Historical returns are no guarantee of future returns. Fund units and other financial instruments can rise and fall in value, and it is not sure that you will get back invested capital. The fund’s value can vary significantly due to its composition and the management methods used by the fund management company. Fact sheets, sustainability-related disclosures, and prospectuses are available on seb.se/fonder.
More information (seb.se, in Swedish)
What has surprised you the most during the year?
What surprised us the most was the large scale on which AI infrastructure was deployed. Many companies have chosen to expand capacity as much as they can. Through their investments, cloud companies Amazon, Google, and Microsoft have already become a “one-stop shop” for AI.
AI infrastructure is very similar to data centres and is sometimes referred to as data centres on steroids. This may explain why the large cloud giants, already leaders in data centres, are also leading the development of AI. However, the difference is often that when building a data centre for AI, the most advanced technology is required, and frequently, it is needed on a larger scale. Another significant difference is that data centres have previously often been built on CPUs (Central Processing Units). In contrast, in AI data centres, GPUs (Graphics Processing Units) perform calculations in parallel and thus go much faster. An AI data centre often requires 5-10 times more energy than a regular data centre.
What does the deployment of AI infrastructure look like?
So far, it looks like the expansion of AI infrastructure is continuing. Although the companies do not report exact numbers about what their data centres look like, it appears that the first model of ChatGPT used about 10,000 GPUs. By the end of this year, Meta is expected to have a model that uses 350,000 GPUs. Microsoft is also rumoured to be planning a data centre with more than 1 million GPUs by 2028. Then, it should also be said that the power you get out of a GPU roughly doubles every two years due to improvements in the process of semiconductors. We will have to see if these figures become reality, but we can be sure that a significant expansion will occur.
What do you think will be the next AI trend?
With the enormous expansion in infrastructure, we see that the next big step in AI is using models or so-called “Deployment.” Although several AI models exist today, their use is relatively low compared to the vast investments made in infrastructure. For example, the American venture capital company Sequoia has estimated that in 2023, more than 50 times more was spent on AI infrastructure than on AI software.
Perhaps the most well-known model many people use is ChatGPT, a “general purpose model”. It is a model trained on a massive amount of data (basically the entire internet) and should be good at most things. What we believe will drive the use of AI in the coming years are “Application Specific” models. These models are trained on a smaller amount of data specific to a particular area or task. This leads to models that are very good at a specific task but are also much cheaper to train and use.
If we are to mention another trend, it is “Edge AI”. This means that you move the process to the device you are using rather than the process taking place in the cloud. This is something that Apple has been working on to move the AI process to your smartphone. We also see that this will be an essential part of the automation of production or self-driving cars.
What type of companies are in the fund?
Some of the companies we have in the fund have specific models. Adobe has a model called Firefly, which can be described as Photoshop with AI tools. This tool can reduce the time it takes to edit images by several hours. Another company with a specific model is RELX, which has the Lexis+ AI model that helps lawyers in their work. Their evaluations have shown that a lawyer can save 6-7 hours per week with their tools, and it is also about 30 per cent more accurate than a general model because it has only trained on data in its field. Another vital area is coding, where Salesforce has reported that they save about 20,000 hours a month by using AI tools when coding.
How did you perform in the first year?
We had an excellent first year and are up 45.2 per cent, significantly better than the fund’s benchmark index, which rose 35 per cent*.
What have been your best and worst investments?
Our best investment is Nvidia, which has increased around 200 per cent since we started the fund. The first step to beginning to use AI is to build the infrastructure required to train and use models, and that’s where Nvidia is crucial with its GPUs. Nvidia has almost a monopoly in this market, and all models rolled out today have been trained with Nvidia’s chips. Other investments that have been good for the fund are the companies with cloud services and their LLMs (Large Language Models), such as Meta, Alphabet (Google) and Microsoft.
One company that has not contributed as positively to the fund is Oracle, which has been one step behind the cloud companies in its database services but has started to catch up recently. Another company is AbCellera, which uses AI to find new drugs and speed up the process of finding new drugs. The company is still very early in its development, which explains the weaker performance.
* Refers to SEB Artificial Intelligence Fund C (SEK) for 30 June 2023 - 28 June 2024.