← Back to blog

RESEARCH

Will AI Ever Be as Energy-Efficient as the Human Brain?

August 1, 2026 · 5 min read

Will AI Ever Be as Energy-Efficient as the Human Brain?

The human brain operates using approximately 20 watts of power. It uses a smaller amount of energy than a device for charging a laptop, but this power is sufficient so that the organ manages how people see, speak, remember and move throughout the day. By contrast, a single instance when a person trains an artificial intelligence system can require gigawatt-hours — this amount of energy is equal to the total electricity that a utility facility produces over a period of many hours or multiple days.

WHAT 20 WATTS ACTUALLY RUNS

VisionHearingLanguageLong-term memoryMotor control & balanceEmotion regulationHeartbeat & breathingContinuous learning

All of it, at once, continuously — on roughly 20 watts, about a fifth of your body's resting energy budget. That figure comes from measured cerebral metabolism, not from a model.

AI will certainly get more efficient over time — that part isn't in question. What's worth asking is whether it can ever reach a state where it performs tasks with as little waste as a brain does. To achieve that, specific technical components of the systems would need to be different than they are now.

The difference in operation

There is also a significant difference when the systems are in operation. If a person runs a model that is already trained to provide a single response, the process consumes much more energy for each "thought" than a brain consumes to perform a similar task. When researchers calculate exact ratios, the results vary based on the specific tasks that they compare. As shown by most formal calculations, the difference is between four and six orders of magnitude. The brain is not slightly more efficient. It is approximately 10,000 to 1,000,000 times more efficient for every unit of calculation that is performed.

CONTINUOUS POWER DRAW

Human brain20 W
One AI GPU1,400 W

A single accelerator draws 70× what your entire brain does, continuously.

Brain at 20 W continuous. GPU figure is the NVIDIA Blackwell Ultra B300 per-GPU TDP.

Why the gap exists

As engineers who work with silicon attempt to close this distance, multiple structural differences remain.

Memory and computation are separate in silicon, combined in biology. In silicon, units for memory and units for computation are separate, but those functions are combined in biological systems. Because of this separation, a GPU must move data between the memory and the processing components at all times. It is this movement of data that consumes the largest portion of energy, rather than the mathematical calculations. In a brain the synapse is the location where both memory and computation occur. There is no physical data bus that requires a constant supply of information.

Neurons fire sparsely; GPUs compute densely. Neurons operate with low frequency and low density, while GPUs operate with high density. When a neural network runs on a GPU, it performs calculations on dense matrices. By this method every unit performs an operation during every cycle, regardless of if the specific task requires it — but biological neurons respond to specific events. They produce signals only when a stimulus occurs, which means that most parts of the brain are not active at any single moment. To compute in a sparse and asynchronous manner is naturally less costly than to compute in a dense but also synchronous manner.

Exactness is expensive. For digital chips, the default state is to calculate with a high degree of numerical exactness, which is a process that requires significant resources. If we look at neurons, they are analog devices that function with noise and a low level of exactness. It is evident that many intelligent processes do not require the specific level of exactness that digital hardware currently provides.

What's actually being tried

This isn't a hypothetical problem — there's real hardware research aimed directly at it:

  • Neuromorphic chips (Intel's Loihi 2) implement spiking, event-driven computation in silicon, showing large efficiency gains on the workloads they're suited for.
  • Near-memory and in-memory compute (IBM's NorthPole, memristors, resistive RAM) attack the memory/compute separation directly, doing the arithmetic where the data already lives.
  • Sparsity and mixture-of-experts models activate only a fraction of a network's parameters per query, borrowing the brain's "only fire what's needed" strategy at the software level.
  • Lower-precision arithmetic (int8, fp8, and lower) trades numerical exactness for a meaningful cut in energy per operation, with surprisingly little accuracy loss for many tasks.
  • Photonic computing uses light instead of electrons for matrix multiplication, which in principle could cut the marginal energy cost of computation dramatically.

Where this lands

In the next ten years, it is likely that hardware will lower the energy difference for specific tasks by 100 to 1,000 times. By using neuromorphic systems and in-memory computing, researchers are showing that those gains are possible now. To close the entire gap and run general purpose intelligence on 20 watts is a separate challenge. And it is probable that this goal requires hardware materials that are fundamentally different from current chips, instead of just being more effective versions of a graphics processing tool.

It is necessary to note that the definitions of success change over time. When silicon technology nears the goals, the systems described in "AI" will likely be different from current large scale models. Because the ability of a system to perform and its energy use change at the same time, those factors are connected.

For a business that creates small and functional software, this perspective is relevant to more than just hardware. The most effective solution is the one that performs the fewest operations to reach a result — this principle is true if the component is a biological cell, a physical circuit or a sequence of instructions.

Sources