Biological computing hardware: what you can actually buy, rent or collaborate on

There are exactly three ways to get neurons acting as a compute substrate today: buy an integrated appliance, rent remote access to someone else’s cultures, or assemble the capability yourself from commodity HD-MEA hardware and cells you source. Only the first two have published access models, and only one has a published price.

This page sets out what each route gives you, what it costs where that is public, and — because it matters more than the hardware — what the published results do and do not establish.

The landscape

Route one: buy the appliance

The CL1 is the only integrated biological computing appliance sold as a product. Its distinguishing characteristic is not the neurons — you can culture neurons — it is that life support, recording, stimulation and the runtime are inside the box.

What the vendor and reporting establish:

• Self-contained. Recordings, applications and life support are all on the device; no external compute required.
• Neuron viability up to six months, maintained by an internal life support system supplying nutrients, controlling temperature, filtering waste and maintaining fluid balance.
• Closed loop. The Biological Intelligence Operating System runs a simulated world, feeds information to the neurons and updates that world from their responses.
• Extensible. USB and ports for cameras, actuators and external systems; a touchscreen for status and pre-packaged assays.
• Low power. A rack of CL1 units is reported to draw 850 to 1,000 watts.

Two constraints buyers should know before budgeting. Cortical Labs requires customers to secure ethical approval to generate cell lines, and buyers of physical units are expected to have a suitable cell culture laboratory. The appliance removes the electrophysiology burden; it does not remove the institutional and biological one.

Route two: rent remote access

This is the more interesting route for most engineering groups, because it removes the wet lab entirely.

Cortical Cloud is browser-accessible with no installation, uses standard data-science tooling including Jupyter notebooks, and deploys custom code to an array of CL1s through a Python SDK. It is network-enabled by design so external sensors can feed the system in real time. The vendor does not publish a rate on its own page; the $300 per unit per week figure comes from IEEE Spectrum’s reporting.

FinalSpark’s Neuroplatform was the first remote-access biocomputing platform and has the fullest published technical description of any system here, because the team published the architecture in Frontiers in Artificial Intelligence. That paper is the most useful document in this entire field for a prospective user, and its numbers are worth knowing:

• Four MEAs, each hosting four forebrain organoids, with eight electrodes per organoid — 32 electrodes per MEA.
• Air-liquid interface: organoids sit directly on electrodes atop a permeable membrane with medium flowing beneath in a 170 µL chamber. The team states this avoids the need for dedicated coating and is less prone to organoid detachment than submerged approaches.
• Recording at 30 kHz, 16-bit, 0.15 µV accuracy, via Intan RHS 32 headstages. Stimulation from 10 nA to 2.5 mA.
• Closed-loop microfluidics supplying medium continuously at 15 µL/min, with automated bubble and overflow detection.
• Per-MEA 12.3-megapixel cameras for monitoring necrosis, displacement, medium acidity via phenol red colour analysis, contamination and overflow.
• Optical uncaging of glutamate, NMDA or dopamine via a 365 nm fibre-coupled LED.
• Over 1,000 organoids used and more than 18 terabytes of data collected over three years at the time of publication; the current site states over 30 TB, offered free to researchers.

Note the electrode arithmetic honestly: 8 electrodes per organoid is two orders of magnitude below an HD-MEA. What FinalSpark sells is not resolution. It is uptime, longevity and access — organoids maintained for months, available 24/7, without you owning an incubator. For a machine learning or robotics group, that trade is often exactly right.

Route three: build it yourself

The self-assembled route is what every published academic result actually used, and it is worth understanding what it entails before dismissing it.

The DishBrain work used the MaxWell MaxOne — 26,000 platinum electrodes over 8 mm² — with cultures from both rodent primary neurons and human iPSC-derived neurons. Notably, the team wrote their own software layer, “DishServer”, replacing and extending MaxWell’s vendor server software so it could receive voltage readings, process them, simulate a virtual environment, encode results as electrode commands and send them back. The published Pong task delivered inputs to a predefined sensory area of 8 electrodes.

That is the honest cost of this route: an instrument, a cell culture capability, and a software engineering project. What you get in return is total control of substrate, protocol and closed-loop logic, which is why every novel result in the field has come from groups that took it.

If you are considering this, the MEA platform compatibility matrix sets out electrode counts and plate formats across all the major systems, and MEA-ready neurons covers the cell side, including the awkward fact that the three-to-seven-week maturation from vial to a network worth recording from is unpriced and entirely your risk. For groups with no wet lab at all, the honest comparison is between building a culture facility and contracting the culture work out — see the contract services taxonomy and assay-ready plating, which covers the case where the buyer has no cold chain at all.

Choosing between the three

What the published results actually establish

This is the section that matters most, because the vocabulary in this field runs well ahead of the evidence and a purchasing decision made on the vocabulary will disappoint.

What is established. Cultured neuronal networks on multi-electrode arrays, placed in a closed loop with a structured environment, change their activity in ways that improve task performance relative to controls. The DishBrain work reported learning within minutes of gameplay in a simulated Pong environment, with cultures self-organising in response to sensory feedback. Separately, brain organoids coupled to high-density arrays have been used as reservoirs in a reservoir computing architecture, with reported performance on speech recognition and nonlinear mapping benchmarks. Both are real, peer-reviewed results.

What is not established. None of this is general-purpose computation, and none of it is competitive with silicon on any conventional benchmark. Reservoir computing extracts useful dynamics from a complex system without programming it; closed-loop adaptation demonstrates plasticity in a task context. Neither demonstrates that neurons will compute an arbitrary function you specify. The efficiency arguments — the very low power draw relative to GPU training — are real as physics and speculative as engineering, because nothing has been run at a scale where the comparison is meaningful.

Where the honest near-term value is. Two places, and they are not “replace your GPUs”.

1. Disease modelling and pharmacology in a closed loop. Cortical Labs is explicit that it views the CL1 as a platform for drug discovery and disease modelling, targeting conditions like epilepsy and Alzheimer’s, and reports work using an in vitro epilepsy model in which antiepileptic compounds improved learning-like performance in impaired cultures. Reading behaviour rather than just electrophysiology is a genuinely new assay axis.
2. Basic research into biological learning. Having a controllable, instrumented, long-lived neural network in a closed loop is a research instrument regardless of whether it ever computes anything commercially useful.

If you are buying against argument (1) or (2), the case is sound. If you are buying against the proposition that neurons will shortly replace accelerators, the evidence is not there and no vendor on this page has published data that would support it.

Ethics and approvals: not optional

Every route here involves human-derived neural tissue, and the approvals are part of the project plan, not an afterthought.

• Cell line provenance and consent. Cortical Labs requires customers to secure ethical approval to generate cell lines. Donor consent scope also propagates to derived products — consent to tissue donation does not imply consent to sequencing or commercial derivation, and restrictions attached to specific donor lineages flow silently into downstream cells.
• Institutional review. Most institutions treat human neural organoid work as requiring review even where no identifiable donor data is involved. Start that process early; it is frequently the long pole.
• The field’s own governance conversation. The organoid intelligence literature, including the Frontiers in Science programme article and the associated Baltimore declaration, sets out an explicit embedded-ethics framework for this research. Whatever your position on the underlying questions, reviewers and funders increasingly expect you to have read it.

What to ask before committing

• Substrate provenance. Which cells, from which donor background, under what consent, and what are you permitted to publish or commercialise?
• Channel arithmetic. Not electrode count — how many channels record simultaneously, and how many can be independently stimulated? The DishBrain team’s experience of routing 8 rather than 32 independent stimulation electrodes is the cautionary case.
• Latency, end to end. Closed-loop work lives or dies here. The CL1 is reported at sub-millisecond, down from 5 ms in DishBrain.
• Culture lifetime and what happens at end of life. Six months on the CL1; over 100 days published for FinalSpark organoids. What is the replacement cost and downtime?
• Data ownership and egress. Who owns the recordings, in what format do you get them, and can you take them with you?
• Reproducibility. Batch-to-batch variability between cultures is the field’s central methodological problem. Ask what the vendor does about it and what variance you should expect.
• Exit. If you start on cloud access and later need on-premises hardware, does the code port?

How we can help

We track this market because it sits directly on top of the cell supply chain we already broker. Where the answer is remote access, we will tell you that and save you a capital project. Where you genuinely need your own instrument and your own cells, we source the cells, normalise the electrophysiology hardware quotes, and can broker the culture work for groups with no wet lab. Tell us what you want the neurons to do, what latency the loop needs, and whether you have culture capability or intend to have none.