The essentials in 30 seconds ⚡
Researchers have developed a nanometre-scale laser that could, in time, allow chips to transmit information using light rather than electrical currents. The stated aim: faster computers, with energy consumption potentially halved. This is a piece of fundamental research, whose industrial application will take years.
Researchers have developed a nanometre-scale laser that could, in time, allow chips to transmit information using light rather than electrical currents. The stated aim: faster computers, with energy consumption potentially halved. This is a piece of fundamental research, whose industrial application will take years.
We regularly devote our environmental articles to data centre consumption, from their escape from the grid to the county that hosts 250 of them. Here is a lead from the other side of the problem: consuming less rather than producing more.
Where the energy really goes
One imagines that most of a chip's energy goes into computing. That is becoming less and less true.
A considerable — and growing — share goes into moving data: between memory and processor, between the components of a single board, between the machines of a single centre. Every transfer through a copper wire produces heat, which then has to be dissipated, consuming yet more.
We have come across this from several angles. The bottleneck in inference is memory access speed, as we explained in our article on dedicated chips. The advantage held by certain manufacturers lies in their internal interconnects, as we noted regarding the agreement between Qualcomm and Amazon, which specifically mentioned optical connections.
Why light
Light carries information with far less loss than electricity through a wire. That is why long-distance networks have used optical fibre for decades.
The challenge has always been scaling down to the size of a chip. You need tiny, efficient light sources that are compatible with existing manufacturing processes. A laser small enough to be integrated as close as possible to the circuits is a missing building block.
The timeline, which must be kept in mind ⏳
Between a laboratory result and an industrial product there is generally a decade: reproducing the result reliably, adapting it to production lines, reducing its cost, integrating it into architectures that were not designed for it. The "potentially halved" figure is a projection over a distant horizon, not a reduction that is available. This announcement should be read as a credible research direction, not as a solution to the current problem.
Between a laboratory result and an industrial product there is generally a decade: reproducing the result reliably, adapting it to production lines, reducing its cost, integrating it into architectures that were not designed for it. The "potentially halved" figure is a projection over a distant horizon, not a reduction that is available. This announcement should be read as a credible research direction, not as a solution to the current problem.
The efficiency paradox
There is a caveat that applies to any improvement of this kind, and it is an old one.
In the 19th century, the economist William Stanley Jevons observed that improving the efficiency of steam engines had not reduced coal consumption: it had increased it, because coal became affordable for new uses.
The same mechanism is at work here. A chip that is twice as efficient will not halve the sector's consumption: it will make profitable uses that were not profitable before. This is exactly what we are seeing with the price of tokens, which keeps falling while total consumption explodes.
What to take away
This research deserves attention because it targets the right place: the movement of data, which is the least well understood and fastest-growing source of consumption.
It also deserves a degree of scepticism about what it promises. Technical efficiency only produces a fall in consumption if something limits demand — a price, a cap, a rule. Without that, it mainly produces more usage.
This is why environmental solutions in this sector will never be purely technical. We noted as much regarding the vote obliging data centres to pay for their connections: what changes trajectories are, first and foremost, the rules that decide who pays.