.10-2.
Known limits: Counts depend on the term list: work that never names the field is missed, and adjacent work that does is included. Conference papers are counted as articles. The latest year is incomplete while indexing catches up.

Method, coverage and limits· source read in full · last read 8 Oct 2026

What is counted: Biomedical papers on brain-computer interfaces, neural interfaces and neuroprostheses, held as individual records and classified by modality, use case, study population and author country.
Method: PubMed search of title, abstract and the MeSH heading, then a relevance check. Modality and use case come from keyword rules over title and abstract; human or animal from MeSH check tags; country from author affiliation text. A paper can carry several modalities, use cases and countries.
Source set: PubMed (NCBI E-utilities). Snapshot 16, source set 2026.10-b, taxonomy 2026.10-2.
Known limits: PubMed covers biomedical journals; engineering conference papers, a large part of the EEG literature, are mostly absent. Keyword rules mis-tag papers that mention a method in passing. Affiliations before 2014 list only the first author's institution.

02Clinical translation

How quickly are human studies starting?

44 registered studies using the field’s terms started in 2025; 18 of them involve an implanted interface. 2026 has already passed that: 27 implant studies started with the year still open.

138 studies are open today, 77 of them with implants.

Level, 2025: 44 studies
One-year change: +38%
Three-year growth: +26% a year; 0% in the three years before
Trajectory: AcceleratingConfidence moderate

Implanted interface
Not implanted

Registered studies by year of start, split by whether an implanted interface is involved
Year	Implanted interface	Not implanted
2010	1	7
2011	3	6
2012	1	5
2013	6	10
2014	1	10
2015	7	8
2016	1	8
2017	3	11
2018	7	12
2019	6	16
2020	2	10
2021	7	16
2022	5	17
2023	5	13
2024	12	20
2025	18	26
2026	27	17

Fig. 02 — Registered studies by year of startClinicalTrials.gov only. The outlined column is 2026 to date. Hover or focus a column for its two parts.

Studies open now, by region of study site

66
N. America
42 with implants
36
Europe
18 with implants
21
China
11 with implants
4
Other APAC
1 with implants
2
Aus / NZ
2 with implants
2
Other regions
1 with implants

Read with careThese are counts from one registry. Registration on ClinicalTrials.gov is required for US studies and voluntary elsewhere. A July 2026 preprint counts 134 BCI trials in China’s own registry alone, several times what appears here for China. Read the briefing.

Regulatory milestones recorded · kept apart by type

18 Sep 2026Breakthrough designationBrain Interchange: second Breakthrough designation · US FDA
22 Jul 2026CE markPRIMA receives CE mark · EU (notified body)
28 Apr 2026Study authorised (IDE)RESONATE authorised under IDE · US FDA
8 Apr 2026Breakthrough designationBrain Interchange: Breakthrough designation for stroke · US FDA
13 Mar 2026Registration (China)NEO registered in China · China NMPA
Nov 2025Study authorised (IDE)Connect-One authorised under IDE · US FDA
Apr 2025510(k) clearanceLayer 7 Cortical Interface cleared · US FDA
24 Feb 2025OtherAdaptive deep brain stimulation approved · US FDA
19 Dec 2024De NovoARC-EX authorised through De Novo · US FDA
Aug 2024Innovative-device review (China)NEO enters the NMPA innovative-device review · China NMPA
Feb 2024Breakthrough designationARC-BCI receives Breakthrough Device designation · US FDA
30 Oct 2023Breakthrough designationbaroloop: Breakthrough designation · US FDA
4 May 2023Breakthrough designationCognixion ONE Axon receives Breakthrough Device designation · US FDA
Apr 2021De NovoIpsiHand authorised through De Novo · US FDA
11 Feb 2021Breakthrough designationBrainQ receives Breakthrough Device designation · US FDA
24 Jan 2020510(k) clearancecortiQ PRO cleared · US FDA
2019CE markrecoveriX certified in Europe · EU (notified body)

A designation, a study authorisation, a temporary-use clearance and a registration are different things. They are listed, never summed into “approvals”.

Method, coverage and limits· source read in full · last read 8 Oct 2026

What is counted: Registered clinical studies that use the field's terms, by the year the study started.
Method: ClinicalTrials.gov search with a relevance check on title, summary, interventions and keywords. Countries are the countries of the listed study sites. "Implanted" marks studies whose description involves an implanted or surgically placed interface.
Source set: ClinicalTrials.gov. Snapshot 16, source set 2026.10-b, taxonomy 2026.10-2.
Known limits: ClinicalTrials.gov only. Registration is required for US studies and voluntary elsewhere, so China, Japan and parts of Europe are strongly under-counted. Recent years rise further as studies are registered late.

03Capital

Where is disclosed funding going?

The 20 rounds recorded here sum to $2.90B. 80% went to companies in North America, 18% to China.

Disclosed funding onlyA sample of announced rounds, each with its source. Not a market total, and not advice of any kind.

Level, 2025: $970M
Trajectory: Not read: a curated sample shows level, not direction.

Disclosed funding in the rounds recorded by the Briefing, by year of announcement, in millions of US dollars
Year	Total
2022	$75M
2023	$280M
2024	$352M
2025	$970M
2026	$1.22B

Fig. 03 — Disclosed funding in the recorded rounds, by year announced20 rounds. The outlined column is 2026 to date.

By region of headquarters · all recorded rounds

North America$2.31B
China$529M
Europe$58M
Japan$2M

What this cannot show

Rounds in other currencies are converted, and the conversion is stated on each round. Undisclosed rounds, public grants and what large companies spend internally are absent. One large round moves a year: read the columns as a record of announcements, not as the size of a market.

Every round, with its source

Method, coverage and limits· a curated sample

What is counted: Sum of the publicly announced funding rounds the Briefing has recorded, by year of announcement.
Method: Each round is entered from a cited announcement or report. Amounts in other currencies are converted, with the method stated on the round.
Source set: Funding announcements and filings; Trade and general press (English); Chinese trade and financial press (动脉网, 36氪, 财联社, 新浪财经). Snapshot 16, source set 2026.10-b, taxonomy 2026.10-2.
Known limits: Disclosed funding only, and only the rounds recorded here: a sample, not a market total. Undisclosed rounds, grants and internal corporate spending are absent. A single large round can dominate a year.

04Company formation

Who is entering the field?

Of the 40 companies and initiatives profiled, 25 have a founding year we could source. 9 were founded since 2020, in 4 countries.

The rest are shown without a year. We leave a founding year out before we guess one.

Level, 2025: 2 founded
Trajectory: Not read: a curated sample shows level, not direction.

Fig. 04 — Profiled companies by founding year and regionOne mark per organisation. Outlined marks at the right have no sourced founding year.

Method, coverage and limits· a curated sample · last read 8 Oct 2026

What is counted: Companies and initiatives profiled by the Briefing, by founding year.
Method: Founding years are taken from cited sources. A company whose founding year we could not source is left out of this series rather than given a guessed year.
Source set: The BCI Briefing curated corpus. Snapshot 16, source set 2026.10-b, taxonomy 2026.10-2.
Known limits: Covers only the organisations profiled here, which are chosen for significance, not completeness. It cannot show how many companies were founded and failed.

05Patent activity

Where is technical work being protected?

855 patent families counted as BCI were first filed in 2023, the latest year that can be read as complete. Most concern decoding and interface and electrodes.

First filed 2018–2020: 1,388
First filed 2021–2023: 2,263
Later years: Incomplete; not read

The patent landscape

Fig. 05 — Patent families by year of first filingOne invention counts once, in the year of its earliest priority. Hatched years are incomplete.

Recent years are incomplete. Applications are published about eighteen months after they are first filed, so families first filed after 2023 are still appearing. A lower bar there is not a decline.

Method, coverage and limits· source read in full · last read 7 Oct 2026

What is counted: Patent families judged relevant to neural interfaces, by the year of their earliest priority. One invention counts once, however many offices published it.
Method: Families are found through targeted searches of the EPO's Open Patent Services, by applicant, by classification group and by wording, scored by published rules, and grouped by DOCDB simple family. Only families the rules count as BCI are included; adjacent neural technology and families the rules could not settle are not. A region is the region of the office where a family was first filed.
Source set: EPO Open Patent Services. Snapshot 16, source set 2026.10-b, taxonomy 2026.10-2.
Known limits: Patent activity is one indicator of technical and commercial activity and should not be read as a measure of technological quality, leadership or value. Applications are published about 18 months after filing, so the three latest priority years are incomplete and are not read as a trend. Chinese domestic-only filings and utility models are under-represented, and relevance is decided by rules that miss inventions described in unusual words.

06Open source and builder activity

Is the builder ecosystem active?

21 of the 23 tracked tools had a commit in the past twelve months, and none has been archived. Together they published 74 releases in 2025, 3 of them new major versions.

Stars are recorded and never used as a measure of quality.

Level, 2025: 74 releases
Trajectory: Not read: a curated sample shows level, not direction.

Releases published by the open-source tools the Briefing tracks, by year
Year	Total
2014	1
2015	9
2016	8
2017	20
2018	26
2019	54
2020	117
2021	104
2022	88
2023	78
2024	103
2025	74
2026	88

Fig. 06 — Releases published by the tracked tools, by year23 repositories. The outlined column is 2026 to date.

Open the tool radar for each project’s licence, language and latest release.

Method, coverage and limits· a curated sample · last read 8 Oct 2026

What is counted: Published releases of the open-source tools the Briefing tracks, by year.
Method: Read from each repository's release list through the GitHub API. Pre-releases are excluded. A major release is one that starts a new major version.
Source set: GitHub. Snapshot 16, source set 2026.10-b, taxonomy 2026.10-2.
Known limits: Covers the tracked tools only. Projects that tag versions without publishing releases are under-counted. Release counts say nothing about scientific quality.

02Ask your own question

Trend explorer

How has speech research changed? How do China, Europe and North America compare on clinical studies? Choose a metric and cut it by geography, modality, use case and time.

Metric

Geography · choose up to four to compare

Or one country

Modality

Use case

Study population

Time

Research · world

03Where it is happening

A field that is no longer in one place

Six layers of activity, from 2010 to today. Move the year to watch companies appear, studies start and research output shift between regions.

Profiled companies at their headquarters, from the year each was founded.

20102026

37 marks on the companies layer in 2026. Hover a mark, or use the list.

Marks are records, not decoration. Country marks sit at the centre of the country, not at an institution. Mark area follows the count, so a few large countries dominate; use the region zoom to read the rest. 3 profiled companies have not disclosed a location and are not placed. 36 countries with a small number of works have no position recorded and are not drawn. Clinical counts come from ClinicalTrials.gov and under-count studies outside the United States.

Region by region

Momentum is uneven, and so is our view of it

The global ecosystem is not symmetrical and this table does not pretend it is. Each region has its own report, with what we see there and what we do not.

All regional reports

Region	Works, 2025	Share of world	Research trajectory	Studies open	Companies profiled	Disclosed capital
North America	656	18%	Accelerating+10% a year, +6% before	66	18	$2.31B
Europe	798	21%	Growing steadily+10% a year, +12% before	36	10	$58M
China	1,120	30%	Slowing+26% a year, +36% before	21	7	$529M
Japan	82	2%	Accelerating+14% a year, −1% before	0	2	$2M
South Korea	126	3%	Accelerating+15% a year, +6% before	0	1	–
Australia / New Zealand	89	2%	Accelerating+16% a year, 0% before	2	1	–
Rest of Asia-Pacific	433	12%	Accelerating+24% a year, +11% before	4	0	–
Middle East, Latin America and Africa	318	9%	Growing steadily+15% a year, +14% before	2	1	–

Works are scholarly works by region of author (OpenAlex). Studies open are from ClinicalTrials.gov. Companies and capital are the organisations and rounds the Briefing has profiled, a curated set that says as much about our coverage as about the region.

04A measured lens

Where science and capital are moving

Where is the literature growing fastest, and has disclosed capital followed? One mark per technology or use case.

Across · scientific momentum: Compound annual growth of PubMed papers tagged with the category, 2022 to 2025. The vertical line is the rate for all BCI papers.
Up · disclosed capital: Sum of the recorded rounds announced 2024–2026 by companies working in the category, on a log scale. A sample. A company in two categories counts in both.
Size: Number of profiled companies working in the category.
What it is not: A valuation, a ranking or advice. More capital is not better and less is not an opportunity: the figure shows where two measurable things stand, and where our sample of rounds is thin.

Size: companies profiled
Dashed: under 30 papers a year

Reading

Stimulation

185 papers in 2025, +17% a year over three years, growing faster than the field (+12%). $1.21B of disclosed funding is recorded across 9 rounds in 2024–2026, in 20 profiled companies. High momentum on both.

Publication growth and disclosed funding by category
Category	Papers	Growth a year	Disclosed capital
Endovascular	20	+115%	$200M
MEG / OPM	24	+39%	none
ECoG	74	+25%	$412M
sEEG	17	+24%	none
Intracortical	160	+21%	$967M
Stimulation	185	+17%	$1.21B
Optical	101	+13%	none
EEG	790	+10%	$316M
Peripheral	133	+9%	$516M
Ultrasound	9	+4%	$336M

Fig. 07 — Publication growth against disclosed fundingSelect a mark to read it. Dashed marks rest on fewer than 30 papers a year.

05Capability over time

Frontier velocity

The Frontier map shows where each capability stands. This shows how it got there: a measured benchmark where studies can be compared, and milestones where they cannot.

Open the Frontier map

Where a benchmark is comparable

Speech: from fifteen words a minute toward conversation

Four peer-reviewed studies report a decoding rate in words per minute. They are plotted as points and not joined, because each comes from one participant with a different vocabulary and pacing.

The latest point is lower than the two before it, and is the bigger advance: the participant spoke at his own pace, at 97.5% accuracy, for eight months. A single curve would have hidden that.

Intracortical
ECoG (cortical surface)

Benchmark point

Card et al. 2024New England Journal of Medicine

Metric: ≈32 words per minute. Self-paced conversation; 97.5% word accuracy over a 125,000-word vocabulary, sustained for 8.4 months.
Participants: One man with ALS
Interface: Intracortical microelectrode arrays
Limitation: A lower rate than 2023 because the participant set the pace; accuracy, not speed, was the advance.
Primary source: PubMed record (opens in a new tab)

Fig. 08 — Speech decoding rate, as each study reports itSelect a point for the publication, participants, context and limits.

Where metrics differ · milestone timelines

Eight capabilities, each on its own path

Most capabilities have no shared yardstick: a robotic arm, a cursor and a sense of touch are not measured in the same unit. For those the record is a sequence of milestones, each with its evidence class, ending in a dashed step that has not happened.

From synthesising speech in people who could talk, to a man with ALS holding daily conversations.

2019
Early milestone
Speech synthesised from cortical activity
A decoder turned recordings from the brain's surface into audible sentences by first reconstructing vocal-tract movements. Participants could still speak.
Anumanchipalli GK et al., Nature
2021
Intermediate milestone
First sentences from a person who cannot speak
A man with anarthria produced sentences from a 50-word vocabulary, decoded directly from attempted speech.
15.2 words/minMedian rate
Moses DA et al., New England Journal of Medicine
2023
Intermediate milestone
Large vocabularies at conversational pace
Two groups, using different implants, published in the same issue: 62 and 78 words per minute over vocabularies of thousands of words.
62–78 words/minDecoding rate
Willett FR et al., Nature
2024–26
Current state
Accurate enough to rely on
Accuracy reached 97.5% and held for more than eight months of daily use. Voice is now synthesised as the person speaks, and one implant can carry speech and gesture together.
97.5%Word accuracy
Card NS et al., New England Journal of Medicine
Next
Next frontier
Everyday conversation on implantable hardware
The open problems are fully implanted wireless systems, more participants, and decoders that do not need hours of calibration for each new user.

06BCI Briefing outlook

What the trajectory suggests

Questions about the next one to five years, each answered the same way: what has been observed, what leads it, what stands in the way, and why we think so. Open one to see the reasoning.

Now

BCI Briefing outlookOutlookImplanted BCIs used at homeHow soon is everyday home use the norm for study participants, not the exception?Confidence high

Observed trajectory

Home use was shown with a fully implanted ECoG system in 2016 and with an endovascular implant from 2020. Commercial feasibility studies now enrol participants who use their systems at home, and China's first registered implant was evaluated for home-based assisted grasp.

Leading indicators

Registered commercial studies of home cursor control run in several countries. Source
A registration trial with home use as its setting completed with 32 participants. Source
A European study planned for late 2026 is designed around one year of home use. Source

Why do we think this?

This has already happened in several programmes and study designs now assume it. The open question is durability and support, not feasibility.

Plausible next milestone

Peer-reviewed reports of unsupervised home use measured in hours a day across a cohort, not an individual.

Horizon Now · Confidence high

Bottlenecks

Home use still depends on research staff for set-up, recalibration and repairs.
Published follow-up beyond a year exists for very few participants.

Depends on

Wireless, fully implanted hardware.
Decoders that stay stable without daily recalibration.

Key uncertainties

What happens to participants and devices when a study ends.
How much support a household needs, and who pays for it.

Evidence it rests onFully Implanted Brain–Computer Interface in a Locked-In Patient with ALS Motor neuroprosthesis implanted with neurointerventional surgery improves capacity for activities of daily living tasks in severe paralysis: first in-human experience Assessment of Safety of a Fully Implanted Endovascular Brain-Computer Interface for Severe Paralysis in 4 Patients: The SWITCH Study An Accurate and Rapidly Calibrating Speech Neuroprosthesis PRIME NEO Digital Autonomy

Trend dataClinical

1–2 years

BCI Briefing outlookOutlookDecoders that work out of the boxCan a BCI be useful on day one, for a new person, without hours of training?Confidence moderate

Observed trajectory

Calibration has shrunk from weeks to minutes in the best single-participant results: a speech BCI was usable after 30 minutes on its first day. At the wrist, generic models trained on thousands of people already work for a new user with no calibration. Decoding across people from brain signals is now being attempted on large open recordings.

Leading indicators

A company reports pretraining cursor decoders on more than 50,000 hours of participant data. Source
A large open MEG speech dataset supports cross-subject benchmarks. Source
Open tooling for neural foundation models is maturing. Source

Why do we think this?

Three independent lines point the same way: shorter calibration in published implanted studies, calibration-free decoding at the periphery, and the data volume inside commercial programmes. What is missing is peer-reviewed evidence of transfer between people with implants.

Plausible next milestone

A peer-reviewed implanted decoder that a new participant uses on the first day at close to its eventual performance.

Horizon 1–2 years · Confidence moderate

Bottlenecks

Implanted data are scarce: a few dozen people, each with a different array placement.
Electrodes do not sit in the same neurons from person to person, so transfer has to be learned, not assumed.

Depends on

Shared datasets and benchmarks with comparable tasks.
Hardware that records for long enough to accumulate training data.

Key uncertainties

Whether pretraining gains reported by companies hold up in independent evaluation.
How far results from the wrist or from MEG carry over to implanted recordings.

Evidence it rests onAn Accurate and Rapidly Calibrating Speech Neuroprosthesis A generic non-invasive neuromotor interface for human-computer interaction FALCON benchmark torch_brain Neuralink Meta Reality Labs

Trend dataResearchOpen source

BCI Briefing outlookForecastFully implanted, wireless, high channel countWhen do the systems with the best signals lose the connector through the skin?Confidence moderate

Observed trajectory

The highest-performance published results still come from percutaneous research systems. Fully implanted wireless systems exist at lower channel counts (epidural and endovascular) and at high channel counts inside commercial studies whose results are mostly company-reported.

Leading indicators

A 1,024-channel thin-film surface array is described in a peer-reviewed paper with an intraoperative pilot. Source
An optical data link reported at up to 50 megabits per second is planned for its first chronic human study. Source
Chinese developers report a fully implanted, battery-integrated wireless system entering a first clinical trial. Source

Why do we think this?

Several independent engineering routes to the same goal are in or near human studies at once, on three continents. That makes the milestone likely; which design reaches it first is not predictable from public data.

Plausible next milestone

A peer-reviewed report of a fully implanted system with several hundred channels or more, with at least a year of follow-up.

Horizon 1–2 years · Confidence moderate

Bottlenecks

Heat, power and bandwidth limits inside the skull.
Hermetic packaging that lasts for years with thousands of feedthroughs.
Little independent data on signal longevity for the new thin-film and flexible electrodes.

Depends on

Low-power electronics and on-implant compression.
Surgical methods that scale beyond a few specialist centres.

Key uncertainties

Whether flexible electrodes hold their signals for longer than rigid arrays in people.
Whether optical or radio links prove more practical in daily life.

Evidence it rests onMinimally invasive implantation of scalable high-density cortical microelectrode arrays for multimodal neural decoding and stimulation An exoskeleton controlled by an epidural wireless brain–machine interface in a tetraplegic patient: a proof-of-concept demonstration Neuralink Precision Neuroscience ABILITY Neurotech StairMed NeuroXess CorTec

Trend dataClinicalPatentsCapital

BCI Briefing outlookScenarioFrom study device to authorised productWhere, and for what, will implanted BCIs be authorised next?Confidence moderate

Observed trajectory

Three authorisations now frame the field, and they are not alike: a Chinese registration of an epidural BCI for assisted grasp in a defined group, a US clearance of a surface array for temporary use, and a European CE mark for a retinal prosthesis. No permanently implanted BCI for communication or computer control is authorised anywhere.

Leading indicators

A preprint counts five BCI-related products approved in China as of June 2026, three of them non-invasive rehabilitation systems. Source
FDA Breakthrough designations continue to be granted; they speed dialogue and are not authorisations. Source

Why do we think this?

The first registration shows that a narrow claim with a bounded risk can be authorised now. Broader claims need trials that have not been registered yet, so the next authorisations are more likely to be narrow than general.

Plausible next milestone

A second jurisdiction authorises a permanently implanted BCI, or the first reports real-world use after registration.

Horizon 1–2 years · Confidence moderate

Bottlenecks

No agreed clinical endpoint for communication or digital independence.
Reimbursement is unsettled in every market, including where a device is authorised.

Depends on

Pivotal trials large enough to support a claim.
Payers' willingness to cover the implant, the surgery and long-term support.

Key uncertainties

How the first registered implant performs outside its trial.
Whether authorities accept each other's evidence.

Scenario envelope

Assumptions: Assumes regulators keep requiring a dedicated pivotal study for any permanent implant with a new function.

Conservative: No further authorisation of a permanent BCI implant within the horizon; activity stays in feasibility studies.
Base trajectory: Further narrow-indication authorisations in China, and the first pivotal trials registered in the United States.
Accelerated: A permanent BCI implant is authorised in a second major jurisdiction.

Evidence it rests onNEO registered in China Layer 7 Cortical Interface cleared PRIMA receives CE mark Connect-One authorised under IDE Neuracle Science Corporation

Trend dataClinical

BCI Briefing outlookOutlookNeural interfaces as consumer productsDoes the first mass-market neural interface read the brain at all?Confidence moderate

Observed trajectory

The neural interfaces sold in volume today read muscle at the wrist or forearm, or EEG at the scalp. A wrist sEMG band generalises across users without calibration, and a Chinese maker of bionic limbs and EEG wearables raised the largest BCI-related round outside the United States and filed to list.

Leading indicators

Peer-reviewed generic sEMG decoding across thousands of users. Source
A non-invasive neural-interface company raises CNY 2 billion and files for listing. Source

Why do we think this?

Scale is arriving first where no surgery is needed and the signal is strong: at the periphery. That is a different technology from the implanted systems in clinical studies and should not be read as progress toward them.

Plausible next milestone

Independent evaluation of a consumer neural interface in everyday use at scale.

Horizon 1–2 years · Confidence moderate

Bottlenecks

Scalp EEG carries too little signal for fast, fine control.
Consumer claims are rarely tested in independent studies.

Depends on

Applications where a modest signal is enough.
Regulatory clarity on where a wellness product ends and a medical device begins.

Key uncertainties

Whether demand exists beyond early adopters and assistive use.
How neural-data privacy rules develop.

Evidence it rests onMeta Reality Labs BrainCo A generic non-invasive neuromotor interface for human-computer interaction Consumer Interfaces Peripheral Neural Interfaces

Trend dataCompaniesCapital

3–5 years

BCI Briefing outlookForecastCommunication at the speed of conversationWhen does a speech BCI become something a person can simply talk with?Confidence moderate

Observed trajectory

Published decoding rates rose from 15 words a minute over a 50-word vocabulary in 2021 to 62 and 78 words a minute over large vocabularies in 2023. In 2024 one participant held 97.5% word accuracy for more than eight months of self-paced use, at about 32 words a minute. Voice is now synthesised as the person speaks, and one implant has carried speech and gesture together.

Leading indicators

A commercial intracortical system reports real-time speech from its first study participant, without figures. Source
A second commercial early feasibility study for communication is recruiting. Source
Public speech-decoding datasets let groups improve decoders without new surgery. Source

Why do we think this?

The science has moved from possible to dependable in five years, in two different interface types, and commercial studies with speech as the stated indication have now started. The remaining gaps are engineering and evidence, not a missing principle.

Plausible next milestone

Peer-reviewed speech results from a fully implanted commercial system in more than one participant.

Horizon 3–5 years · Confidence moderate

Bottlenecks

Every peer-reviewed high-performance result comes from one participant at a time, on a wired research system.
Speed and accuracy have not peaked together: the most accurate system is not the fastest. Natural conversation runs near 160 words a minute.
Decoders are trained per person, and evidence across causes of speech loss is thin.

Depends on

A fully implanted system with enough channels over speech cortex.
Language models that run with low delay on the device or beside it.
A regulatory route for a communication indication.

Key uncertainties

Whether accuracy holds as a disease such as ALS progresses.
How much performance survives outside structured sessions, in noise and fatigue.

Scenario envelope

Assumptions: All three cases assume no serious safety signal in current studies. They differ in how quickly commercial systems reproduce academic performance and how regulators treat a communication endpoint.

Conservative: Speech BCIs stay in early feasibility studies through the horizon, with published results from a handful of participants.
Base trajectory: A pivotal trial of an implanted BCI for communication is registered, and peer-reviewed multi-participant results appear.
Accelerated: A communication BCI receives its first marketing authorisation in one jurisdiction.

Evidence it rests onNeuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria A high-performance speech neuroprosthesis A high-performance neuroprosthesis for speech decoding and avatar control An Accurate and Rapidly Calibrating Speech Neuroprosthesis A streaming brain-to-voice neuroprosthesis to restore naturalistic communication Simultaneous speech and gesture decoding for multimodal communication in paralysis Speech & Communication Paradromics

Trend dataResearchClinical

BCI Briefing outlookOutlookLess invasive routes to the brainHow much function can be delivered without opening the dura, or the skull?Confidence moderate

Observed trajectory

Three less invasive designs have reached people: an endovascular array with a 12-month peer-reviewed safety record, epidural implants that never touch the brain, and thin films slid under the skull through a slit. The first implanted BCI to be registered anywhere is an epidural one.

Leading indicators

An epidural BCI is registered in China for assisted grasp. Source
An endovascular company raises a round to prepare for pivotal work. Source
A surface array holds a US clearance for temporary use of up to 30 days. Source

Why do we think this?

Regulators and investors have both moved toward the less invasive end first. That says these designs are easier to authorise, not that they will prove the most useful.

Plausible next milestone

A pivotal trial of a permanently implanted, minimally invasive BCI registered in the United States or Europe.

Horizon 3–5 years · Confidence moderate

Bottlenecks

Lower channel counts and coarser signals than penetrating electrodes.
Functions shown so far are simpler: clicks, grasp, cursor control.

Depends on

Decoders that extract more from fewer, coarser channels.
Evidence that a simpler procedure widens the group of surgeons and patients.

Key uncertainties

Whether the functional ceiling of low-channel systems is high enough for patients to choose them.
How regulators outside China weigh a narrower function against a lower surgical risk.

Evidence it rests onSynchron Neuracle Precision Neuroscience Clinatec Assessment of Safety of a Fully Implanted Endovascular Brain-Computer Interface for Severe Paralysis in 4 Patients: The SWITCH Study Endovascular

Trend dataClinicalCapital

BCI Briefing outlookOutlookTouch that closes the loopWhen does a brain-controlled hand feel what it holds?Confidence low

Observed trajectory

Stimulating somatosensory cortex evokes touch felt in the hand, and feedback halved the time to complete robotic-arm tasks in one participant. Later work has made evoked sensations more stable and better localised. Dedicated commercial investment remains small.

Leading indicators

Peer-reviewed work reports tactile edges and motion evoked by patterned stimulation. Source
In three participants, evoked sensations are stronger, easier to localise and stable over the study. Source

Why do we think this?

The science is consistent and steadily improving, but it is carried by a handful of laboratories and has no commercial programme built around it. Progress is therefore likely to be real and slow.

Plausible next milestone

A registered study of a bidirectional implant with sensory feedback as a stated endpoint.

Horizon 3–5 years · Confidence low

Bottlenecks

Sensations are still coarse compared with natural touch.
Stimulation and recording interfere on shared hardware.
Few participants, all in academic studies.

Depends on

Bidirectional implants with enough channels in sensory as well as motor cortex.
A clinical indication in which feedback changes an outcome that matters.

Key uncertainties

Whether feedback improves daily function enough to justify a second array.
Long-term safety of chronic cortical stimulation at these sites.

Evidence it rests onA brain-computer interface that evokes tactile sensations improves robotic arm control Tactile edges and motion via patterned microstimulation of the human somatosensory cortex Evoking stable and precise tactile sensations via multi-electrode intracortical microstimulation of the somatosensory cortex Somatosensory FeedbackUniversity of PittsburghUniversity of Chicago

Trend dataResearchCapital

Beyond

BCI Briefing outlookOutlookReading and writing the brain with ultrasoundCan an interface without electrodes reach useful bandwidth?Confidence low

Observed trajectory

Two companies founded since 2025, one in the United States and one in China, raised large early rounds for ultrasound-based interfaces. Neither has published human or device data. Functional ultrasound in people has so far needed a window in the skull.

Leading indicators

A US company emerges with a $252 million seed round. Source
A Chinese company reports RMB 570 million raised in its first six months. Source

Why do we think this?

Capital has arrived well ahead of evidence. That is a signal about ambition and about investors' appetite for non-surgical routes; it is not yet a signal about what works.

Plausible next milestone

A first published human demonstration of closed-loop control or therapy through an ultrasound interface.

Horizon Beyond · Confidence low

Bottlenecks

The skull strongly attenuates and distorts ultrasound.
The signal follows blood flow and is slower than electrical activity.
No human performance data exist to judge either company by.

Depends on

Transcranial imaging at useful resolution, or a minimally invasive acoustic window.
Evidence that haemodynamic signals can drive control fast enough to be useful.

Key uncertainties

Whether the first applications will be therapeutic (pain, mood) rather than control.
Whether capital committed now outlasts the time the physics needs.

Evidence it rests onMerge Labs Gestala tFUS Merge Labs Seed

Trend dataCapitalResearchPatents

Issued 5 Oct 2026. Only the paragraph headed “Observed trajectory” describes what has happened. Everything else is a forward-looking judgement by The BCI Briefing and may prove wrong. No trend line on this page is extended into the future, and past growth is not treated as a promise of more. When a topic is revised the earlier version is kept and compared with what occurred.

07How this page knows what it says

Method, sources and limits

A trend is only as good as the count under it. This is how the counts are made, when each source was last read, and what changed in our coverage along the way.

How a trajectory is read

A trajectory describes how the growth rate has changed, not only whether a number went up. Two rates are compared: compound annual growth over the latest three complete years (2022 to 2025), and over the three before. A difference of 3 percentage points a year counts as a change.

2026 never enters a reading. Curated samples (capital, companies, tools) are shown as levels and not read for direction, because a new entry in our corpus is not a change in the world.

Accelerating: Growing, and faster over the latest three years than over the three before.
Growing steadily: Growing at about the same rate as in the three years before.
Stable: Within three percent a year of flat.
Slowing: Growth over the latest three years is clearly lower than over the three before.
Declining: Falling by three percent a year or more over the latest three years.
Emerging: Started from fewer than ten a year three years ago and has at least doubled.
Too little data: Fewer than seven complete years, or too few items a year to read a direction.
Confidence: Set by how much data is behind a reading, never by its direction: high at 200 or more items in the latest year from a source read in full, moderate from 30, low below that and for every sample.

A change in coverage is not a change in the world

If we connect a new source tomorrow, our counts will jump. The field will not have. Every addition of a source, backfill or change of rules is logged with its date, and each snapshot records the source set and taxonomy it was computed under, so a step in a series can be traced to its cause.

Charts are drawn from stored snapshots, not recomputed from today’s database. Snapshot 16 was taken on 8 Oct 2026.

Coverage log

5 Oct 2026BackfillClinicalTrials.gov backfill: 339 matching studies read, full registry history.
5 Oct 2026BackfillOpenAlex yearly aggregates from 2010 added: world, eight regions, countries, preprints and reviews.
5 Oct 2026BackfillPubMed backfill from 2010: 15638 papers indexed. Counts before this date were not held by the Briefing.
5 Oct 2026Collector changedImplant studies are now also recognised from the description of the intervention, and sponsors are matched to organisations without their legal form. Counts of implant studies rise where a registry entry said "implantable" only there.
5 Oct 2026Taxonomy changedThree classes added (bioelectronic medicine, biohybrid, enabling technology) and one rule stated for what is counted as a BCI company: a primary class of core or bidirectional BCI.
7 Oct 2026BackfillPatent families published from 2015 added for the first time, found through targeted EPO OPS queries by applicant, class and wording.

80%
of PubMed papers placed in a country: 86%
identified as human or animal studies: 72%
given at least one modality: 62%
given at least one use case

The rest could not be classified by the rules and appear in totals only. A paper that is not tagged is not evidence that a topic is absent.

Sources read automatically

Collectors run on a schedule against public interfaces. Each run is logged; a source that keeps failing is marked here.

ClinicalClinicalTrials.govPublic API. Registration here is voluntary for studies run outside the United States, so non-US studies are under-represented.Read 8 Oct 2026checked daily
CompanyCompany discovery (sponsors, applicants, leads)Reads records the pipeline already holds and lists organisations that appear in them without a profile. A candidate is a pointer for an editor, never a published fact.Read 8 Oct 2026checked daily
EventsOrganisers' event pagesEach event's official page is read once a day, and organisers' listing pages once a week, only where robots.txt allows. A program compares the dates a page shows with its last reading and asks an editor; it copies no fact into the calendar.Read 8 Oct 2026checked daily
Open sourceGitHubPublic API. 60 requests an hour without a token. Stars are stored as descriptive metadata and never used to rank quality.Read 8 Oct 2026checked daily
PatentEPO Open Patent ServicesRequires a registered key (free tier, with a weekly volume allowance). Used for discovery and for checking: targeted searches by applicant, class and wording, then each relevant family's members and recorded events. Never crawled. Chinese utility models and domestic-only filings are less complete than invention patents.Read 7 Oct 2026checked weekly
ScientificOpenAlexCC0 data. Without an API key the daily allowance covers yearly aggregates only, so this source is used for counts by year and country, not for item records.Read 8 Oct 2026checked weekly
ScientificPubMed (NCBI E-utilities)Public API. At most 3 requests a second without a key, 10 with one. Abstracts are read for classification and not stored.Read 8 Oct 2026checked daily

Sources read by editors

Used for company facts, funding, regulatory actions and news. Nothing from these enters the site without an editor.

CompanyCompany and institution announcementsPrimary for what an organisation says; not independent evidence that it is so.Read by an editor
DatasetDataset repositories (DANDI, OpenNeuro, PhysioNet, DataCite)Read by an editor
FundingFunding announcements and filingsDisclosed rounds only. Amounts are recorded as stated; undisclosed and approximate amounts are not turned into numbers.Read by an editor
PressChinese trade and financial press (动脉网, 36氪, 财联社, 新浪财经)zhRead in Chinese. Titles are kept in the original with a translation.Read by an editor
PressTrade and general press (English)Used to discover events and as independent reporting. Primary documents are preferred where they exist.Read by an editor
RegulatoryEuropean notified bodies and national authoritiesRead by an editor. CE certificates are not published centrally; most are known through company statements.Read by an editor
RegulatoryNational Medical Products Administration (国家药品监督管理局)zhRead by an editor, in Chinese.Read by an editor
RegulatoryUS Food and Drug AdministrationRead by an editor. Clearances and approvals are public; IDE authorisations and Breakthrough designations are confidential unless the company discloses them.Read by an editor
ScientificPreprints cited individuallyRead by an editor

Known gaps

Sources we know we should read and do not yet. Their absence shapes every regional comparison on this page.

ClinicalAustralian New Zealand Clinical Trials RegistryNot yet connected.Not connected
ClinicalChinese Clinical Trial Registry (ChiCTR)zhNot yet connected: no public API. Most Chinese BCI studies are registered here and not on ClinicalTrials.gov.Not connected
ClinicalEU Clinical Trials Information System / EUDAMEDNot yet connected. CTIS covers medicines; device investigations are only partly public through EUDAMED.Not connected
ClinicalJapan Registry of Clinical Trials (jRCT)jaNot yet connected.Not connected
ClinicalWHO International Clinical Trials Registry PlatformNot yet connected. Aggregates national registries; bulk access requires an agreement.Not connected
PatentGoogle Patents Public Data (BigQuery; optional, not in use)Not in use. A provider for it exists and could add a second discovery source later; it needs a Google Cloud project, and queries are billed by the columns read.Not connected
ScientificarXivNot yet connected.Not connected
ScientificbioRxiv and medRxivNot yet connected.Not connected
ScientificCrossrefNot yet connected. Planned for DOI metadata and retraction notices.Not connected
ScientificEurope PMCNot yet connected. Would add preprints indexed from bioRxiv and medRxiv.Not connected

What we count as BCI

Counts depend on where the line is drawn. Research and clinical counts follow the field’s own terms, so they include whatever calls itself a BCI, neural interface or neuroprosthesis. Organisations are sorted by hand into seven classes, so that a wrist band, a stimulator and a speech implant are never presented as the same kind of thing.

The full definitions

Core BCI19 profiled
Records activity from the brain and decodes it into a command, text, speech or movement for the person it is recorded from.
Bidirectional BCI1 profiled
Reads from the brain and writes back to the nervous system in the same system, with the reading shaping the stimulation.
Neural interface8 profiled
Hardware that records from or stimulates the brain and is being developed toward BCI use, without a demonstrated closed decoding application yet.
Neuroprosthetic3 profiled
Replaces a lost sensory or motor function by stimulating the nervous system, without decoding the user's intent from the brain.
Neuromodulation5 profiled
Stimulates the nervous system to treat a condition. Systems that sense neural activity and adapt to it (closed loop) are where the class meets BCI. Open-loop systems are profiled only where they bear on the field, and are never counted as BCI.
Peripheral / neuromotor interface2 profiled
Reads intent outside the brain, from muscle or peripheral nerve, and uses it to control a device.
Bioelectronic medicine1 profiled
Interfaces with a peripheral nerve, most often the vagus nerve, to treat disease in an organ system. Nothing is decoded from the brain and no sense or movement is restored.
Biohybrid0 profiled
Couples electronics with living neural tissue that is grown or kept outside a body. Related science, not an interface with a person.
Enabling technology0 profiled
Makes the instruments, components or platforms that BCI research and products are built with, where that is the organisation's main business.
Adjacent technology1 profiled
Surrounds the field without being an interface to the nervous system for control or restoration.

Is the science still growing?

Where it is published · scholarly works by region of author

By use case · PubMed papers, 2025

By modality · PubMed papers, 2025

Who the studies are in · PubMed, 2025

Institutions with the most works · 2021–2025

How quickly are human studies starting?

Studies open now, by region of study site

Regulatory milestones recorded · kept apart by type

Where is disclosed funding going?

By region of headquarters · all recorded rounds

What this cannot show

Who is entering the field?

Where is technical work being protected?

Is the builder ecosystem active?

Where science and capital are moving

Speech: from fifteen words a minute toward conversation

Eight capabilities, each on its own path

Speech synthesised from cortical activity

First sentences from a person who cannot speak

Large vocabularies at conversational pace

Accurate enough to rely on

Everyday conversation on implantable hardware

Now

Observed trajectory

Leading indicators

Why do we think this?

Plausible next milestone

Bottlenecks

Depends on

Key uncertainties

1–2 years

Observed trajectory

Leading indicators

Why do we think this?

Plausible next milestone

Bottlenecks

Depends on

Key uncertainties

Observed trajectory

Leading indicators

Why do we think this?

Plausible next milestone

Bottlenecks

Depends on

Key uncertainties

Observed trajectory

Leading indicators

Why do we think this?

Plausible next milestone

Bottlenecks

Depends on

Key uncertainties

Scenario envelope

Observed trajectory

Leading indicators

Why do we think this?

Plausible next milestone

Bottlenecks

Depends on

Key uncertainties

3–5 years

Observed trajectory

Leading indicators

Why do we think this?

Plausible next milestone

Bottlenecks

Depends on

Key uncertainties

Scenario envelope

Observed trajectory

Leading indicators

Why do we think this?

Plausible next milestone

Bottlenecks

Depends on

Key uncertainties

Observed trajectory

Leading indicators

Why do we think this?