Neurophos and the promise of optical compute in AI.

Data plus context creates information, which allows us to build knowledge and ultimately yields intelligence. By continually adding more context, we have developed powerful AI. Let me explain.

In the breakthrough AlexNet deep-neural network from 2012, this image recognition model was trained by feeding it a large number of pictures, each with a human-generated label that, for example, said “this is a cat.” The label provides the essential context that turns the pixel data into information. By then finding related features—or parameters—across this large amount of information, we can build an AI knowledge model that is able to recognize cats. But this “supervised” learning approach, with its human labels, was holding AI back.

Then, in 2017, the famous academic paper “Attention Is All You Need” was published, which showed how context actually forms part of language. This allowed “self-supervised” training, where the model was able to learn directly from the contextual information captured in language. It became possible to feed in vast amounts of language and build bigger models. Then, with more compute applied to the problem, we were able to create powerful artificial intelligence—like ChatGPT and Claude.

Recently, this scaling effect has started to reach certain limits, and the solution was to add a reasoning process that could go off and find more context associated with the specific task the AI was being asked to solve—creating agentic AI. Adding more context has now become the key factor in making AI both more powerful and more accurate.

Processing input context is where the major portion of compute and energy is now consumed in AI inference. These “input tokens” need to be mapped to an internal “value” that the AI model can understand. To do this, each input token is encoded with a Key (K) and a Value (V) vector—in what’s called the Key-Value cache, or KV-cache. For example, the word “swordfish” will be broken into two tokens, “sword” and “fish,” with each assigned its own KV values. Then, when we input the word “cuttlefish,” we can reuse the encoding we have already calculated for “fish” and compute a new KV value for “cuttle,” adding this into our cache. As the number of unique input tokens grows—i.e., as the context window gets bigger—the KV-cache grows, and the amount of compute required increases dramatically. Today, this means using more GPUs, with huge amounts of energy needed.

With AI inference, demand is currently growing by around 7x each year, and context window sizes are growing by 30x per year. Providing the energy and compute demanded by this input token processing is becoming one of the biggest challenges to solve in AI data centres.

But one potential solution does exist…

Instead of doing this KV-cache encoding in silicon on GPUs, we could instead use optical computers. Using light to do complex AI multiplications is technically possible and the incredible “theoretical” benefit is that optical processing consumes zero energy. However, there are two very big technical challenges that we must overcome:

First, we need to convert the tokens, which are stored as digital information, into light, which means converting from digital to analog and then generating light sources that match these analog signals. Then, after performing the “zero-energy optical compute,” we will need to capture the resulting answers and convert them back from analog light to digital data. This conversion process can consume significant amounts of energy, potentially offsetting the compute energy savings.

Second, we need to ensure that the maths precision through this optical compute path matches the precision of the maths that would otherwise be done on a silicon GPU; otherwise, the system might end up giving us the wrong answer.

To solve the first problem, we need to be able to do a substantial amount of compute in the analog domain to overcome the energy cost of converting from digital to analog light and then back again. This is where other companies, that have tried to solve this optical compute problem, have failed.

Neurophos has developed a new metamaterial-based optical modulator that is over 10,000 times smaller than any other optical modulator, or optical “transistor,” previously built. This allows Neurophos to build an optical AI processor that is over 100 times more energy efficient than a traditional silicon GPU will ever achieve.

Neurophos has also solved the second problem too. Its optical compute path can match the Normalized Mean Square Error (NMSE) in this complex AI mathematical processing of the leading GPUs—specifically for this energy-demanding pre-fill inference task.

The incredible breakthroughs that Neurophos has achieved are why I invested in the company, alongside the likes of the Gates Foundation. They are why I am delighted to announce that I have joined the company’s Board of Directors. And they are why the company is getting so much attention from the major AI labs and hyperscalers.

Neurophos is now delivering the potential for much more AI compute with far lower energy use. Its technology will allow AI to become even more intelligent. But perhaps more importantly, with more context available, we can build AI at lower energy that is also more accurate.