The NSF Bets on Human Neurons to Solve AI’s Energy Crisis

A medium-sized data center consumes between 5 and 20 MW, which represents the power supply of approximately 3,500 to 14,000 households — it is only for a large data center (50-70 MW) that the comparison with a city of 50,000 inhabitants becomes relevant. A human brain organoid, a few million neurons cultivated in a Petri dish, runs on a few milliwatts. The gap is so vertiginous that it has finally attracted serious money.

In 2025, the National Science Foundation committed $14 million to seven biocomputing projects. This is not symbolic funding for basic research. It is a signal that the field has crossed a threshold: the point where an academic idea begins to resemble an industrial bet. The question that opens up is no longer “can miniature brains compute?” It has become “at what price — energetic, ethical, social — are we willing to do it?”

The Essential Points

  • The NSF invested $14 million in seven biocomputing projects in 2025, signaling a shift from laboratory to industrial traction.
  • The Brainoware system, developed at Indiana University, demonstrated in 2023 in Nature Electronics the capacity of human brain organoids to recognize speech and solve differential equations.
  • A brain organoid operates on a few milliwatts, versus gigawatts for current AI infrastructure — a theoretical energy advantage of several orders of magnitude.
  • The field raises unprecedented ethical questions about the status of human nervous tissue used as a computing tool, which existing regulatory frameworks do not cover.
  • First commercial applications remain 5-10 years away, but several private laboratories began forming partnerships with academic teams in 2025-2026.

The Brain Consumes What GPUs Waste for Nothing

To understand why this bet now attracts public funding, we must first measure the scale of the problem it claims to solve. Electricity consumption by U.S. data centers is projected to double by 2030, according to projections from the International Energy Agency. Demand driven by AI is the primary driver. Each query to a large language model mobilizes hundreds of specialized chips for fractions of a second. Multiplied by billions of daily interactions, this equation begins to pose a problem of physics, not just economics. The race for AI is becoming an electrical battle that network operators are struggling to anticipate.

The human brain processes information using approximately 20 watts. It does so with an efficiency that the best current chips do not approach. This comparison has been known for decades, but it long remained in the register of impotent admiration. What is changing since 2022-2023 is that we now have brain organoids that are stable and complex enough to begin connecting them to real computing systems.

A brain organoid is a three-dimensional cluster of human stem cells reprogrammed to differentiate into neurons. It does not resemble an adult brain. It has no consciousness, no organized structure like a cortex. But it spontaneously develops active synaptic connections, forms networks, responds to stimuli. And most importantly, it learns — in a way that remains poorly understood, but is documented.

Brainoware: 2023, the Moment It Ceased to Be Science Fiction

The team led by Feng Guo at Indiana University published in December 2023 in Nature Electronics a demonstration that changed the tone of discussions in the field. Their system, called Brainoware, connected a human brain organoid to an electronic chip via an electrode network. The organoid received electrical signals encoding human speech sounds. After several training sessions, it was able to distinguish different speakers with a significant recognition rate. It was also capable of solving differential equations of the type used in physical modeling.

These performances remained modest compared to a dedicated algorithm running on a GPU. But they were not aimed at head-to-head competition. They were aimed at proof of concept: human nervous tissue can be integrated into a computing loop, receive structured inputs, produce interpretable outputs, and improve with exposure. This is the functional definition of a learning system.

What the article did not fully reveal is the colossal technical difficulty that precedes this moment. Organoids die. They do not behave reproducibly from batch to batch. The interface between living tissue and silicon poses biocompatibility problems that electrophysiology teams have worked for years to solve. Brainoware was not a turnkey solution. It was a demonstration that the fundamental lock was not physical, but engineering.

From Johns Hopkins to the NSF: Three Years to Move from Manifesto to Check

In 2022, a Johns Hopkins team published what the field retained as the first structured manifesto of biocomputing. The document, signed by a coalition of researchers, laid the conceptual foundations of a new discipline: organoid intelligence, or OI. It argued that brain organoids constituted a promising biological computing substrate, combining energy efficiency, learning capacity, and scalability potential. The tone was deliberately programmatic: the goal was to convince funders, not just peers.

Three years later, the NSF followed. The $14 million committed in 2025 covers seven distinct projects, addressing complementary aspects: improved organoid stability, development of neuro-electronic interfaces, design of algorithms capable of interpreting biological signals, and ethical modeling. The last component is not ornamental. The NSF integrated it into funded projects because the questions it raises condition the field’s trajectory.

This public funding comes in a context where several private laboratories have begun to take interest in the subject. Startups like Cortical Labs in Australia have developed hybrid neuron-silicon systems capable of playing simple games — Pong, notably — which made waves in 2022. The fact that the NSF is now entering the game with multi-year funding signals that the U.S. government does not want this field to develop without public oversight.

A Few Milliwatts Against Gigawatts: The Energy Advantage Is Real but Incomplete

The central argument for biocomputing is energetic. It deserves careful examination, as it is often presented too simply.

A brain organoid consumes a very small amount of energy to maintain its cellular activity. A training GPU (e.g., H100) consumes between 350 and 700 watts; a complete GPU server integrating eight of these chips can reach between 5,000 and 10,000 watts, sometimes up to 80 kW per rack in hyperscale configurations. The gap is real and it is massive. But it is incomplete for two reasons.

The first: the consumption of an organoid does not account for the entire infrastructure needed to make it function as a computing tool. Electrodes, amplifiers, signal reading systems, maintaining the tissue’s viability (temperature, CO₂, nutrients) have their own consumption. What is measured when we speak of a few milliwatts is the expenditure of nervous tissue alone, not the complete system.

The second: current organoids cannot do what GPUs do. They do not process matrices with billions of parameters. They adapt to specific tasks after training, but their generalization capacity is currently very limited. Comparing their consumption to that of a data center is comparing a calculator to a supercomputer.

What researchers in the field defend is an argument of trajectory, not immediate performance. Silicon has reached physical limits close to miniaturization. Biological computation, on the other hand, has barely begun to be optimized. The argument is not “the organoid beats the GPU today.” It is “the organoid can become competitive on adaptive learning tasks at a fraction of the energy cost, if interface and stability problems are solved.”

The Ethical Frontier That Committees Have Not Yet Drawn

This is where the subject changes nature. As long as we discuss energy efficiency and performance, we remain in the usual register of technological innovation. But biocomputing puts on the table a question that regulators had not anticipated: what is the moral status of human nervous tissue used as a computing tool?

Current organoids are not brains. They do not have the structures necessary for consciousness as we understand it. They do not suffer in the way an organism suffers. But they are made of human cells. They develop spontaneous electrical activity. And the more complex we make them to increase their computing capabilities, the wider this gray zone becomes.

The issue is not hypothetical. Alysson Muotri, neuroscientist at the University of California San Diego, published in August 2019 in Cell Stem Cell observations of electrical activity in brain organoids resembling oscillations recorded in premature infant brains. His work did not imply that organoids were conscious. But it was enough to trigger serious controversy in the bioethics literature. As organoids gain complexity for computational needs, this question will return with greater force.

Several researchers in the field, including members of the Johns Hopkins team, advocate for an ethical framework established upstream, before commercial applications create pressure to move quickly. NSF funding explicitly includes projects on this component. But to date, there is no specific regulation for organoids used for computing purposes, neither in the United States nor in Europe. Institutional ethics committees treat these projects under existing protocols for research on human stem cells, which were not designed for this situation.

This void is not merely a philosophical problem. It is practical. Companies that want to commercialize biocomputing systems in five to ten years will need to know what is allowed, what is forbidden, and who decides in case of dispute. The absence of a framework today is a source of legal and commercial risk for field development, not just a matter of principle. This is in fact one of the reasons why structured public funding, with safeguards, is better than exclusively private development advancing under the regulatory radar.

We find a similar pattern in other fields of synthetic biology, such as microrobots repairing the spinal cord: innovation runs ahead of frameworks, and it is often public money that finances both the research and the safeguards that allow it to be regulated.

What the Seven NSF Projects Really Seek to Resolve

The details of the seven projects funded by the NSF in 2025 are not all public, but the communicated research axes allow us to understand what the field considers priority.

The first challenge is stability. An organoid survives on average a few weeks to a few months under standard laboratory conditions. For a computing tool to be operational, much longer lifespans and reproducible behaviors are needed. Several projects work on improving culture media and bioreactors capable of maintaining functional organoids over periods of one to two years.

The second challenge is the interface. Connecting biological tissue to a digital system without destroying it, and reading its signals with sufficient precision to extract useful information, is a major engineering problem. Current electrodes read the activity of a few hundred neurons simultaneously. Researchers aim for tens of thousands for serious applications.

The third challenge is algorithmic. It is not enough for the organoid to produce signals. Methods are needed to interpret them, train them, and exploit them in computing loops. These are machine learning problems applied to a substrate whose behavior is fundamentally stochastic, requiring approaches different from training methods for artificial neural networks.

These three locks are serious. None seems insurmountable on a ten-year horizon, but none is solved today. What is underway is not a race toward an imminent product. It is the patient construction of a scientific foundation on which applications could one day rest.

The Horizon Is Ten Years Away, but Bets Are Being Placed Now

The history of AI itself teaches that fields that seem purely academic can shift quickly. Deep neural networks were a niche subject for twenty years before the combination of data and computing power made them operational. Biocomputing is not in the same situation: it is not waiting for data, it is waiting for stable organoids and reliable interfaces. But the logic of the shift is comparable.

What is happening in 2025-2026 looks less like an imminent revolution than the establishment of intellectual and financial infrastructure. Researchers working on these subjects publish, recruit, train doctoral students. Public funders enter. Startups watch. The first conferences specifically dedicated to biocomputing are taking shape. This is the phase where a nascent discipline decides whether it will exist or remain a laboratory curiosity.

For now, the trajectory leans toward existence. And if AI’s energy problem continues to worsen at its current pace, pressure on the field will only increase. This raises the most difficult question: if biocomputing delivers on its promises, who will decide under what conditions human nervous tissue can be used as a tool, for which applications, with what limits?

This is a question that the seven NSF projects cannot solve alone, and that no industrial actor has an interest in raising too loudly while the technology is not mature. Yet it will be posed — and it will be easier to answer if the debate has begun before the first commercial systems reach the market.

Sources

  1. Undark — Brain Organoids and Big Questions (2026): https://undark.org/2026/02/26/brain-organoids-big-questions/
  2. Cai H. et al., “Brain organoid reservoir computing for artificial intelligence” — Nature Electronics, December 2023 (Brainoware, Indiana University): https://www.nature.com/articles/s41928-023-01069-w
  3. Smirnova L. et al., “Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish” — Frontiers in Science, 2023 (Johns Hopkins): https://www.frontiersin.org/journals/science/articles/10.3389/fsci.2023.1017235/full
  4. International Energy Agency — Electricity 2024, datacenter and AI projections; see also: Energy and AI Report: https://www.iea.org/reports/energy-and-ai/executive-summary
  5. Cortical Labs — DishBrain, Neuron, 2022
  6. NSF — Official press release on $14M in biocomputing: https://www.nsf.gov/news/nsf-invests-14m-bioengineered-systems-ethical-biocomputing
  7. Cell Stem Cell 2019 — Muotri, organoid oscillations premature infants: https://www.scientificamerican.com/article/can-lab-grown-brains-become-conscious/
  8. Indiana University IU Impact Blog — Brainoware and Feng Guo: https://blogs.iu.edu/iuimpact/2023/12/15/human-brain-tissuebioengineers-are-building-the-intersection-of-organoids-and-ai/