It is not a complete human brain
Around 800,000 cultured neurons versus roughly 86 billion in a human brain — five orders of magnitude, with no brain regions, senses or body.
BIOLOGICAL COMPUTING · TECHNICAL DOSSIER
CL1 Biological Computer
When life begins to compute.
Explore the technology that fuses living neurons with silicon: from electrical spikes to programmable biological neural networks, and how a next-generation hybrid computing system actually works.
01 DEFINITION
The CL1 is a biological computer from Cortical Labs (Melbourne): human induced pluripotent stem cell (iPSC) derived neurons are cultured on a microelectrode array. Software stimulates them, reads their activity back through the same array, and closes the loop in real time.
It is not a general-purpose processor, nor a server that simply “uses cells for compute”. A more accurate description: a programmable neurophysiology platform that treats a living neural network as its computational medium. The unit still contains full life support, data acquisition electronics and a software stack.
Biological computing is often sold as a “mini-brain”. Before the engineering details, the boundaries matter.
Around 800,000 cultured neurons versus roughly 86 billion in a human brain — five orders of magnitude, with no brain regions, senses or body.
Neurons have no instruction set, clock or deterministic addressing; they process information analogically through spike timing and synaptic weights.
Observable electrical activity, plasticity or even “learning” behaviour is not evidence of consciousness, sentience or self-awareness.
Stimulation, amplification, sampling, denoising, decoding, experiment control and storage are all handled by conventional electronics and software.
Biological networks rely on synaptic plasticity, homeostasis and network dynamics — no loss function, no gradient descent. The analogy is loose.
This dossier separates published specs, peer-reviewed findings, engineering estimates and undisclosed internals. Undisclosed parts are schematic only.
Founded in Melbourne, Australia, with the goal of turning living neurons into a programmable computing substrate.
A closed-loop system embodied cultured neurons in a simulated game world; performance improved within minutes. Published in Neuron.
Cortical Labs disclosed a rack-scale unit integrating neurons, life support and a compute interface.
First units shipped from summer 2025 at a public price of US$35,000, alongside cloud wetware access.
| Dimension | Biological computer (CL1 class) | Conventional computer | Artificial neural network |
|---|---|---|---|
| Computational medium | Living neurons + silicon interface | Silicon transistors | Floating-point matrix math |
| Information carrier | Spike timing, synaptic weights, field potentials | Binary voltage levels | Numeric tensors |
| Programmability | Shaped indirectly via stimulation, environment and reward — not instruction-level | Full instruction set, deterministic | Defined by code and training |
| Determinism | Low: biological variability and noise | High | Reproducible after training, near-deterministic on hardware |
| Power | Tens of watts per unit (reported) | Tens to hundreds of watts | Training needs large GPU clusters |
| Maturity | Early research platform, limited volume | Mature industry | Mature industry |
Wetware Computing
An umbrella term for using biological tissue (cells, neural networks) as the computing or information-processing medium, as opposed to hardware and software. It stresses the biological substrate; it is not “biologically inspired algorithms”.
Synthetic Biological Intelligence
A term used by Cortical Labs for cultured neural networks that can be shaped through external interfaces and show adaptive behaviour in a closed loop. It blends engineering and marketing; its scientific boundaries are still debated.
“It is a human brain in a box”
No. It is a dispersed cultured network with no brain regions, senses, body or global organisation.
“The cells write code and run LLMs”
No. Tasks are defined by biOS and external software; neurons provide a shaped dynamical response.
“Biological computing has replaced GPUs”
Far from it. Current scale, stability and programmability cannot support general workloads.
“Buy one and culture neurons at home”
Human cell culture involves biosafety, ethics review and specialist facilities. A home setting is not compliant.
This is an independent technical dossier, not affiliated with Cortical Labs. CL1 is a product name; trademarks belong to their owners.
02 SYSTEM
Five subsystems must work together: living medium, silicon interface, life support, electronics and software. If one fails, no usable data exists.
The medium itself: human iPSC-derived neurons that differentiate, mature and self-organise into a network.
Cultured neurons age, vary between batches and can be contaminated — a root cause of biological computing's instability.
The only physical interface between neurons and electronics: it converts ionic currents into readable electronic signals and delivers stimulation current back.
Public CL1 material cites a 64-channel read/stimulate interface at 25 kHz; some reports say 59 channels. The discrepancy is unresolved.
The environmental engineering that keeps neurons alive for months: temperature, gas, nutrients, pH and sterility. If one fails, the experiment ends.
Cortical Labs states its internal life support sustains a culture for up to about six months — far shorter than a server's service life.
It turns analog biological signals into a controllable, storable, reproducible data stream.
Public material does not disclose CL1's controller models or board architecture; descriptions here are general engineering inference.
biOS translates a “world” into stimulation, then translates neural responses into observable behaviour and data.
biOS is closed-source commercial software; this dossier does not claim its internals.
03 CLOSED LOOP
Click any step for detail, or run the loop to watch one full stimulate–respond–decode cycle. Every waveform and number is simulated in your browser.
04 TECHNICAL PRIMER
This chapter reads the CL1 on four layers: why neurons can compute, what hardware the unit contains, how a run actually operates, and where the technology may go. Every figure is tagged as published, peer-reviewed, an engineering estimate or unverified; undisclosed engineering detail is schematic only.
There is no instruction set. Information lives in membrane potentials, spike timing and synaptic weights, and processing is the superposition of physical and chemical processes. Four stages form the chain from ions to behaviour.
Membrane potential & action potential
A neuron stores ionic gradients as a transmembrane voltage (resting about −70 mV). When input drives the membrane past threshold, voltage-gated sodium channels open, producing a 1–2 ms action potential, followed by potassium-driven repolarisation. The spike travels along the axon and triggers transmitter release at its terminals.
Values are textbook ranges for mammalian cortical neurons. Peer-reviewed
Synapse & excitation/inhibition balance
A presynaptic spike releases transmitter that either depolarises the postsynaptic membrane (excitatory, AMPA/NMDA receptors) or hyperpolarises it (inhibitory, GABA receptors). Stable firing depends on excitatory/inhibitory balance: too little inhibition gives seizure-like bursts, too much makes the network unresponsive.
Mechanism is established neuroscience; actual weights depend on cell type and culture conditions. Peer-reviewed
Plasticity & STDP (not backpropagation)
Synaptic strength changes with history: long-term potentiation/depression (LTP/LTD), spike-timing-dependent plasticity (STDP) and homeostasis. If the presynaptic spike slightly precedes the postsynaptic one, the connection strengthens; reversed order weakens it. “Learning” in a closed loop is the superposition of these local rules and network dynamics.
Here “learning” means improved task metrics, not supervised learning. Peer-reviewed
Network dynamics & encoding
Hundreds of thousands of neurons form a recurrent network with spontaneous oscillations, bursting and synchrony. Stimulation does not write an instruction; it perturbs a dynamical state, and readout decodes population rate, local field potential or synchrony. This is why identical stimulation rarely yields bit-identical responses.
Encoding remains an open research question; there is no agreed “neural instruction set”. Peer-reviewed
Signal chain: from ionic current to a digital action
| Stage | Physical quantity | Typical magnitude | Engineering constraint |
|---|---|---|---|
| Ionic flow | Transmembrane current | pA – nA | Only indirectly observable with an MEA |
| Action potential | Extracellular voltage | 50 – 500 µV | Noise floor must sit below ~10 µV |
| Electrode interface | Impedance | 50 kΩ – 1 MΩ | Higher impedance amplifies thermal noise |
| Pre-amplification | Gain | 100 – 1000× | Common-mode rejection > 80 dB |
| Sampling | Sample rate | 20 – 30 kSa/s | Anti-alias filtering required |
| Stimulation | Current pulse | 1 – 100 µA | Charge balance plus artefact blanking |
| Decoding | Population features | 10 – 100 ms 延迟 | Must be shorter than the task time constant |
These are textbook and engineering magnitudes for generic in-vitro MEA systems. They explain the signal chain and are not published CL1 specifications.
A CL1 is not a single chip but an integration of biological, analog, digital and fluidic subsystems. The layers below run from wetware to software, followed by the unit-level figures that can be publicly checked.
| Layer | What it is | Function | Evidence |
|---|---|---|---|
| Wetware | Human iPSC-derived cortical neurons, ~800,000 per unit | Provides a plastic, recurrent computational medium | Official / public |
| Interface | Microelectrode array, 59–64 channels (public figures differ) | Bidirectional: record field potentials, inject current | Unverified |
| Analog front end | Low-noise amplifiers, filters, ADCs | Moves µV-level signals into the digital domain | Engineering estimate |
| Digital control | FPGA/MCU, clock synchronisation, protection logic | Schedules stimulation and acquisition, blanks artefacts | Engineering estimate |
| Life support | Medium loop, thermal control, CO₂/O₂ mixing, contamination monitoring | Keeps cells near physiological state for ~6 months | Official / public |
| Software | biOS simulated world, loop scheduling, logging and API | Defines tasks, encodes stimuli, decodes responses | Official / public |
“Not disclosed” means the vendor has not published the detail; this page does not infer it.
A run rests on three things: millisecond-scale loop control, a culture workflow measured in weeks, and quality metrics enforced throughout. If any one fails, the data is unusable.
Closed-loop timing budget
| Stage | What happens | Budget | Basis |
|---|---|---|---|
| World update | Advance the simulated world one tick | 1 – 16 ms | Engineering estimate |
| Encoding | Map state to stimulation parameters | < 1 ms | Engineering estimate |
| Injection | Current pulse delivered to electrodes | 0.1 – 1 ms | Engineering estimate |
| Neural response | Synaptic delay and conduction | 5 – 50 ms | Peer-reviewed |
| Acquisition | Amplify, sample, transfer | 1 – 5 ms | Engineering estimate |
| Decoding | Feature extraction and action selection | 1 – 10 ms | Engineering estimate |
| Total | One full loop | 10 – 100 ms | Engineering estimate |
A budget inferred from published sample rates and typical neural latencies to judge loop feasibility; it is not a measured CL1 latency.
Run lifecycle (six stages)
Thaw iPSCs, expand and run basic QC.
Days – 2 weeks
Induce cortical neurons; monitor morphology and markers.
3 – 6 weeks
Plate onto the MEA and form adhesion and an initial network.
3 – 7 days
Record spontaneous activity; screen units and channels.
1 – 3 days
Run closed-loop tasks and collect behavioural and neural data.
Days – weeks
Archive data, inactivate cells, clean the chamber and replace consumables.
1 – 2 days
Durations are engineering estimates for common iPSC neural workflows and vary widely by line and lab.
Key quality metrics
| Metric | Target range | If out of range | Basis |
|---|---|---|---|
| Electrode impedance | 50 – 500 kΩ | Higher noise, lower stimulation efficiency | Engineering estimate |
| Signal-to-noise ratio | > 4 | Single units become indistinguishable | Engineering estimate |
| Spontaneous rate | 0.1 – 10 Hz | Network too quiet or seizure-like bursting | Engineering estimate |
| Temperature | 37 ± 0.5 °C | Drift in dynamics, risk of protein denaturation | Engineering estimate |
| pH | 7.2 – 7.4 | Cell stress and death | Peer-reviewed |
| Osmolarity | 280 – 320 mOsm/kg | Cell swelling or shrinkage | Engineering estimate |
| Contamination | No bacteria/fungi | Whole run invalidated | Official / public |
Common control ranges for in-vitro neural culture; CL1's actual thresholds and calibration routine are not published.
Four typical failure modes
Blocks medium flow; local cells die of hypoxia and channel signals vanish suddenly.
Mitigation — Bubble traps, inline sensors, degassing before plating
Bacteria or fungi consume nutrients and shift pH; activity collapses within hours.
Mitigation — Aseptic technique, scheduled sampling, terminate on detection
Impedance drift or open circuits permanently lose channels, leaving gaps in stimulation coverage.
Mitigation — Redundant channels, impedance tracking, per-channel remapping
Different lines, passages and histories yield different dynamics; cross-batch comparisons distort.
Mitigation — Same-batch controls, randomisation, complete metadata
Separating what works now, what needs specific engineering breakthroughs, and what remains speculative is more useful than a generic roadmap. Each horizon lists its key gate and maturity.
| Horizon | Direction | Key gate | Maturity |
|---|---|---|---|
| Near · 1–3 yrs | Disease modelling, neurotoxicity screening, plasticity research, teaching | Batch reproducibility and standardised readouts | Peer-reviewed |
| Mid · 3–7 yrs | Embodied control, reservoir-computing benchmarks, bio-hybrid sensing | Stable sub-50 ms loops and long-term culture stability | Engineering estimate |
| Far · 7 yrs+ | Low-power edge computing, 3D organoid computing, personalised digital twins | 3D interfaces, programmability, ethics and regulation | Unverified |
No horizon implies replacing general-purpose computing: biological networks still trail silicon in numerical precision, addressability and reproducibility.
Figures in this chapter fall into four classes: published specifications, peer-reviewed consensus, engineering estimates, and undisclosed or unverified items. Unlabelled qualitative statements are explanatory and imply no performance claim.
05 TEARDOWN
Below is a rotatable, zoomable, exploded schematic. It explains system layering and part relationships; it is not CL1's actual engineering drawing.
Drag to rotate · Scroll to zoom
Illustrative model — not the actual internal structure of CL1
06 MANUFACTURING
Building one is not “putting chips and cells in a box”. Seven parallel engineering tracks converge into a single integration. Each phase below lists goals, work items and deliverables.
Any work with human cells requires biosafety level, ethics review, trained staff and compliant waste handling. This page explains principles, equipment needs and compliance only — never home culture protocols.
GoalProve cells differentiate, attach and form a recordable network reliably.
Deliverable / gateReproducible culture protocol, baseline viability data and batch release criteria
Key riskBatch variability, low differentiation efficiency, contamination
GoalBuild an MEA with suitable impedance, controlled noise and long soak life.
Deliverable / gateManufacturable MEA with impedance and noise test reports
Key riskDelamination, electrode corrosion and low yield under long soak
GoalHold a stable environment unattended for weeks to months.
Deliverable / gate30-day environmental stability log and contamination rate
Key riskBubbles, leaks, evaporation, thermal failure
GoalTurn the analog front end, timing and protection into a reliable unit.
Deliverable / gateElectrical test and 72-hour continuous run
Key riskNoise coupling, ground loops, timing jitter
GoalMake experiments programmable, reproducible and analysable.
Deliverable / gateReproducible experiment scripts, data format spec and regression tests
Key riskInsufficient real-time performance, data loss, irreproducibility
GoalMake biology, fluidics, electronics and software cooperate on one timeline.
Deliverable / gate7-day closed-loop run with recordable neural activity
Key riskAssembly contamination, interface mismatch, vibration
GoalTurn a one-off experiment into an operable device lifecycle.
Deliverable / gateMaintenance manual, spare-parts list and health dashboard
Key riskCulture expiry, contamination, spare-parts shortage
07 BILL OF MATERIALS
Browse components by category: function, principle, specifications, estimated price, alternatives and public usage. Prices are public-channel estimates, not quotations.
Items without reliable quotes or specs are marked estimated or unverified; no vendor data is fabricated.
08 SOFTWARE
Programming a biological computer is not writing instructions; it is designing input encoding, closed-loop policy and decoders. Three kinds of code are separated below: real hardware control, pure simulation, and the in-browser demo on this page.
The following is an interface sketch, not the official Cortical Labs SDK. Real hardware must use vendor drivers, current limits and safety procedures.
# Closed-loop pseudocode (interface sketch, not an official SDK)
from cl1_sdk import Device, MEA, LifeSupport # hypothetical interface
dev = Device.open("cl1-01")
mea = MEA(dev, channels=64, sample_rate_hz=25_000)
life = LifeSupport(dev)
with life.maintain(temperature_c=37.0, co2_pct=5.0, ph_range=(7.25, 7.40)):
mea.check_impedance(max_kohm=500)
for trial in range(120):
ball = world.next_state() # 1. world state
stim = encode_rate(ball, n_channels=16) # 2. state -> stimulation
mea.stimulate(stim, amplitude_ua=8, biphasic=True) # 3. inject charge
spikes = mea.record(duration_ms=50) # 4/5. record activity
action = decode_population(spikes) # 6/7. decode to action
world.apply(action) # 8. feedback to world
if score(action, ball) > 0:
apply_reward(mea, pattern=stim) # 9. reward shaping
# Real hardware must use vendor current limits and safety procedures.This code simulates a spiking network only. It never touches living cells, and its output says nothing about real biological experiments.
# Pure software simulation: LIF network + STDP (no living cells)
import numpy as np
N, dt, tau_m = 200, 0.1e-3, 20e-3
v = np.zeros(N); v_rest, v_th, v_reset = -65e-3, -50e-3, -70e-3
w = np.random.rand(N, N) * 0.3
for t in range(2000):
I = np.random.randn(N) * 0.9 # simulated input current
v += dt / tau_m * (-(v - v_rest) + I)
fired = v > v_th
if fired.any():
v[fired] = v_reset
w[:, fired] += 0.01 # toy STDP placeholder
w = np.clip(w, 0, 1)
print("This simulation says nothing about real biological experiments")Runnable closed-loop demo SIMULATED
A leaky integrate-and-fire network with a simplified STDP rule demonstrates a minimal stimulate → respond → decode → reward loop in your browser. Fully simulated.
09 LABORATORY
Behind one device sits an entire facility and governance layer. Functional zones, staffing, qualifications and compliance are outlined below.
Controlled clean environment, biosafety cabinet, CO₂ incubator, inverted microscope, centrifuge.
Engineering estimateBSL level per local regulation; negative pressure, HEPA, autoclave and compliant waste disposal.
Official / publicLithography/sputtering/bonding can be outsourced; in-house impedance testing, probe station and clean bench.
Engineering estimateOscilloscope, source meter, low-noise supplies, pre-compliance EMI and rework stations.
Engineering estimateHost computers, storage, remote access and scheduling; network and permission isolation.
Engineering estimateHigh-bandwidth storage, time sync, backup and integrity checks.
Engineering estimateSpares, calibration tools, cleaning and sterilisation flow, physically separated from culture.
Engineering estimateExact levels and approvals vary by jurisdiction; follow local regulators and institutional rules.
10 INDEPENDENT R&D
The goal is not to clone the CL1, but to build a modular testbed that validates the principles. Three tiers progress step by step, each valid on its own.
Entry / no biological material
Build a spiking network with LIF or Izhikevich neurons, simulating encoding, closed-loop decoding and STDP.
This tier involves no living cells and must not be marketed as a real biological computer.
Intermediate / interface & life-support emulation
Build “electronic neurons” in analog circuitry, emulate MEA stimulation and acquisition with multichannel DAC/ADC, and monitor temperature, humidity and fluid level.
This tier involves no living cells and must not be marketed as a real biological computer.
Professional / real neurons and MEA
Work with a qualified laboratory, using compliant cells, a real MEA, life support and bidirectional closed-loop software.
Budgets are engineering estimate ranges and vary widely with region, FX, used equipment and in-house work.
11 APPLICATIONS
Separating validated research uses from long-range speculation is essential to judging where this technology really stands.
Study plasticity, network dynamics and behaviour in controlled loops.
Observe drug effects on human neural activity patterns, complementing animal models.
Patient-derived iPSC networks for electrophysiological phenotypes in epilepsy and neurodegeneration.
Explore encoding, decoding and energy limits of living networks as a computational medium.
Map neural responses to control signals and study adaptation in embodied systems.
Compare energy scaling of biological and silicon approaches on narrow tasks.
Derive sparse, event-driven computing ideas from biological dynamics.
Not feasible today: scale, stability and programmability fall far short.
No scientific evidence; it should not be an engineering claim.
12 COST
The calculator below lets you edit quantities and unit prices, totals live, and switches between CNY and USD. Prices are midpoints of publicly visible ranges — engineering estimates, not quotations.
| Stage | Item | Qty | Unit price (editable) | Subtotal | Basis |
|---|
Items without formal quotes are inferred from public retail prices or comparable products; error may reach ±50%.
Over weeks or months, one connector, bubble or handling error can invalidate an entire run.
Microvolt signals coexist with millivolt stimulation; artefact blanking and noise control decide how much data is usable.
Batch, passage and history change network dynamics; designs need controls and replication.
The same stimulation may yield different responses across cultures; cross-lab replication remains a core challenge.
Cultures last months, so core biological components and consumables must be replaced periodically.
Human cells, data governance and cross-border material transfer add institutional and time cost.
13 LIBRARY
Filter by type, then open an entry for its abstract and source link. Links point to official pages, journal DOIs or vendor documentation.
Official product page: closed-loop system, integrated life support (~6 months), biOS simulated world and cloud access.
Open source ↗Company overview and research direction, including DishBrain background.
Open source ↗The DishBrain paper: cortical neurons show measurable adaptation in a closed-loop Pong task.
Open source ↗A roadmap for organoid intelligence covering opportunities, ethics and technical barriers.
Open source ↗A classic high-density CMOS MEA implementation, illustrating density, multiplexing and noise trade-offs.
Open source ↗A systematic review of CMOS MEAs interfacing electrogenic cells.
Open source ↗The foundational human iPSC work underpinning the cell-sourcing route used by CL1.
Open source ↗Representative brain organoid work showing the possibilities and limits of neural self-organisation.
Open source ↗Shows complex spontaneous oscillatory activity emerging in organoids as networks mature.
Open source ↗Search published families by assignee. Status and claims must be checked per document; no legal conclusion is drawn here.
Open source ↗Commercial high-density CMOS MEA platform commonly used for high-resolution neural recording and stimulation.
Open source ↗Classic 60/64-channel MEA systems and chips, suited to teaching and small-to-mid-scale experiments.
Open source ↗Multiwell MEA recording platforms widely used in drug screening and network analysis.
Open source ↗A Python spiking neural network simulator, suitable for the tier-A prototype on this page.
Open source ↗A simulator for large-scale spiking networks with multiple neuron and synapse models.
Open source ↗Foundational reference on neuronal electrophysiology and synaptic transmission for the signal chain described here.
Open source ↗Nothing saved yet. Use the “Save” button on any reference or term.
14 GLOSSARY
Search or browse by category. Each entry gives a concise definition and where it appears.
Concept Peer-reviewed
Computing that uses biological tissue as its medium.
Concept Official / public
Cortical Labs' term for adaptive behaviour shown by cultured neural networks.
Biology Peer-reviewed
Induced pluripotent stem cells, reprogrammed from somatic cells and differentiated into neurons.
Interface Official / public
Microelectrode array: a grid of electrodes for recording and stimulating electrogenic cells in vitro.
Software Official / public
Cortical Labs' Biological Intelligence Operating System.
Signal Peer-reviewed
Local field potential: low-frequency signal reflecting population synaptic activity near an electrode.
Signal Peer-reviewed
The extracellular signature of a neuronal action potential, a brief high-frequency event.
Signal Peer-reviewed
A cluster of synchronous spikes across neurons in a short window.
Learning Peer-reviewed
Spike-timing-dependent plasticity: synaptic strength changes with pre/post spike timing.
Learning Peer-reviewed
Long-term potentiation / depression: persistent changes in synaptic strength.
System Official / public
Stimulation, response, decoding and re-stimulation form a real-time loop.
System Engineering estimate
Delay from acquisition to feedback; it determines whether a loop is viable.
Engineering Peer-reviewed
Equal positive and negative charge per stimulus pulse, preventing electrode polarisation and cell damage.
Engineering Peer-reviewed
AC resistance at the electrode–electrolyte interface; affects noise and stimulation efficiency.
Engineering Peer-reviewed
Interference on recording channels during stimulation; removed by blanking or template subtraction.
Compliance Official / public
Laboratory containment levels classified by biological risk.
Compliance Official / public
Material Transfer Agreement governing biological material between institutions.
Model Peer-reviewed
Leaky integrate-and-fire, one of the most common simplified neuron models.
Biology Peer-reviewed
A three-dimensional miniature tissue self-organised from stem cells.
Business Official / public
Wetware-as-a-Service: remote cloud access to biological computing resources.
No. It contains roughly 800,000 cultured neurons with no brain regions, sensory input or body, and no evidence of consciousness.
No. Current scale, stability and programmability cannot support general training; it targets neuroscience research and closed-loop experiments.
Cortical Labs cites up to about six months with internal life support. Maintenance includes medium exchange, temperature and gas control, contamination monitoring, impedance tracking and calibration.
Neither advisable nor compliant. Human cells require biosafety, ethics review and specialist facilities. Individuals should start with tier A simulation or tier B electronics.
No. Every animated trace is simulated in the browser to explain mechanisms. Real data must come from controlled experiments and publications.
No. Task adaptation can be explained by low-level plasticity and network dynamics; it does not imply general intelligence or consciousness.