Biocomputing vs Silicon: The Energy Efficiency Case

Biocomputing could ease AI’s energy problem because living neurons compute with extraordinary efficiency: the whole human brain runs on about 20 watts, while one high-end data-center GPU can draw up to 700 watts. Neurons store and process information in the same place and learn from little data, but today’s biological systems remain far from matching silicon.

Key takeaways

The human brain runs on roughly 20 watts with about 86 billion neurons, while the IEA estimated that data centers used around 415 TWh of electricity in 2024, about 1.5% of global demand.
A single NVIDIA H100 SXM GPU has a configurable maximum thermal design power of up to 700 W, around 35 times the brain’s entire budget.
The brain’s efficiency comes from sparse spiking, in-memory computation at synapses, massive parallelism, and sample-efficient learning.
Biocomputing and organoid intelligence try to tap these properties directly, while neuromorphic chips imitate them in silicon.
Honest caveats matter: life-support overhead, slow speed, noisy responses, short lifespans and tiny scale mean no organoid system currently beats a GPU on any practical task.

Why is AI’s energy use a problem?

Artificial intelligence has become one of the fastest-growing sources of electricity demand in computing. The International Energy Agency (IEA), in its 2025 report Energy and AI, estimated that data centers worldwide consumed about 415 terawatt-hours (TWh) of electricity in 2024, roughly 1.5% of global electricity use. The IEA projected that data-center demand could more than double by 2030, to around 945 TWh, with accelerated servers for AI a major driver.

Two phases of AI consume energy. Training builds a model by running enormous numbers of calculations over large datasets, sometimes for weeks across thousands of chips. Inference is the cost of using the model afterward, every time someone asks it a question. As AI is deployed more widely, inference adds up. Strubell, Ganesh and McCallum drew wide attention to the energy and carbon costs of training large language models in a 2019 paper at the Association for Computational Linguistics conference, and the models have grown substantially since.

The energy demand is not just an environmental concern. It shapes where data centers can be built, how much AI costs, and how quickly it can scale. That is why researchers are looking at alternative computing approaches, including biocomputing, that might do more with less.

How much power does the human brain use compared with a GPU?

The brain is the obvious benchmark. In a 2021 PNAS perspective titled “Brain power”, Vijay Balasubramanian discussed the brain’s roughly 20 watt power budget. Raichle and Gusnard, also in PNAS (2002), noted that the brain accounts for about 20% of the body’s resting energy consumption while representing only about 2% of its mass. Azevedo and colleagues (2009) estimated that the human brain contains about 86 billion neurons.

Compare that with modern AI hardware. NVIDIA’s specification page for the H100 Tensor Core GPU lists a maximum thermal design power (TDP) of up to 700 W, configurable, for the SXM form factor. That is one chip. Large AI training clusters use thousands to tens of thousands of such GPUs, plus networking, storage and cooling.

A back-of-envelope comparison

Some simple arithmetic, using the sourced numbers above, puts the scales in perspective:

One H100 SXM GPU at full TDP versus one brain: 700 W divided by 20 W is 35. A single high-end GPU can draw as much power as about 35 human brains.
Power per neuron: 20 W divided by 86 billion neurons is roughly 0.23 nanowatts per neuron. This is a crude average that also pays for glial cells, blood flow and housekeeping, but it shows how little power each unit uses.
Data centers versus brains: 415 TWh per year corresponds to an average power draw of about 47 gigawatts (415 TWh divided by the 8,760 hours in a year). That is the equivalent of more than two billion brains running at 20 W.

These comparisons are not like for like. A GPU performs dense, precise arithmetic that brains do poorly, and a brain does many things no GPU can. The point is not that a brain is “35 GPUs” but that biology reaches sophisticated perception, learning and control on a power budget that would barely run a laptop.

Energy per operation or energy per task?

Engineers usually quote chip efficiency as operations per joule, or its inverse, energy per operation. That metric works well when comparing two chips that run the same software. It breaks down when comparing silicon with neurons, because a spike is not a floating-point operation and a synapse is not a multiply-accumulate unit. Attempts to convert one into the other depend heavily on assumptions about what a neuron “computes”.

A more meaningful measure is energy per task: how many joules a system needs to reach a defined level of performance on a defined problem, including the energy spent learning it. By that measure the brain’s advantage is less a matter of cheap arithmetic and more a matter of needing far less of it. This is also the only measure that will let biocomputing prove itself, since it captures both the efficiency of the hardware and the efficiency of the learning.

Most of the brain’s budget is ongoing activity

One subtlety from the physiology literature is worth knowing. Raichle and Gusnard argued in their 2002 PNAS paper that the brain’s energy use rises only modestly when a person performs a demanding task compared with resting quietly. Most of the budget supports continuous, intrinsic activity. In other words, the brain is always “on”, and its efficiency comes from how much it achieves within a nearly constant power envelope, not from switching off between jobs.

Why are living neurons so energy efficient?

Several design features explain the brain’s efficiency. None of them is magic, and some are already being copied in silicon.

1. Sparse, event-driven signaling

Neurons communicate with brief electrical pulses called action potentials, or spikes. Most neurons are silent most of the time, firing only when they have something to signal. Attwell and Laughlin’s 2001 energy budget for brain grey matter, published in the Journal of Cerebral Blood Flow & Metabolism, found that signaling (action potentials and synaptic transmission) accounts for most of the energy used in grey matter. Because spikes are costly, evolution appears to have favored codes that use as few of them as possible.

Conventional digital chips, by contrast, are clocked: large parts of the circuit switch on every cycle whether or not there is useful work. Spiking neural networks and neuromorphic hardware try to capture the brain’s event-driven style.

2. Memory and computation in the same place

In a standard computer, the processor and memory are separate. Data is fetched, processed, and written back, over and over. John Backus described the resulting limit in his Turing Award lecture, “Can programming be liberated from the von Neumann style?”, published in Communications of the ACM in 1978; it is now known as the von Neumann bottleneck.

The energy cost of this shuttling is large. In a 2014 keynote at the IEEE International Solid-State Circuits Conference, “Computing’s energy problem (and what we can do about it)”, Mark Horowitz showed that fetching data from off-chip memory costs orders of magnitude more energy than performing an arithmetic operation on it. For many AI workloads, moving weights and activations dominates energy use.

Brains do not have this problem. Information is stored in the strengths of synaptic connections, and those same synapses do the computing when signals pass through them. Memory and processing are physically the same thing. This is the biological version of what engineers call in-memory or near-memory computing. When learning changes a synapse through synaptic plasticity, it updates the stored “weight” right where it is used.

3. Massive parallelism at low speed

Neurons are slow. They signal on a millisecond timescale, while transistors switch in well under a nanosecond. The brain compensates with parallelism: billions of neurons operate simultaneously. Running many slow elements at once can be more energy efficient than running a few fast ones, because power in electronic circuits rises steeply with switching speed and voltage.

4. Analog, noise-tolerant computation

Neural signals are analog in many respects, and biological networks are tolerant of noise and the failure of individual components. Digital chips spend energy guaranteeing that every bit is exactly right. For tasks like perception, where approximate answers are fine, this precision may be unnecessary.

5. Local learning rules

Training a deep neural network typically uses backpropagation, which computes error signals across the whole network and requires storing intermediate values. Biological learning instead relies largely on local rules, such as Hebbian learning and spike-timing-dependent plasticity, where each synapse adjusts based on the activity of the neurons it connects. Local rules avoid much of the data movement that global training requires.

What is sample efficiency and why does it matter for energy?

Energy per operation is only half the story. The other half is how many operations you need to learn something.

Sample efficiency describes how much data or experience a learner needs to reach a given level of performance. In a widely cited 2017 article in Behavioral and Brain Sciences, “Building machines that learn and think like people”, Brenden Lake and colleagues argued that people can often learn new concepts from just a few examples, and that deep learning systems of the time typically needed far more data and experience to reach comparable performance on tasks such as recognizing handwritten characters or playing video games.

If biological networks are more sample efficient, the energy advantage compounds: fewer examples means fewer training passes, which means less computation overall.

What experiments in a dish suggest

Early experiments hint at this property in cultured neurons. In the DishBrain study, Kagan and colleagues reported in Neuron (2022) that cultures of human and mouse neurons on a high-density electrode array improved their performance at a simulated game of Pong within minutes of gameplay, using feedback designed around the free energy principle. In the Brainoware study, Cai and colleagues reported in Nature Electronics (2023) that an organoid used as a reservoir for reservoir computing could support speech-sound classification and that its responses changed with training. Our article on reservoir computing with brain organoids explains the setup.

These are small, specific results. They show that learning-like changes can be induced quickly in tissue, not that organoids are generally more sample efficient than AI. Measuring sample efficiency fairly, with matched tasks, inputs and baselines, is still an open research problem. For the mechanics of training cultures, see how neurons learn in a dish.

How does silicon compare with biological computing?

The table below summarizes the key differences between conventional silicon, neuromorphic silicon and biological computing with neurons. It reflects general properties, not benchmarks.

Property	Conventional silicon (CPU/GPU)	Neuromorphic silicon	Biological neurons (organoids, cultures)
Basic unit	Transistor	Silicon neuron and synapse circuits	Living neuron and synapse
Signaling speed	Nanoseconds or faster	Typically designed for real-time, event-driven operation	Milliseconds
Memory and compute	Separate (von Neumann)	Co-located on chip	Co-located at synapses
Precision	Exact, deterministic	Low precision, often deterministic	Analog, noisy, variable
Learning	Global training (e.g. backpropagation), usually offline	Offline training or local on-chip rules	Native local plasticity
Power of the compute element	Hundreds of watts per high-end GPU	Very low for suitable workloads	Very low for the tissue itself
Overhead	Cooling, power delivery, networking	Similar to other chips, smaller	Incubation, gas, media exchange, sterility, electronics
Reliability and lifespan	Years, highly reliable	Years	Weeks to months, variable between samples
Scale today	Billions of transistors per chip, huge clusters	Up to millions of neurons per chip	Organoids with far fewer neurons than a brain
Programmability	Mature software ecosystems	Specialized toolchains	Early-stage protocols and tools

Where neuromorphic chips fit

Neuromorphic engineering borrows biology’s principles without the biology. IBM’s TrueNorth chip, described by Merolla and colleagues in Science (2014), implemented a million spiking neurons on a single chip with very low power consumption. Intel’s Loihi, described by Davies and colleagues in IEEE Micro (2018), added on-chip learning. Neuromorphic computing is far more mature than biocomputing and is manufacturable at scale, which makes it a serious competitor for low-power AI. Biocomputing’s distinctive bet is that real neurons offer learning capabilities and adaptability that are hard to reproduce in silicon.

What are the honest caveats?

The efficiency argument for biocomputing is real but often overstated. Anyone evaluating the field should keep these caveats in mind.

Life support costs energy

Neurons themselves use tiny amounts of power, but keeping them alive does not. An organoid system needs an incubator held at body temperature, controlled gas levels, regular replacement of culture medium, sterile handling, pumps or microfluidics, and the electronics for stimulation and recording. Producing the culture media and growing the organoids also has an energy and resource footprint. Today, the overhead is far larger than the power used by the tissue. Any fair comparison must count the whole system, just as GPU comparisons should include cooling.

Speed

Neurons work on millisecond timescales. For tasks that require billions of precise operations per second, such as rendering graphics or multiplying large matrices, silicon is vastly faster. Biocomputing could only make sense for workloads where adaptive learning matters more than raw speed.

Reliability and reproducibility

Biological responses are noisy and drift over time. Organoids grown from the same protocol vary in size, composition and activity. A computer whose behavior changes as it ages, and that differs from its neighbor, is hard to build products on. Researchers address this with careful quality control, electrophysiology benchmarks and good data practices, but it remains a core challenge.

Lifespan

A GPU can run for years. Organoids in culture survive for months, and their properties change as they mature. Jordan and colleagues’ 2024 description of the FinalSpark Neuroplatform in Frontiers in Artificial Intelligence showed that organoids can be maintained and recorded remotely over long periods, but replacing and recalibrating living tissue is a fundamentally different maintenance model from swapping hardware.

Scale and interfaces

Organoids contain many orders of magnitude fewer neurons than the brain’s 86 billion, and lack blood vessels, which limits growth (see organoid vascularization). Interfaces are also narrow. A microelectrode array can address thousands of points, but that is a tiny fraction of the connections inside the tissue. The brain’s efficiency is a property of an enormous, vascularized, mature organ with sensory input. It is not guaranteed that a millimeter-scale organoid captures it.

No current performance advantage

To be direct: no published biological system currently beats a GPU, or even a modest CPU, on a practical computing task, in speed, accuracy or energy per useful result. The 2023 Frontiers in Science paper by Smirnova and colleagues that defined organoid intelligence framed efficiency as a long-term research goal, not a present-day achievement. Any claim otherwise should be treated with skepticism.

Where could biocomputing realistically help first?

Given those caveats, where might living neurons make sense? Researchers have suggested a few directions:

Adaptive, low-data learning tasks, where a system must adjust continuously to changing inputs and labeled data is scarce.
Reservoir computing, where the tissue provides a rich, nonlinear dynamical system and only a simple digital readout is trained, keeping digital computation small.
Hybrid systems, in which biological tissue handles a narrow function and conventional chips do everything else.
Research into efficient learning itself, using organoids to study the mechanisms that make biology efficient, which could then inspire better algorithms and neuromorphic chips.

The last point may be the most important in the near term. Even if biocomputers never replace silicon, understanding how living networks learn so efficiently could change how silicon systems are designed. Our article on the future of biocomputing looks at this roadmap in more detail.

How should we measure biocomputing energy efficiency fairly?

A credible efficiency claim needs a clear method. A good comparison should:

Define the task precisely, with the same inputs, outputs and success criteria for both systems.
Count total system energy, including life support for biology and cooling for silicon.
Account for training energy as well as operating energy.
Report variability across multiple organoids or cultures, not just the best one.
Compare against a reasonable silicon baseline, not an inefficient one.
Share data and code, ideally in standard formats such as Neurodata Without Borders, so others can check the result.

These standards are not yet routine in the field. Building them is part of making biocomputing a real engineering discipline.

The bottom line on biocomputing vs silicon

The case for biocomputing energy efficiency rests on solid biology. The brain does extraordinary things on about 20 watts by computing with sparse spikes, storing memory where it processes, running in parallel and learning from little data. Silicon AI, by contrast, is consuming a growing share of the world’s electricity, with a single high-end GPU drawing up to 700 watts.

But today’s brain organoids are not miniature brains, and today’s biocomputers are research instruments, not competitors to GPUs. The honest version of the argument is that living neurons are an existence proof of efficient learning and a promising, uncertain path toward it. Turning that promise into measurable gains will take better tissue, better interfaces, and rigorous, reproducible measurement. Making those experiments easier to design, run and compare is what we are building toward at Organoid Programming; you can read about the platform or start with our complete guide to organoid intelligence.

Frequently asked questions

How much power does the human brain use?

The human brain uses roughly 20 watts, about the same as a dim light bulb, while accounting for around a fifth of the body’s resting energy use despite being about 2% of body mass. This figure, discussed in sources such as Balasubramanian’s 2021 PNAS perspective and Raichle and Gusnard’s 2002 PNAS paper, covers all of its sensing, thinking, memory and control functions.

How much electricity do data centers use?

The International Energy Agency’s 2025 report Energy and AI estimated that data centers worldwide consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of global electricity use. The IEA projected that demand could more than double by 2030, with AI-focused servers a major driver of the growth.

Is biocomputing more energy efficient than GPUs?

Not today in any practical sense. Biological neurons use very little energy per signaling event, and the brain as a whole is remarkably efficient, but current organoid systems need incubators, pumps and electronics that use far more power than the tissue itself, and they solve only simple tasks. The efficiency argument is about long-term potential, not current performance.

Why is the brain so energy efficient?

Several features contribute: neurons communicate with sparse, brief spikes instead of constant activity; memory and computation happen in the same place, at synapses, so data is not shuttled back and forth; signals are analog and tolerant of noise; and computation is massively parallel. The brain also learns from relatively little data, reducing the cost of training.

What is the von Neumann bottleneck?

The von Neumann bottleneck is the limit created when a computer’s processor and memory are physically separate, so data must constantly move between them. In modern chips, moving data often costs much more energy than computing with it, as Mark Horowitz described in a 2014 ISSCC keynote. Brains avoid this by storing and processing information in the same synaptic connections.

What is sample efficiency and why does it matter for energy?

Sample efficiency is how much data or experience a learning system needs to reach a given level of performance. Humans can often learn a new concept from a handful of examples, while deep learning systems typically need far more. Fewer training examples means fewer computations, so a more sample-efficient learner could reduce the energy cost of training, not just of running a model.

How does neuromorphic computing compare with biocomputing?

Neuromorphic chips such as IBM’s TrueNorth and Intel’s Loihi copy the brain’s design principles, like spiking and local memory, in silicon. They are reliable, manufacturable and already very low power for some workloads. Biocomputing uses living neurons directly, gaining native plasticity but facing problems of life support, variability and lifespan. The two approaches are complementary research paths.

Biocomputing vs silicon: why living neurons could fix AI's energy problem

Why is AI’s energy use a problem?