A few years ago, I helped build a system at Instagram that could notify hundreds of millions of people almost at once.
Most of the interesting engineering was invisible. There were queues, databases, networks, and backups for when things failed. But all of it depended on something more basic: electricity.
Computers turn energy into decisions.
The human brain does this using about twenty watts. Modern AI systems can require data centers with power measured in megawatts or even gigawatts.
This is not a fair one-to-one comparison. A data center does different work from a single brain. Still, the difference is large enough to be interesting. It suggests that our machines are wasting energy somewhere.
For a long time, programmers did not have to care much about this.
As transistors became smaller, computers became faster without using proportionally more power. Software could become slower and more complicated, while new hardware covered the cost.
That stopped working around 2005. Processor clock speeds stopped rising quickly. Since then, most improvements have come from using more processors at once.
At a large enough scale, computing becomes an energy problem. You are no longer asking only how many servers you need. You are asking how much electricity a building can receive and how much heat it can remove.
Why are computers so inefficient?
One reason is that they keep computation and memory separate.
A processor often spends less energy performing an operation than retrieving the data needed for it. The brain is organized differently. A synapse stores information close to where that information is used.
A second reason is that computers do too much work.
Traditional chips are controlled by a clock. Large AI models often activate most of their parameters for every piece of input. The brain is more selective. Most neurons are inactive most of the time. Work happens when an event requires it.
A third reason is precision.
Computers were designed to produce exact and repeatable answers. But neural networks often do not need exact arithmetic. They can work with low-precision numbers and tolerate noise.
So we have built approximate models on top of hardware designed for exact computation. That may be an expensive mismatch.
This does not mean future computers should copy the brain
Airplanes do not flap their wings. But they still obey the same physical laws as birds. In the same way, efficient computers may not look biological. They may simply follow some of the same principles:
- Keep data near the place where it is used.
- Do work only when something changes.
- Use only as much precision as the problem requires.
We do not yet know what the best architecture will be. It may involve specialized chips, analog computation, event-driven systems, or something that has not been invented.
But the direction is clear.
For decades, transistors were cheap and programmer time was expensive. It made sense to waste computation in order to make software easier to build.
Now energy is becoming one of the main constraints.
That will change what we optimize for. It will change our hardware first, then our software, and eventually the abstractions programmers use.
The brain does not tell us exactly how to build the next computer. It tells us that much greater efficiency is possible.
A working example already exists.
It uses about twenty watts.