Marc Andreessen is, besides being the founder of Netscape during the internet's adolescence, a successful venture capital investor and, not least, known for his 2011 essay ’Why Software is Eating the World’ . We have indeed seen fantastic development in many software companies since then. The 2010s were unequivocally the decade of software, and we continue to see very good opportunities for these business models to grow structurally and scalably with good profitability.

However, during the 2020s, software seems to be sharing the spotlight with hardware companies. Both the pandemic and the war in Ukraine served as reminders that hard things are needed. Not just vaccines and tanks, but robust supply chains and energy systems. Dictatorship-free electricity production, if nothing else. Software tends to drive prices downward (deflation), as it solves an existing problem at a lower cost. Conversely, (excessive) demand for hardware often drives inflation. We can all see the shift from the 2010s' deflation and zero interest rates to the 2020s' inflation, which at least partly appears to be becoming structural. This is a strong indication that hardware, in a broader sense, has made a comeback.

Artificial intelligence, currently the best example of software that impacts (or "eats") the world, also requires massive investments in hardware. AI is born, trained, and thrives in data centers, solid buildings filled with tangible things. More precisely, AI exists on hardware in the form of processors and semiconductors, just like all other software.

The semiconductor sector was for a long time a difficult sector to invest in – enormous investment budgets, oversupply, and cutthroat competition, constant price pressure fundamentally driven by Moore's Law, and a strongly cyclical business. Companies have been forced to handle volume and price fluctuations that are accentuated through the entire chain from factory to a final product in a consumer's hand. Often, decisions have been complicated by poor visibility – meaning it's difficult to see how a change in volume or price affects the chain further down and a relatively long time after the decision.

Let's address these problems one by one. The industry has tackled high investment requirements by becoming fabless, that is, sticking to designing semiconductors and letting TSMC and other so-called foundries bear the investment costs. Thanks to this focus, these companies have been able to afford to stay at the forefront of technology development, which has gradually reinforced the tendency and created deep moats over time. On the back of this outsourcing, there arose a need for tools and design software to create tomorrow's semiconductors. Here, Cadence and Synopsys (both holdings in TIN World Tech) have taken the lead and grown strongly, both organically and through acquisitions. Synopsys recently made a bid for another simulation company in the fund: Ansys. The semiconductor industry itself has also consolidated through acquisitions and mergers, which in the case of Xilinx (acquired in 2020 by AMD) has contributed to the returns in TIN World Tech. This has greatly contributed to more reasonable competition and a better balance between supply and demand.

Sand is the New Oil

Moore's Law — the expectation that computer performance can double approximately every two years — has been one of the cornerstones of technical development over the past 50 – 60 years. The sharply decreasing cost of computing power ("cost of compute") has served humanity well, but has been challenging for the semiconductor industry to balance. Price erosion has been consistently high and gross margins difficult to defend over time. New and more demanding tasks are really the only countervailing force. A company like Nvidia has successfully navigated from one specialized niche to another over the decades, where the increase in demand for computing power has more than offset Moore's Law. From gaming computers (3D increases calculations cubically, while Moore's Law is quadratic) through data centers and crypto to artificial intelligence.

Data, which is what all AI requires to be trained and become efficient, has been called the New Oil. According to Sam Altman of OpenAI, it is rather computing power ("compute") that is today's most valuable commodity. If you will: sand (silicon) is the new oil. The demands placed on training and inference — all the questions thrown at the model — are increasing far faster than Moore's Law, which currently gives Nvidia strong pricing power. Those who can help meet the need for compute will be able to generate value for a long time to come. Nvidia (and its investors) can expect future challengers, as the incentives are too strong for this not to happen.

The sensitivity to economic cycles and the cyclical element remain and can probably never completely disappear. Again, industry consolidation has somewhat alleviated the situation, as has the shift to contract manufacturing. Right now, we see few signs of economic movements for semiconductor companies with operations related to AI. Herein lies the challenge for investors: in certain segments, there is currently no cycle to be seen. At all major shifts, people in general, and investors in particular, tend to overestimate the process in the short term, while underestimating the power of change in the long term.

We strive to balance risks against opportunities by trying to find ultimate winners, especially related to AI. Preferably by investing in picks and shovels rather than gold diggers. Nvidia remains a medium-sized position in the fund. During March, TIN World Tech bought shares in ASML, Applied Materials, and TSMC. TSMC is the contract manufacturer that stands head and shoulders above almost all others. For 30 years, they have built a position where they can stay at the forefront of production technology and bear investment costs of over 100 billion for each new factory. In the metaphor where cloud companies like Amazon, Microsoft, Google, and Oracle are railroad owners, Nvidia is best at assembling sleepers, rails, and catenaries into a finished system. TSMC is here the steel company that delivers the very latest and best rails.

Applied Materials and ASML are subcontractors to TSMC and many others that manufacture semiconductors, each of them leaders within their niches. Studying flagship customer TSMC and how they distribute their purchases, ASML and Applied Materials are the largest and take a considerable part of TSMC's investment budget. In the railroad metaphor, they thus become the mining companies that supply the steel mills with coal and iron ore, respectively.

A Platform with a View

Will it be OpenAI (and thereby Microsoft) that becomes the most successful train operator, or one of the many other projects? How easy will it be to run a train over land that could be seen as mined with conflicts around copyright and intellectual property (IP)? Who will be the first provider of ethical AI, traceably free from infringement on someone else's property? And, can anyone cost-effectively operate a train line when each passenger litters the train carriage and increases the operating cost?

We will eagerly engage in intensive train spotting and seek answers to these questions. Thanks to the investments we have already made and the network we possess, we have a reasonable spot on the platform from which to observe. In the meantime, we content ourselves with investing upstream, among the companies that form the technical foundation and enable the development of train lines. Hopefully, they will also give us passengers more efficient lives.

What do we mean by "littering the train carriage"? Back to computing power and the cost of compute. Each question or prompt consumes computational power and, therefore, money. According to Sam Altman, it can often take up to a thousand prompts before a professional user is satisfied with an AI-generated image. If users were charged the true cost, they would be discouraged, but at the same time, train operators cannot sustainably run operations like a bank in Öbberöd.

We all need to learn to become better at prompting, to use these tools in the best way possible to reap the benefits of increased productivity. Even though we collectively have the drive and the incentives, this may take too long to save a shaky train operator. Either the cost of computation must come down significantly, or perhaps someone can develop new tools for AI consumers. Why not an AI that learns to write prompts exactly as you want them? A bit like sending your avatar on the train instead of traveling yourself. As said, it will be exciting to see what we find on our platform with a view.