An AI Winter Could Be Good for AI

The AI boom has been built on abundant capital and ever more computation but are we approaching an AI Winter? A financial correction could make efficiency matter again.

There is a lot of talk about an AI bubble at the moment, mostly centred on whether the hundreds of billions being spent on AI infrastructure will ever generate an adequate return. I think there is a more interesting question: what happens to AI if money and compute stop being abundant?

So far, much of the extraordinary progress in AI has come from scale. Bigger models, more data, more GPUs and more memory have produced better results across a surprisingly wide range of problems. If spending more on computation produces an answer that is sufficiently better to justify the cost, there is nothing wrong with that. But it does mean that we have spent the past few years developing AI during a period when there has been relatively little incentive to economise.

The amounts now being committed are extraordinary. The International Energy Agency reports that capital expenditure by five large technology companies exceeded $400 billion in 2025 and estimates a further 75% increase in 2026. That isn’t all AI expenditure, but AI is a major reason for the increase. The IEA also expects worldwide data-centre electricity consumption to rise from 485 TWh in 2025 to around 950 TWh in 2030. Those figures cover all data centres, not just AI.

None of this tells us that AI is a bubble. What it does tell us is that the financial hurdle is getting higher. The more capital, silicon, memory and electricity required to provide AI services, the more economic value those services eventually have to produce.

That is why I am less interested in whether we have another AI winter than in what one would do to the technology. Previous AI winters followed disappointment with what the technology could actually achieve. This time we already know that AI is useful. ChatGPT isn’t suddenly going to forget how to write, coding assistants won’t stop being useful and machine vision won’t disappear from factories because investors decide they have overpaid for AI infrastructure.

A future AI winter could therefore be primarily financial. There is plenty of precedent for a useful and rapidly growing technology producing terrible returns for some of the people who financed it.

Telecoms provides a particularly brutal example. In one Federal Reserve study, capital expenditure by publicly traded telecom service companies rose from $47 billion in 1995 to $121 billion in 2000. By 2002 it was back to $49 billion. Internet use hadn’t collapsed. The fibre being put into the ground wasn’t useless. Too much capacity had simply been financed too quickly and at prices that didn’t survive contact with the market. Much of that infrastructure subsequently became extremely useful to businesses that had not paid to build it.

AI isn’t telecoms. A GPU that becomes obsolete after a few years is a very different asset from fibre sitting in a trench for decades. But the basic lesson holds: being right about the growth of a technology doesn’t necessarily mean that you will make money financing the infrastructure behind it.

There is also a useful side effect when money becomes harder to find. People start paying attention to efficiency.

What is the cheapest way of knowing it?

I’ve been around technology long enough to see the same thing happen several times. Memory used to be precious, so programmers worried about memory. Storage was expensive, so we worried about storage. Bandwidth was limited, so systems were designed around bandwidth. As each became cheaper, it made less sense to spend expensive engineering time conserving it.

That is generally a sign of progress. There is little virtue in saving a resource that costs almost nothing. But software also has a tendency to expand into whatever resources are available, and AI has taken that logic further than most previous technologies. If processing another million data points gives you a better answer at an acceptable cost, processing them is perfectly rational.

Things get more interesting when the next increment of computation is expensive enough to notice.

Take a robot picking up a wine glass. A single 4K image at 3,840 × 2,160 contains 8,294,400 pixels. At 30 frames per second that is nearly 249 million pixel observations every second. In raw 24-bit RGB it amounts to roughly 746 MB/s, or just under 6 Gbit/s.

Real robotic vision systems obviously don’t have to send an uncompressed 4K stream to a giant model somewhere. Images can be compressed and cropped, frame rates reduced and processing performed locally. But the calculation illustrates something that I think gets overlooked. The amount of information available to the machine can be enormous compared with the amount it actually needs.

If the robot is already holding the glass, the important question may simply be whether it has started to slip.

It is possible to answer that with vision. With good cameras, good models and sufficient processing it may be possible to answer it very well. But that isn’t how we solve the same problem ourselves. I don’t watch my fingertips to work out whether a glass is slipping. I feel the movement where my fingers meet the glass.

This is one reason robot touch interests me. Not because touch replaces vision, but because the two are providing different information. Vision is very good at telling a robot where an object is and how to approach it. Once contact has been made, some of the information the robot needs exists at the contact itself. Measuring it there may be easier than reconstructing it from images.

The same argument applies elsewhere. A force sensor measures force rather than requiring software to infer it. Event cameras report changes rather than repeatedly recording a complete conventional frame. Dedicated electronics can extract a particular signal before a large stream of raw data has to be moved and processed.

There is no general rule that direct sensing is cheaper. Sensors cost money, need integration and calibration and can fail. A camera is extraordinarily useful because the same piece of hardware can answer many different questions. In some applications it will be cheaper to infer another variable from a camera that is already there than to add another sensor.

Which brings me back to the question I think matters: what is the cheapest way of knowing it?

If a hypothetical £10 sensor removes £100 of computation, bandwidth and hardware while giving an equally reliable answer, that is worth doing. If £100 of sensors saves 2p of computation, it isn’t. The interesting comparison is not between sensing and AI, or between dedicated hardware and general-purpose computation. It is between the total cost of different ways of obtaining the information needed to do the job.

Doing more with less

We already make these choices inside computers. General-purpose processors are useful because the same hardware can do an enormous range of things, but flexibility has a cost. If the same operation is going to be performed billions of times, designing hardware specifically for it can make sense. Modern AI itself depends on specialised accelerators for precisely this reason.

As the cost of AI becomes more important, I would expect the same pressure to work through the rest of the system. A smaller model may be perfectly adequate for a particular task. Sparse architectures can avoid using parts of a model that aren’t needed for every query. Edge processing can keep data close to where it is produced. Specialised processors can do repetitive work more efficiently. Better sensors may reduce the amount of inference needed in the first place.

The savings aren’t confined to electricity. Halving the computation required for a task might allow the same GPU to serve more requests, avoid another accelerator purchase, reduce memory traffic or release cooling capacity. At sufficient scale it may allow another workload to fit within an existing electrical connection.

This is where efficiency starts to look less like an incremental operating saving and more like infrastructure. If the thing preventing a data centre from doing more useful work is a shortage of accelerator capacity, memory bandwidth, cooling or power, reducing demand for that particular resource can have much more value than the direct cost saving would suggest.

There are limits to this. Saving processor cycles doesn’t recover the money already spent on a building, and physical infrastructure still has to be designed around resilience and peak loads. What matters is whether an efficiency improvement releases something that is actually constrained.

We will probably respond to cheaper AI by using more AI. That has happened throughout computing. If inference becomes ten times cheaper, developers will find applications that couldn’t previously justify the cost, users will make more requests and models will be asked to do more. Total electricity consumption can continue rising while the energy required for an individual task falls.

That doesn’t make efficiency pointless. The useful measure is not simply how much electricity the industry consumes, but how much useful work it gets from the capital, computation and energy available to it.

What comes after the boom

Technology booms are very good at getting things built. They are less reliable at establishing what those things are actually worth.

When capital is plentiful and demand appears limitless, speed quite reasonably takes priority. Get the model working, secure the GPUs, build the capacity and find the customers. There will be time to optimise later.

If investors become less willing to finance growth without convincing evidence of returns, later will have arrived. Unit economics, utilisation and hardware life will matter more. So will the amount of computation required to produce something a customer is prepared to pay for.

That could be uncomfortable for parts of the AI industry without being bad for AI itself. We may see more specialised processors, smaller models that are exceptionally good at particular jobs, more processing at the edge and more thought given to whether all the data being collected actually needs to be collected and processed.

For robotics, I think this will put more emphasis on sensing. A machine doesn’t always need a richer representation of the world. Sometimes it needs a particular piece of information, and the cheapest way to get it may be to measure it directly.

Large models aren’t going away, and there will be plenty of applications where spending more on computation produces results valuable enough to justify the cost. There will also be applications where it doesn’t. As that distinction becomes financially important, engineers will look further down the stack for savings: smaller models, specialised processors, local processing and, where it makes sense, better measurement.

This is why an AI winter could turn out to be useful. The past few years have shown what can be achieved when enormous amounts of capital and computation are available. A period in which both are harder to come by would put more emphasis on what engineers have always done when resources become expensive: finding ways of doing the same job with less.

In robotics in particular, that means starting with what the machine actually needs to know, rather than how much data we are capable of giving it. The useful question is then a fairly simple one: what is the cheapest reliable way of finding out?

References and calculation basis