Perspectives on artificial intelligence

We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.

– Amara’s Law, Roy Charles Amara

Artificial Intelligence (AI) is a term used rather loosely, but at its essence, it can be used simply as an umbrella term for strategies and techniques you can use to make machines more human-like. The recent emergence of Generative AI tools using Large Language Models (LLMs) that can receive instructions and compose responses using natural language has catapulted this field into mainstream awareness and unlocked excitement around its future potential.

Overall, AI is a tool, as software has been a tool, and as a tool, it will be incorporated by businesses and consumers to solve problems and drive productivity. The productivity gains that AI will unlock are expected to provide the next wave of economic growth, and consumers, businesses, and countries that leverage this best will be positioned to pull ahead. There is much capital pouring into this space, many new businesses and business models emerging, and the disruption that AI may create could provide an opportunity for us as a Fund to find new champions. But it is also critical that our portfolio companies capitalise on this changing landscape to drive their businesses, and that we leverage it as an effective tool even within our own Fund operations.

The evolving AI industry

The large technology companies such as Meta, Google, Microsoft and Amazon that have strong balance sheets and cash generation from legacy businesses are now racing each other to execute massive capex drives to build out processing power for AI applications, with Nvidia’s GPU chip being the current beneficiary of this investment wave. In fact, at this pace, their business models are at risk of transitioning from a high ROI platform play with deep network effects to an infrastructure-based model as they search for ways to ensure they are not disintermediated by this next wave of disruptive innovation.

However, the application layers and the business models that will be structured around them are still a work in progress. Estimates are that last year, the industry spent around US$50bn in capex on chips but only generated around US$3bn in revenue. Regardless, capex investment is set to accelerate, with current quarterly forecasts showing a tripling of capex spend. Questions now are being fairly asked as to where the revenues will come from that will justify this enormous spending. The industry seems to be faithfully following the classic Gartner Hype Cycle, and we seem to be heading quickly from the Peak of Inflated Expectations into the Trough of Disillusionment.

Gartner hype cycle

Stepping back from the noise and viewing this sector in context in the long term appropriately frames it as a continuation of long-running technology development, points to where some of the direct and indirect impacts may be, the potential impact on our portfolio companies and the opportunities that may arise for disruption and investment. The longer-term implications (the Plateau of Productivity) will provide the most powerful opportunities.

Within this space, there is a complex value chain but simplistically it can be thought of as encompassing:

1. Supply chain to the chip industry (e.g. ASML, and the suppliers to ASML)
2. Core compute engine designers and manufacturers (various chip configurations such as Nvidia and other GPUs/TPUs/LPUs etc).
3. Cloud computing providers (e.g. Google, Meta, Amazon)
4. Training data (e.g. search data, social data, research)
5. Language models (e.g. GPT, Llama, Claude, Gemini)
6. Application layer (e.g. Microsoft’s Copilot, Salesforce’s Einstein, Perplexity for search)

The strategic dynamics within each are important to consider when evaluating knock-on effects and opportunities within and outside of AI, with different capital needs, expertise depth and network effects important for success in each. But at this stage, we expect that for venture and growth capital, the most interesting opportunities are likely to be found in areas 1 (in deep-tech businesses) and 6 (especially valuable niche-level applications), with some supporting opportunities elsewhere (e.g. cooling and energy management technology). For our portfolio companies, we expect that areas 4 (training data) and 6 (application layer) will be where they will need to capitalise.

The technology foundation of AI

Computing capability has been driven up exponentially over time. Moore’s law (the observation that the number of transistors in an integrated circuit doubles about every two years) is slowing as transistor size has bumped up against the limits of physics, but other innovations have allowed computational power to continue to rise on a power law basis (i.e. exponentially). Simultaneously, other critical drivers have also continued to change on a power law basis: storage cost declines, the volume of data stored, and data transmission speeds. This has had an inexorable positive impact on our capability to conduct complex calculations since the first electronic computers were developed in the 1940s, and consequently impacted a plethora of human endeavours. Generative AI is built on a foundation of sufficient stored data, highly sophisticated models, and computational hardware that now has the ability to process the data at sufficient volume and speed to be useful. Previous cost and capacity breakthroughs facilitated the rise of Netflix thanks to streaming video, and Instagram thanks to phone cameras and images. Working in the US on private equity projects in video compression, fibre networks and mobile networks in the early 2000s, I was in the fortunate position of being a direct witness to the powerful effects here through the dot-com boom, bust and re-emergence of new business models. As an example, the rapidly collapsing cost of data transmission and storage opened up a radically different video transmission experience that resulted in Netflix streaming and YouTube.

Overall, computational power, cost speed, and the application of complex models to a variety of problems have been core drivers to this phase of human technological progress for some time and will likely be for some time further. See below for a perspective on how these dramatic functional and economic exponential shifts have been playing out for some time (note that the y-axis is on a log scale, representing 10x changes in each step).

For now, based on history and current innovations, it can be comfortably extrapolated that computational power and speed, data storage, transmission speed, and volume are likely to continue increasing dramatically, opening up ever-increasing application possibilities and improving speeds and costs.

However, note that the observation behind Moore’s second law also applies: which is that the capital cost of a semiconductor fabrication plant also increases exponentially over time (essentially doubling every four years). Suppliers such as ASML, which makes highly complex photolithography machines, have been beneficiaries of this rising investment. For now, the sheer increasing scale of the industry has ameliorated this cost per unit, but at some point, the cost of investment may start to compress the economics.

Drivers of success in AI

AI can be thought of simplistically as a multivariate regression model which essentially takes existing data to calibrate a model which can predict an outcome based on a new scenario. The accuracy of the model is a factor of two things: the amount of data, and the number of variables (complexity) that can be handled in creating the model. The more data you have, the better the model. The more complexity you can handle in the calculation, the more variables can be added, and then the more accurate the model. These models have been rapidly rising in computational capability as the technical foundation has strengthened.