Could AI break the internet? Some reasons to be cheerful .

The massive increase in the size and sophistication of large language models such as ChatGPT have occurred alongside an exponential growth in computing power, energy consumption and raw data. With existing technology reaching its limit, how might the internet find ways to accommodate cutting-edge AI in the future? We examine the challenges and consider some potential solutions. 

The computing universe is expanding and accelerating at breathtaking pace. Currently, the world has over 1.1bn gigaFLOPS of computing power (a gigaFLOP is one thousand million calculations per second)– up 6,150% in a decade. Yet demand for more power continues unabated and, with 4nm chips now available, our ability to shrink technology further is fast diminishing. How much longer can Moore’s Law last? 

In addition, new computing methods are arriving all the time, from quantum to new CPU architectures. The game-changer is AI. Large language models (LLMs) such as ChatGPT offer powerful and versatile human-like intelligence that have revolutionised computing. But these models are hungry: ChatGPT-3 required 175bn parameters; three years later, GPT-4 devoured 1trn parameters, a six-fold increase. Now the compute required to train a cutting-edge AI model is estimated to double every five months. 

Power needs power

Generating such computing power takes massive amounts of energy. Training ChatGPT-3 consumed 1,287 MWh of electricity and generated 552 tons of CO2e, equal to driving 123 petrol-powered cars for a year. And that’s before anybody used it; data from Google and Meta suggest that training large AI models accounts for only 20-40% of their total energy consumption. Furthermore, datacentres running AI compute already use 1.5-2% of the world’s energy, roughly equivalent to that used by the UK. This is expected to double by 2030. 

The computing revolution, therefore, poses big questions. Some are clear: how can compute continue to meet the exponential growth in demand? And how can this happen without draining the world’s energy supply? 

Supply lags demand 

Others are less obvious. For example, the amount of data created, captured, copied and consumed globally is growing exponentially, from 64.2ZB in 2020 to more than 180ZB in 2025. How can we continue to move so much data quickly and efficiently to the end user? And with ever more computations happening ever faster in ever smaller spaces, how can we avoid bottlenecks into and out of processors?

More fundamentally, network bandwidth is lagging this increasing demand. This imbalance will only increase – wireless capacity is fixed by nature, while physical infrastructure is expensive and slow to roll out. So how can the internet expand quickly enough to handle the accelerating volume of data caused by cutting-edge LLMs? How can AI and the internet grow in synergy?

Two useful developments

Some developments are helping. Most modern computing, including some smaller AI models, does not require datacentres. By processing data locally, edge computing uses less bandwidth and reduces latency – essential for time-sensitive tasks such as autonomous driving. And by operating independently of the cloud, it offers users greater reliability and privacy and, often, lower costs.

Equally, the rapid growth of Low-Earth Orbit (LEO) satellites, such as Starlink, increase network capacity by providing fresh bandwidth. Their rapid expansion has seen costs plummet, possibly as low as $1m per unit. With direct-to-cellular services expected soon, enabling smartphones to connect with LEO satellites as well as base stations, capacity constraints will ease a little.

Nevertheless, LEO satellites cannot alone overcome the mismatch between network demand and supply. Meanwhile, edge compute cannot replace datacentres’ role in distributing computing power. And neither tackle the central challenge of making future compute faster and more sustainable.

Light might work

One of IP Group’s portfolio companies does. By using optics instead of electronics, Lumai can perform AI calculations far more efficiently than traditional electronics. Unlike companies trialling silicon photonics, Lumai uses 3D optics – ideal to reach the scales needed for AI. “The operations underlying AI are just simple arithmetic – basically multiplication and addition, but with billions or trillions of numbers per second,” says Lumai’s co-founder, James Spall. “Using optics instead of copper cables for communication gives you much higher throughput and much lower latency, with far less energy consumption – the same is true for performing arithmetic.”

Having developed the world’s first optically-trained neural network, Lumai is now focusing on inference – “using the model to do something useful once it’s trained, which our processor does very well,” says James. Better processing is only part of the solution, however. “There’s also the challenge of memory and interconnection,” he explains, “and as we move more towards cloud compute, moving data around becomes increasingly important.” 

One answer: AI? 

Deep Render, another portfolio company, addresses this issue very effectively by using AI to compress video data, which accounts for 70% of all internet traffic. “It’s choking the internet’s capacity but traditional compression is hitting its limits,” says Sebastjan Cizel, Machine Learning Lead at Deep Render.

Enter AI-based compression. “The fact that we’re creating more and more video data lets us train better compression models,” Sebastjan explains: “In just five years we produced a compression algorithm at least five times better than the current standard, effectively increasing global bandwidth five-fold.” And it’s incredibly cost-effective: “We’re able to provide the equivalent of trillions of dollars of bandwidth investment just with a compression algorithm.”

Deep Render was the first company to build an AI compression-backed video chat app that can encode and decode simultaneously in real time at 30fps. And it recently won the Intel Ignite Innovation award: “Millions more people now know what we’ve achieved so it gives us a lot of exposure to potential customers,” says Sebastjan: “That’s very valuable.”

Many future solutions

We believe that Deep Render and Lumai offer viable routes ahead for 21st century computing. That’s why we’re backing them. However, the challenges are complex and there will be no single winner. Other solutions may improve spectral efficiency or replace flash memory. Meanwhile, future innovations will surely emerge around processing, data handling, transmission and storage. We will continue working with entrepreneurs, accelerators and like-minded investors to identify and support viable solutions to come; solutions that will enable future compute to become more sustainable and for AI and the internet to grow in synergy.