Machine-learning-based artificial intelligence is now available to everyone, and it became an especially hot topic with the spread of ChatGPT. We have, however, been using AI-based solutions for some time. Banks, and the Estonian Tax and Customs Board, use AI systems that analyse transactions and behaviour patterns to detect fraud and suspicious activity. Many companies use chatbots to automate customer service, and machine-learning methods can be blamed for the selection of advertising shown to users online. Amid all these and dozens of other applications, very little attention has been paid to how an AI is actually put to work.
It all starts with technology
Broadly speaking, applying AI requires two components: computing power and data, both in large quantities. Over the decades of AI development it has been the lack of computing power that has caused standstills. Those pauses in development even have their own term, AI winter, and they have sometimes lasted years. Every standstill has been followed by progress, though, and it is fairly safe to say that no technology-driven "winter" is in sight right now.
A shortage of data has also played its part in AI winters. Training AI algorithms takes a large amount of high-quality data, and that resource has grown abundantly in recent times. Today, information from social media, Internet of Things devices and public digital databases can be used to train AI. So-called synthetic data is also created to train machines, and people are hired to give the AI feedback on its decisions against set parameters.
Where do the large AIs live?
Large AIs are often housed in data centers because they demand a great deal of electricity, a controlled environment and high security. The requirements and expectations placed on the infrastructure naturally grow over time. The US technology giant Meta, for example, has announced that it intends to double the number of its data centers by 2028. The company already owns 21 "data campuses" around the world, representing a total of 16 billion dollars of investment and more than 3.7 million square metres of data center floor space. Because the success of AI applications depends on data-processing capability, this is the only possible way forward.
Into each next-generation data center Meta adds more computing power, using a new chip designed in-house that, by Meta's own account, is twice as efficient as the GPU (graphics processing unit) used in most AI infrastructure.
What does AI demand of data centers?
"The more complex and demanding the IT equipment, the higher the requirements on the environment it runs in. An average office building cannot provide the power and cooling needed to run a large-scale AI. That is why, in some cases, entirely separate data centers are built for them," says Kert Evert, the originator of the idea for the Baltics' largest data center. According to him, there are companies developing and applying AI in their building too. One good and internationally successful example is R8 Technologies, which grew out of a garage, so to speak, and now operates in 15 countries across Europe, using AI to control the indoor climate and energy use of large, technically complex buildings. You can read more about their success story in another blog post.
"Given that we have enough electrical capacity for an average Estonian town, we can host even more power-hungry AIs," Evert confirms.
The most common hardware used for AI is the graphics processors, or GPUs, mentioned earlier in the Meta example, because they can perform many calculations in parallel. In an AI context GPUs therefore often do their job several hundred times faster than classic processors (CPUs). Highly efficient (and expensive) special versions of these processors have of course been developed by now.
The same applies to storage devices capable of handling both structured and unstructured data. Connectivity is, naturally, also critical in this equation: the links between the processing units working in parallel have to carry enormous volumes.
The rule here is that the more specialised and the newer the device, the more it lightens your wallet. Taking processors alone, at the beginning of this year 2,048 Nvidia A100 GPUs were used to train Meta's LLaMA AI. At the time of writing, one such GPU costs around 10,000 euros, so the processors alone exceed the 20 million mark. "In that context it is fair to say that, alongside reliability, physical security is undoubtedly important for training and running AI. Both the equipment itself and the intellectual property on it have to be protected," Evert adds.
In summary, the AIs reshaping our future world rest largely on two components: computing power and data. The first requires specific hardware and data centers to house it. While the world's data keeps growing on its own, data centers and the technology running in them do not appear by themselves, and rapid development awaits us here in the near future.