A system built from human brain cells learns to recognize speech using 90% less training time than a conventional silicon processor. This result, published in Nature Electronics by the University of Indiana team, is not a laboratory demonstration reserved for specialists: it describes a technology that, by late 2025, fits on a desk and is being sold.

The biocomputer is no longer a promise. It is a product. And the legal framework to govern it does not exist anywhere.

The Essentials

  • The University of Indiana demonstrated voice recognition with 90% less training time than a silicon equivalent, published in Nature Electronics.
  • Cortical Labs (Melbourne) delivered its first desktop biocomputer CL1 in late 2025; FinalSpark (Geneva) offers paid remote access to its organoids via its Neuroplatform.
  • According to estimates available in the literature, neural organoids consume between 100,000 and 1,000,000,000 times less energy than their silicon equivalents for comparable learning tasks, depending on the task and comparison performed.
  • No legal framework exists to define the status of an organoid, the ownership of data from a donor’s cells, or the consciousness thresholds above which protection applies.
  • Europe has a window of approximately eighteen months before this market becomes structurally difficult to regulate retroactively.

What Organoids Do That Silicon Cannot

A neural organoid is an aggregate of a few thousand to several million human neurons cultivated in a laboratory from stem cells. It does not resemble a brain. It lacks the brain’s structures, functions, or consciousness. But it shares with the brain a property that silicon chips cannot imitate: synaptic plasticity.

When an organoid receives a signal, its neural connections strengthen or weaken in real time. It learns, in the literal biological sense of the term, without requiring programming of this capacity. Silicon, meanwhile, simulates learning through massive mathematical operations. It does not learn: it calculates what a system that learns would learn. The difference is fundamental. It explains most of the energy gap.

A large language model like GPT-4 consumes, during training, several gigawatt-hours. An organoid, for a comparably complex representation task, consumes a few microwatts. The order of magnitude is not 10% or 50%: according to available estimates — FinalSpark cites a factor of one million, Johns Hopkins a range from one million to ten billion, and the original Nature Electronics publication citing 20 watts for the human brain against 8 million watts for a comparable artificial neural network — the energy advantage is between several hundred thousand and several billion times depending on the task considered. In the specific case of voice recognition tested at Indiana, the 90% reduction in training time is accompanied by an energy footprint incomparable to current systems.

It is not that generative AI models are inefficient in an absolute sense. It is that they do something fundamentally different, and biology offers a parallel path for certain learning tasks that do not require the raw power of GPUs. The question of AI’s energy consumption has become central to the debate on its long-term viability, and organoids constitute the first serious biological response to this challenge.

Cortical Labs and FinalSpark: From Laboratory to Catalog

Cortical Labs is an Australian company founded in 2019 in Melbourne. Its CL1, delivered to its first customers in late 2025, is a desktop device containing living human neurons cultured on an electrode chip. These neurons receive electrical signals, respond, and their response is read in real time by the system. The entire machine fits in a housing comparable to a compact workstation computer. It requires a controlled temperature and nutrient environment, but it does not require a clean room, supercomputer, or industrial infrastructure.

The selling price is not public. The existence of the product is.

FinalSpark, a Swiss company based in Geneva, chose a different model. Its Neuroplatform is not sold: it is subscribed to. Clients access living organoids remotely hosted in the company’s facilities via a digital interface. They send computational tasks to it, receive results, and never see the cells. It is a cloud model applied to living matter.

These two commercial trajectories say something important about the maturity of the sector. We are no longer talking about isolated experiments published in academic journals and forgotten three years later. We are talking about supply chains for stem cells, large-scale culture protocols, customer support, contracts. The industry exists. It is small, but it is real.

Other players are advancing in parallel. Johns Hopkins is working on more complex organoids to test neurological drugs. The University of Graz in Austria is exploring real-time neural interfaces. Startups in South Korea and China are publishing results on biological-silicon hybrid systems. The field is becoming structured, with different rhythms depending on the country and regulatory questions that vary considerably.

The Legal Vacuum Is Not a Technical Detail

Here is the question that no one in European, American, or international law has yet resolved: at how many neurons does an organoid acquire a particular status?

The current answer is: we do not know, and existing texts do not answer this question. The European AI Act, which entered into force in 2024, regulates artificial intelligence systems in the computational sense of the term. It says nothing about biological systems that produce intelligence without being software. The Oviedo Convention on Biomedicine protects persons, not cellular aggregates. The directive on personal data protection covers data from identifiable persons, but what does it cover when data is generated by cells from an anonymous donor and transformed into computational signals?

These questions are not rhetorical. They have immediate contractual implications. When a FinalSpark client sends a task to an organoid and receives a result, who owns that result? The company, which hosts and maintains the cells? The client, who formulated the task? The donor, whose DNA structures the neurons used? The question of intellectual property over the productions of a biological system has no clear answer in any current jurisdiction.

Even deeper is the question of moral status. Current organoids have no central nervous system, no thalamus, no integrated cortex. The dominant scientific consensus is that current organoids are not conscious, do not feel anything, and do not suffer — but this consensus is not unanimous: a significant minority of researchers, including recent publications in Cell Patterns and PMC (2025), contest this rejection as premature and call for caution. This question concerns current systems, but it is all the more open because more complex, more integrated organoids, equipped with structures closer to a functional brain, are on the roadmaps of several laboratories. The demarcation line between “cellular aggregate” and “entity to protect” is not defined, and no one is seriously working to define it before the question becomes urgent.

The regulation of generative AI has shown the limits of the reactive approach: when rules arrive after large-scale deployment, they chase accomplished facts that have shaped markets, habits, and power relations. Biocomputing is following exactly the same trajectory, with one difference: the questions it raises are not only economic or security-related, they are anthropological.

Europe Has Eighteen Months to Not Repeat Its Usual Mistake

The European Union has a history with transformative technologies. It often arrives late to creation, but sometimes ahead on regulation. The AI Act was adopted before the United States had an equivalent federal framework. The GDPR reshaped data protection practices worldwide far beyond European borders. This model has its merits and limitations: it produces global standards, but it can also slow European actors while their competitors operate without constraints.

On biocomputing, Europe starts from an interesting position. FinalSpark is Swiss, at the frontier of European law. Several world-class university laboratories are working on organoids in Germany, Austria, the Netherlands, and France. Fundamental research is solid. What is missing is regulatory coordination.

The window is narrow for a simple reason: companies commercializing biocomputers today are building contractual precedents, business models, and market expectations. In eighteen months, if Cortical Labs’ CL1 has been adopted by enough research institutions and pharmaceutical laboratories, regulating it will become as difficult as regulating Uber after five years of operation in every major city. Technical and commercial accomplished facts create political resistance to change that does not yet exist.

What Europe could concretely do: define a framework for classifying organoids according to their neurological complexity, establish consent rules for cell donors used in commercial systems, and create a provisional intellectual property regime for data generated by biological systems. None of these projects is technically beyond reach. They demand political will and coordination between regulators of biomedicine, data law, and intellectual property that does not yet exist.

What Energy Efficiency Changes in Geopolitical Balances

The energy argument deserves to be taken seriously, not as a technical curiosity, but as a strategic variable.

Generative AI in its current form is so power-hungry that it is reshaping national electrical policies. Microsoft, Google, and Amazon have announced massive investments in new energy generation capacity, including nuclear, to power their data centers. The energy constraint of AI is not a peripheral issue: it touches the balance between cheap energy-producing countries and countries seeking to develop an AI industry without the electrical base for it.

Organoids do not solve this problem entirely. They will not replace large language models for tasks requiring raw power and massive corpora. But they open a path for specific, repetitive learning tasks that do not need the generality of a GPT-4 and that currently consume a disproportionate fraction of global computational resources. Pattern recognition, anomaly detection, certain classification tasks, reinforcement learning in constrained environments: all these applications could migrate to biological substrates without performance loss, with massive energy gains.

For a country like France, seeking to position its AI industry while managing real electrical constraints despite nuclear power, or for Germany, needing to reduce its industrial emissions, biocomputing is not an academic curiosity. It is a strategic option that deserves to enter national research and innovation roadmaps.

The Questions Science Has Not Yet Resolved

It would be misleading to present biocomputing as a mature technology whose only delay is the legal framework. Several technical barriers remain.

The lifespan of organoids is limited. Cells age, synaptic connections degrade, nutritional support systems are complex to maintain at industrial scale. FinalSpark has solved this problem for its own use, in its facilities, with its own protocols. Scaling that to an industry is another matter. Standardization of culture protocols, reliability of electronic interfaces with living neurons, reproducibility of results across different cell batches: these are open engineering problems.

Programmability is also an outstanding question. An organoid learns in a plastic manner, but you cannot yet give it precise instructions the way you program a microprocessor. Research on bidirectional neural interfaces is advancing, notably thanks to Cortical Labs’ own work and academic teams in Europe and the United States. But fine control of an organoid’s computational behavior remains incomplete.

Finally, the question of interpretability arises with even greater acuity than for artificial neural networks. When an organoid produces a result, that result can be measured. Understanding why it produced it, tracing the path in synaptic connections that led to that output, is currently impossible. For critical applications, in medicine or infrastructure, this opacity is a serious obstacle.

These limitations are real. They do not negate the results obtained. They set the horizon for the coming decade: not the replacement of silicon, but the coexistence of different computational substrates, each adapted to specific classes of problems.

The real question for the next five years is not whether neural organoids work. The results from the University of Indiana and Cortical Labs’ products have answered that question. The real question is the one posed by all technologies that progress faster than institutions: who decides the rules of the game, and when? Europe has a chance to answer before the answer is imposed upon it by facts.

Sources

  1. TechXplore / Science Alert — Scientists use human brain cells for computing: https://techxplore.com/news/2025-12-scientists-human-brain-cells.html
  2. Nature Electronics — University of Indiana study on voice recognition with organoids (Brainware)
  3. Cortical Labs — CL1 presentation (cortical.com)
  4. FinalSpark — Neuroplatform, commercial access to organoids (finalspark.com)
  5. Regulation (EU) 2024/1689 — AI Act, Official Journal of the European Union
  6. Original Brainoware publication – Nature Electronics: https://www.nature.com/articles/s41928-023-01069-w
  7. Cortical Labs – Wikipedia: https://en.wikipedia.org/wiki/Cortical_Labs
  8. FinalSpark – Official Neuroplatform Site: https://finalspark.com/neuroplatform/
  9. FinalSpark – BusinessWire press release: https://www.businesswire.com/news/home/20240515701469/en/FinalSpark-Launches-the-First-Remote-Research-Platform-Using-Human-Neurons-for-Biocomputing
  10. PMC – Legal challenges of brain organoids: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11882709/
  11. Cell Patterns – Consciousness in HBOs: https://www.cell.com/patterns/fulltext/S2666-3899(25)00213-2
  12. BigGo Finance – CL1 shipping 2025: https://finance.biggo.com/news/202603010220_Cortical_Labs_CL1_Bio_Computer_Runs_Doom
  13. UOC – Neurotechnologies and European AI Act: https://www.uoc.edu/en/news/2026/neurotechnologies-european-legal-challenge