The year 2023 will go down in history as the first time the mass public came into contact with artificial intelligence (AI) through the launch of OpenAI’s app: ChatGPT. The company was the quickest to reach 100 million users in its user base, a significant milestone for any B2C (business-to-consumer) technology company, requiring only 2 months to achieve this feat. Instagram, for example, Meta’s popular social network, took 30 months, i.e. 15 times longer to achieve the same milestone.
The reason behind why so many people were attracted to it was an AI model called the Large Language Model (LLM). Its use has unlocked the potentially huge productivity gains that can be achieved by using AI-based technologies.
But what is an LLM?
A Large Language Model is an advanced type of artificial intelligence that can understand a question and generate answers in a sophisticated way. They are trained with large amounts of text to learn language patterns and structures, allowing them to perform tasks such as answering questions, writing texts and even carrying out translations. These models have the ability to understand complex contexts and produce contextualized responses, simulating in a certain way a conversation with a real person.
To give you an idea, the paragraph you have just read was generated 100% by ChatGPT 3.5 (the model is now in its 4th generation) using as a parameter: “Explain to me in a simple paragraph what a Large Language Model is”.
The evolution of ChatGPT began by predicting what the next word in a sentence would be. As more training was carried out, based on data feeds, the model became more sophisticated. For example, it managed to outperform 90% of humans on the SAT (an American university entrance exam) [1]. This was a truly remarkable achievement for a computer program. As a result, it quickly began to be used by its user base for everyday queries (e.g. putting together a travel itinerary, making a recipe with items from the fridge) and making simple productivity gains at work (e.g. summarizing or translating a large block of text).
It is important here to point out that the AI model we are referring to and which is the focus of all the market euphoria is a generative model which can be both image and text-based, such as DALL-E and ChatGPT, both of which belong to OpenAI.
Despite OpenAI’s admirable achievement, what is the real use of the LLM in the corporate world and how important is it?
To give a good practical example, consider GitHub, a Microsoft company for code developers, which launched GitHub Copilot in 2022. Through an LLM, GitHub Copilot helps programmers to write code more productively at a price of just $10/month. This feature is responsible for US$100 million in revenues [2], which still represents only 1% of the user base and 10% of the total revenue [3]. This demonstrates the significant growth potential as technology adoption continues to expand. It is also worth watching Adobe’s Firefly video [4] which illustrates the impact of artificial intelligence on the creative world (in this case the app is applied specifically to the world of design).
Now that we have briefly discussed the promising signs of artificial intelligence in generating value through increased productivity, it's time to address the extensive value chain that involves i) software/applications, ii) infrastructure, and iii) semiconductors. It's worth noting that companies can be present in more than one point of the chain, as is the case with Amazon, Google, and Microsoft, which are leaders in cloud computing with a combined market share exceeding 60% [5].
Perhaps the best way to discuss AI is from the end user’s point of view. A person may have demands for various software such as Adobe, Apple, Google, Instagram, Microsoft Excel, Oracle, Salesforce, SAP, Tesla and Uber. An AI model can be built within these companies (e.g. Google Bard) or they can choose to rent an external model, such as that provided by OpenAI. In this option, it is also possible to train the model with internal data, so that it addresses the company’s needs in a tailor-made way, while also offsetting the risks of compliance and data sharing by keeping the model within the company’s security system and maintaining the information private.
Listed companies in the AI chain
In software, we can identify which listed companies in the sector have the advantage of being first movers for two main reasons: i) speed and ease of distribution and ii) access to volumes of data. In the first case, the best illustration is Microsoft which is the absolute leader in productivity applications (e.g. Excel, Teams, Outlook). Its alliance with OpenAI was important in consolidating its position on the issue. The move was strategic from a defensive point of view (incremental innovation is essential to create barriers to new entrants) and from the point of view of increasing revenues (the client has to pay for an upgrade to use it), while OpenAI gains speed in distributing its model. Databases, on the other hand, represent both an opportunity and a threat since they are increasingly interchangeable or even public, giving companies the chance to take on the incumbents with more efficient AI tools. The first challenge is already underway, with Microsoft’s Bing using ChatGPT to make Internet searches more assertive and in doing so challenge Google search.
Moving along the chain, we arrive at infrastructure and the story of the emergence of AWS (Amazon Web Services). This was the first cloud computing company whose business model was based on renting servers to others at a cost as competitive as owning a server. While the origin of cloud computing dates back to a common outsourcing process, it has developed and become increasingly more specialized. Furthermore, in the AI era, financial sums involved have become superlative.
A high-tech server, like the one used by OpenAI to train the model, can cost US$4 billion [6], an amount that will be paid out by Amazon, Google and Microsoft, among other cloud computing companies. These will, in turn, rent the processing capacity to various companies. In OpenAI’s case, it will rent these servers for both training and inferences (each ChatGPT query represents an inference). Although this variable cost is low, around US$0.02 [7] per query, it becomes representative in scale. In one month, OpenAI could have spent US$40 million on inferences. This is clearly an advantage and a market that is still in its infancy for these three Big Techs.
One more step along the chain takes us to semiconductors which are the main content of these high-tech servers. Their complex ecosystem can be considered a separate value chain. The chart below shows how some of the main companies are organized.
Nvidia is perhaps the most talked about company at the moment as it was the first semiconductor company to exceed a market value of US$ 1 trillion [8]. The reasons why it has achieved such a feat may be both its absolute leadership with over 90% market share [9] in supercomputer chips and the fact that these chips account for over 70% of the cost of a server [10].
The chips that Nvidia produces are called GPU (Graphic Processing Unit). The reason why this chip has managed to outperform the CPU (Central Processing Unit) in training LLMs is its ability to perform simultaneous calculations, known as parallelism. In this process, parts of the model are trained at the same time to speed up the calculation process and thereby reduce the time needed to train the entire model, improving computational efficiency and the final cost.
Despite its current hegemonic position, Nvidia is already facing challengers on the market. Some are older, like AMD, and others are relatively new, like Amazon and Google. It is no coincidence that the latter are using greater verticalization in an attempt to gain more differentiation and profitability in cloud computing.
Another feature of the sector is the outsourcing of chip manufacturing. Nvidia only designs the chips, but the final manufacturing is the hands of other companies such as TSMC (Taiwan Semiconductor Manufacturing Company). TSMC is the leader in vertical manufacturing with more than 50% market share [11] – another natural beneficiary of the AI movement.
Chip manufacturing is highly capital-intensive and extremely complex. An advanced chip factory requires around US$10 billion just to be built. The complexity, in turn, can be illustrated by the involvement of the fourth state of the matter, plasma, in the manufacture of semiconductors. The result is the meticulousness required to achieve high production yields and specialized machines that cost tens or hundreds of millions of dollars, as is the case with Applied Materials, ASML, KLA, Lam Research or Tokyo Electron. Given the complexity of the topic, we won't delve too deeply here, but I believe it's clear that there is a significant barrier to entry for new players in this segment of the chain, highlighting the importance that current players hold in it.
The table below presents an interesting overview of the universe of public companies linked to AI through operating profit (EBIT - Earnings Before Interest and Taxes) in which the 50 largest reached US$ 688 billion in 2021.
Private companies in the AI chain
That said, we cannot fail to mention the impact of AI on the private market since most of the innovations are native to Venture Capital market deals.
OpenAI itself is still a private company that raised $10 billion in January of this year in a round led by Microsoft for its expansion into the LLM universe, as mentioned.
On the Enterprise Software side, as we are talking a lot about efficiency and productivity gains, Databricks is one of the leaders in the sector. It has raised $3.5 billion over its 10 years of existence with well-known Venture Capital managers such as Andreessen Horowitz, Coatue and NEA. Databricks provides a unified cloud-based platform for processing large blocks of data, with the aim of making data processing and analysis simpler and more efficient. This company’s IPO which is expected to take place soon (probably in 2024) will be an important milestone for Venture Capital managers and the AI segment.
As for the semiconductor segment, which as we said is an essential “raw material” for processing AI models, we should mention ARM, whose IPO was eagerly awaited by the technology market this year. The IPO price in September valued the company at USD 54 billion. The company operates in the same value chain as Nvdia and was invested in by Softbank, another major player in the Venture Growth market.
Although AI investing is booming today, the Venture Capital market has been allocating capital to the segment for many years. In 2018 (long before the launch of ChatGPT 4.0) more than USD 50 billion was allocated to companies whose business models are directly linked to AI. The figure in 2022 was USD 73 billion and by the middle of this year we are at USD 40 billion invested in almost 3,000 deals.