For investors and funders · September 2026

Back a brain that learns by itself.

BrainCore is building a complete human brain in software, region by region, from published neuroscience. It is at research stage: pre-revenue, one founder, a validated foundation. This page is written for the reader who will check every number.

At a glance

The company on one screen.

GoalA brain that sees, hears, remembers and learns from its own experience, as a child does. No language model, no pretrained weights, no backpropagation.
StageResearch, pre-revenue. Stage 1 of 5 complete: Paper 1, vision and hearing, frozen on 11 September 2026 (tag sensory-report-2026-09-11).
BuiltRegion core; 23 acceptances against published papers, 11 green and 12 red; a 79-service anatomical atlas; fixed-size tissue storage with lossless packing.
CodeGo. About 49,000 lines in the V3 region core, 74 test files, 167 connectome and acceptance fixtures, 20 source audits. Private; open to reviewers and partners.
TeamVitalii Cherepanov, founder. Engineer, fifteen years of production software.
LocationNovi Sad, Serbia. Today a sole proprietorship; the round funds an operating company, jurisdiction to agree with the lead investor.
Raising€500 K pre-seed (equity), 24 months. Separately, a €150 K research grant, 12 months.
Contact[email protected]
Thesis

Intelligence you grow, rather than download.

Every large AI system today is trained once on a fixed corpus and then frozen. It cannot learn from its own experience without being retrained, and it cannot tell what it has learned from what it has merely read.

A brain does the opposite. BrainCore’s bet is that the most direct route to machines that learn continuously, know what they do not know and can be audited is to build the organ itself, from the part lists neuroscience has already published, and to hold every part to its paper.

Learns without retraining

Local plasticity at each synapse, all the time, from its own senses. There is no training run to repeat.

Honest by construction

Knowledge exists only as learned synapses. What was never learned cannot be recalled, so it cannot be invented either.

Auditable to the bit

A SHA-256 over every spike, the same bits on any number of cores or processes, tests preregistered in git before the run.

Native to new hardware

Spiking point neurons are exactly the workload neuromorphic chips are built for. Those chips exist; brain-grade models for them are rare.

Built on public science

Two dozen published region models already compose into sensory paths. The parts list is public; the assembly and the grading are the asset.

Measured, not promised

Twelve of 23 acceptances are red and published as red. Investors see the failures before they see the pitch.

Pros and cons

What you are buying, and what it does not have yet.

Every pro carries a label: measured today, built into the design, or still only a goal. How it differs from an LLM, how it will learn, and whether the brain will have the same problems are on the home page.

Pros
  • Learns on site, without retraining. goal · Paper 2
  • Knows only what it has seen: the origin of everything it knows is on record. by design
  • No third-party training data, so no dispute over rights to it. by design
  • Bit-exact: one SHA-256 on 1 core or 12, 16 of 16 digests from a clean copy. measured
  • Checkable part by part: 23 acceptances against published papers. measured
  • Installs at the customer: ordinary processors, no GPU, no cloud, no outside calls. measured
  • What it has learned can be copied and versioned as a tissue file. by design
  • Honest reporting: 12 of 23 acceptances are red and published as red. measured
Cons
  • It does not speak or write. Today it tells apart four letters and four tones.
  • Learning in V3 is not shown yet; that is Paper 2.
  • It cannot take ready-made knowledge from books or the internet: everything is learned from experience, and experience today runs slower than real time (2.7 s of model in 4.5 s).
  • Scale: 179,298 neurons against 86 billion in a human.
  • On processors it does not save energy. Spiking chips could change that, but BrainCore has not run on one yet.
  • 12 of 23 region models do not yet pass their acceptances.
  • No validation on disease data for medicine or pharma.
  • No revenue, no customers, one founder.
Why now

Four facts from the last two years.

  • Hardware for spikes exists; brain-grade models for it are scarce.

    Intel, 17 April 2024

    Intel’s Hala Point supports 1.15 billion neurons and 128 billion synapses on 1,152 Loihi 2 processors at a maximum of 2,600 W.

  • The largest brain-reconstruction programme has handed over.

    Blue Brain Project · Open Brain Institute

    EPFL’s Blue Brain Project ran from 2005 and ended in December 2024. Its work continues as the non-profit Open Brain Institute, launched in 2025.

  • Buyers already pay for brain-like compute.

    IEEE Spectrum, 3 June 2025

    Cortical Labs’ CL1 holds 800,000 lab-grown human neurons and sells for $35,000 a unit, or $20,000 in 30-unit racks.

  • Europe funds this category.

    EBRAINS, June 2025

    SpiNNcloud, maker of SpiNNaker2 neuromorphic systems, secured €10 M of blended grant and equity funding from the EIC Accelerator.

Market

Two markets: one still being defined, one already paying.

Analyst forecasts for neuromorphic computing in 2030 differ by a factor of sixteen: a sign of a young market, and of room for the company that defines its software. In-silico drug discovery and clinical trials are markets that pay today.

$1.3 Bneuromorphic computing in 2030, up from $28.5 M in 2024. MarketsandMarkets
$20.9 Bneuromorphic computing in 2030, up from $4.99 B in 2023. Next Move Strategy Consulting
$2.7 Mfrom the Gates Foundation to Numenta in 2024 to test brain-based AI. Semafor
$6–7.2 Bin-silico drug discovery by 2030, up from $3.3–3.6 B in 2024; two estimates. Market Glass · Next Move
$5.93 Bin-silico clinical trials by 2031, up from $3.87 B in 2025. Mordor Intelligence
$2.56 Bthe average cost of bringing one new drug to market, in 2013 dollars. DiMasi et al. 2016, J Health Econ 47:20

Industries, in the order we expect

Installed at the customer, not in the cloud

Ordinary processors, no GPU, no cloud, no calls to outside services: BrainCore installs on a company’s own servers, and patient, production or customer data never leaves the company’s perimeter. Proven on a laptop; packaging for enterprise installation does not exist yet.

HorizonIndustryWhat they pay forNeeded first
2026–2027Computational neuroscience labs and universitiesRegion models graded against their papers; reproducible simulation; runs certified by digest; teaching labsDocumentation for outside users; design partners
2027–2029PharmaModels of neural circuits in disease and under drugs (basal ganglia and dopamine, hippocampus, serotonin and noradrenaline nuclei) on the company’s own servers. US law since 2022 counts computer modeling as a nonclinical testA partner and validation on drug and patient data
2027–2029Medicine and medtechSound coding for hearing implants, vision models for prostheses, teaching modelsA manufacturer partner, clinical validation
2027–2028Neuromorphic hardware makersBrain-grade spiking workloads and reference models for their chipsA first region on a chip
2028 onIndustry and roboticsA learning core: learn a new object, defect or sound on site, with no cloud and no retrainingPaper 2, live senses
2028 onRegulated industriesA component that logs every event and repeats its result bit for bit; the AI Act requires logs and transparency from high-risk systems from 2 December 2027Useful decisions; a conformity assessment

BrainCore is pre-revenue. None of these industries has paid yet; the round exists to learn which pays first. Sources for the law and the AI Act are on the home page; every quote is in SOURCES.md.

Business model

The core is open. The brain built on it is paid.

An open core earns trust and users; revenue comes from certified regions, custom models, installation at the customer and support. The ways in, in the order we will test them.

WhenSourceWho paysFormBefore the first payment
NowGrants and joint projectsFunders, EU programmes, institutesNon-dilutive: EIC Pathfinder Open up to €4 M per project, ERC Starting up to €1.5 MSubmit; the €150 K grant is ready
2027Regions and supportLabs, institutes, R&DThe core is open; certified regions, new regions and support are paidRelease the core; first users
2027Installation at the customerPharma, medicine, industryAnnual site licence and support; data never leaves the company’s serversEnterprise packaging; a first pilot
2027–2028Custom modelsPharma, medtechProject: build and certify a circuit for the customer’s questionOne published example
2027–2028Hardware partnershipsNeuromorphic chip makersPorting contract, licence for reference workloadsA first region on a chip
2028 onA learning core for devicesRobotics, device makersSDK licence or per-device royaltyPaper 2, live senses

Sources: EIC Pathfinder · ERC Starting Grant. No prices until the first paying user.

Open core

The region engine, the acceptances and the digest are open: anyone can check a result and build a region of their own. Everything on top is paid: certified regions, custom models, installation and support.

Data stays with the customer

For pharma and medicine this is the entry ticket: the model runs on their servers, and no showing and no patient record ever leaves.

The money goes far

5 days from the V3 plan to a frozen Paper 1, one founder, one laptop. A checkable result every quarter.

Landscape

Everyone else builds a tool, a chip or a model. We build the organ.

WhoWhat it isSubstrateRelation to BrainCore
Large language model labsPretrained transformer modelsGPUs in data centresTrained once and frozen. BrainCore uses none of them, not even as a teacher.
NEST, Brian 2Open-source spiking simulatorsCPU, HPCTools, not a brain. NEST is BrainCore’s numerical reference.
Blue Brain → Open Brain InstituteDetailed reconstructions of rodent brain tissue; a non-profit platform since 2025HPCTwo decades of public science, ended December 2024. BrainCore targets the human atlas with point neurons.
Thousand Brains Project (Numenta)Sensorimotor learning from a theory of the neocortexSoftwareTheory first. BrainCore is anatomy first, graded region by region.
Intel Hala Point, SpiNNcloudNeuromorphic hardwareSiliconChips built for spikes. They need models like BrainCore’s.
Cortical LabsLiving human neurons on a chipBiologyBiological tissue, not software; cannot be copied, versioned or audited bit for bit.
BrainCoreThe human brain in software, region by region from published models, each graded against its paperCPU today; spikes suit neuromorphic chipsLearns only from its own experience. Bit-exact, preregistered, red results published.
Traction

What exists today.

23acceptances against published papers: 11 green, 12 red
79services in the anatomical atlas, 15 built all-green
16 / 16digests reproduced from a clean copy of the committed tree
5 daysfrom the V3 plan, 6 September, to a frozen Paper 1, 11 September

An 11-page preprint with 53 references checked against PubMed and Crossref is written; the venue is being decided. The evidence behind every claim on this site is on the home page and in Paper 1.

Not yet: revenue, customers, a public repository, learning in V3. BrainCore is one person, the founder.

Milestones for the round

Twenty-four months, a checkable result every quarter.

Targets, not promises, for one founder working at the pace that produced Paper 1: five days from plan to a frozen paper. Each milestone gets its acceptance written before the work starts, and is reported whether it passes or fails.

MonthMilestoneDone when
3Live sensesA letter on paper in front of the camera is read in ≥ 75 % of showings; a tone from a phone lights its column of area 41 in ≥ 8 of 10; picture and sound files reproduce the digests of the fixture runs
6Paper 2: letter ↔ sound learned through the hippocampusThe hippocampal loop passes 23 of 23 first; the preregistered rule holds on 20–50 seeds; unlearned letter, shuffled pairing and plasticity-off each break it
9The thalamus in the sensory pathsAn LGN relay between the eye and area 17, and SOC, IC and MGN as their own regions; relay transfer inside the measured 0.30–0.62; reading and hearing still pass their acceptances through the relays
12A brain across machinesThe same digest on four machines as on one; at least one million neurons in a single run
15Shape and actionAreas 18 and 19 read the four letters away from the centre and at other sizes, ≥ 75 %; area 4 writes a letter in spikes and the examiner reads it
18A first concept learned from experienceAn object, its sound and its written name joined on one loop, with ablations that break the link
21Half the atlas40 of the 79 services built, each passing the acceptance of its paper (15 today)
24Open platform and partnersCore and fixtures released; three design partners among labs or hardware makers; ready for a seed round
The ask

Two ways in.

Pre-seed · equity
€500 K24 months

SAFE or priced round; terms to agree with the lead investor.

  • One founder, with external scientific review and independent replication in the budget
  • Stages 2–4: live senses, Paper 2, a brain across machines
  • First design partners and a seed round at month 24
Research grant
€150 K12 months

For foundations, public programmes and research partners who fund science rather than equity.

  • One question: can this brain learn an arbitrary sight–sound association through its own hippocampus?
  • Preregistered, with three ablations and an independent replication
  • Published whether it passes or fails

Pre-seed: use of funds

LineWhat it buysShare
FounderFounder salary, 24 months180,00036 %
ComputeFour high-memory machines for the brain across machines, and cloud time for 20–50-seed sweeps150,00030 %
OperationsCompany formation, legal, accounting, intellectual property40,0008 %
Science and validationPart-time review by an external computational neuroscientist, independent replications, open-access fees, conferences90,00018 %
ReserveContingency40,0008 %
Total500,000100 %

Research grant: work packages

PackageMonthsWorkDone when
WP1 · Hippocampal recall1–3Close the six red criteria of the loop: CA3 recall from half a cue, four CA1 criteriaLoop acceptance 23 of 23
WP2 · Cortex ↔ hippocampus3–7Spikes from areas 17 and 41 into entorhinal cortex; the path back from CA1 to cortexThe recalled pattern changes the evoked signature of the target field
WP3 · The association6–10An arbitrary table of four letters and four tones; three of four trained; recall both ways4 of 4 forward and back on 5 seeds, then 20–50 seeds; all three ablations break it
WP4 · Release and replication9–12Tagged tree, fixtures and digests released; an external researcher reproduces the result from the tagPaper 2 preprint and a replication report

Research grant: budget

LineDetail
Principal investigator12 months full time66,000
Computational neuroscientistReview and co-design, 6 person-months30,000
ComputeOne high-memory server (512 GB or more) and cloud seed sweeps30,000
PublicationOpen-access fees, preprint, two conferences10,000
Independent replicationContract with an external researcher8,000
AdministrationLegal, accounting, audit6,000
Total150,000
Risks

What could go wrong, and what we do about it.

  • Learning is not yet shown in V3.

    Paper 1 is perception with plasticity off.

    Paper 2 is the test, preregistered with a plasticity-off control. The grant and the first six months of the round are built around it, and the answer is published either way.

  • Scale and energy.

    179,298 neurons today; a human brain has 86 billion.

    Regions are services with a bit-exact exchange, so the brain spreads over machines without changing a spike. Milestone at month 12: one million neurons on four machines. At human scale on ordinary processors that means petabytes of memory and thousands of servers, with their power bill. Spiking chips could bring it down; BrainCore speaks their signal but has not yet run on one.

  • Fidelity is not intelligence.

    A faithful brain may be slow to become useful.

    Every stage is a publishable, checkable result on its own, and the research-platform segment pays for fidelity before capability.

  • Published parts do not always compose.

    The geniculate relay transmits 0.124 against a measured 0.30–0.62.

    That is measured, diagnosed and published. The acceptance protocol finds such failures early instead of hiding them inside a demo.

  • One founder.

    Key-person risk.

    Everything is reproducible from tagged commits by someone other than the founder. The round pays an external neuroscientist to review the work, and the grant pays for an independent replication.

Next step

Ask for a live run.

The sensory paths run on a call and reproduce to the digest while you watch. Acceptance reports and code access for due diligence on request.

This page describes goals, plans and a proposed use of funds. They are intentions, not current capabilities, and may not be achieved. It is not an offer to sell, or a solicitation of an offer to buy, any security. External figures are quoted from the linked sources as of September 2026.