Andreessen Horowitz are a legendary Silicon Venture firm who are famous as early backers of success stories like Facebook. In 2011 their eponymous founder Marc Andreessen presciently noted that ‘software is eating the world’. Companies producing useful software would seize an increasing proportion of the economic pie from legacy industries. The best example of this is the evisceration of traditional advertising mediums like print media. Advertising is largely a digital phenomenon now because the way people searched for information changed. Simply put, the physical world surrendered to the digital.
Why is software development such a good business
Digital products like software make fantastic businesses. Yes, there is the cost of creating them, and until recently this required well-paid coders prepared to labour day and night behind lit screens to create the software itself, but beyond their salaries not much more was necessary. Assuming your software worked and lots of people used it the cost of acquiring each additional customer was very low. Google is an example of the economics of software development. Their search engine was excellent and therefore virtually everyone used it. Each additional user cost Google almost nothing to acquire. Remember that the code was already written, it did not need to be re-written for new users, nor was there an increasing requirement for ‘physical stuff’ as more Google searches were performed. Of course, Google was obliged to build additional physical infrastructure to power their search engine as it grew, but the additional cost in doing so was less than the revenue they could expect to earn. Furthermore, costs tended to fall across users because the sophistication of a Google Search had reached a limit. The precise information being sought was provided to users as fast as one might reasonably require, in most cases in less than a second. The incentive to invest more to improve search results was no longer present. By any measure Google won the search war. It was as good as it was going to get.
An AI arms race
Large Language Models (AI) are, of course, at their core a software technology, but the economics of adding new users is different. For the purposes of this analysis, we are treating a Google search as roughly analogous to making an enquiry from an AI. In both cases the user is looking for information. The difference is that AI is actually doing something far closer to actual cognition rather than performing a mere ‘fetch and retrieve’ exercise. There are several AI tools now in widespread use, ChatGPT, Gemini, Grok, DeepSeek etc. It is too early to say whether any particular tool will emerge as preeminent. Perhaps each of them will have a niche. Will it matter to me that you are using X or Y AI Tool? Is there likely to be a network effect akin to that which made Google search so effective? Who can say, but the backers of each AI tool are behaving as if there will be a winner. Why else would they be spending huge amounts of money (more about this later). We are witnessing an AI Arms Race.
Hardware is eating the world
With each successive successful iteration of AI, the costs per user enquiry do appear to fall. This is to be expected. However, because we are a long way from perfecting the technology, or even understanding what is possible, the incentive to continue investing is very powerful. This is where AI becomes a ‘brute force’ problem, which in technology speak is one where computation is the key constraint. Improving AI is not principally a software question, instead it is a hardware one. In order for AI to ‘think’ necessitates increasing quantities physical infrastructure, data centres, microchips, water, copper, silver and power. This is why Nvidia is the world’s largest company. They make the chips that keep the AI show on the road. Perhaps if asked now, Marc Andreessen might rightly claim that ‘hardware is eating the world’.
Is AI a Bubble?
The largest cultural event in 2025 was arguably the Oasis Reunion Tour, therefore it seems apt to answer this question with reference to their most famous album, which is of course Definitely Maybe. Without wishing to devote too much of this piece to an exhaustive analysis of the history and psychology of financial exuberance, there appears to be too much scepticism about AI for us to conclude that this is definitively a bubble. Why have we drawn this conclusion. Not a day goes past without some luminary warning about an impending crash. Previous episodes have taught us that there is usually a total lack of scepticism about the technology in question before things get really dangerous. Barring a global recession, or a significant geopolitical shock, US stocks, which are the ground zero for AI can probably grind higher. Even in the face of these risks, a sell-off precipitated by either would arguably be an AI buying opportunity. As long as investors continue to believe in the potential of the technology the sky is just about the limit. We aren’t presumptuous enough to count ourselves as among the luminaries, but we are sceptical too. Sooner or later the financial equivalent of gravity will assert itself, which is to say, investors will ask in unison ‘show us the money’. We do not pretend to know when that might be. No one does.
Capital expenditure galore
Big tech companies expect to spend in excess of $405bn this year, which is 58% more than 2024 [1] . The list of requirements to propel AI forward is endless, chips, switches, cooling units to name but a few. Even in an economy the size of the United States where the consumer is ordinarily the engine of growth, AI spending might be of sufficient magnitude to offset weakness elsewhere [2]. The stock market gives the impression of being mostly disinterested in the deteriorating employment picture and early signs of consumer weakness.
Portfolio positioning
In our previous year end missive, we saw reasons for caution across three dimensions. First, tariffs seemed likely to cause upset, and they did just that on Trump’s Liberation Day announcements. The tariff matter is not behind us. We were surprised how quickly markets rallied after the announcements. Second, we felt the Chinese had limited scope for economic stimulus and that has proved correct. Finally, like now, we felt that US mega cap tech stocks were expensive and consequently we have been taking profits were appropriate throughout 2025. The proceeds have been deployed into commodities, notably physical Silver and mining stocks too. Given what we have said above, general infrastructure and that associated with AI have also found their way into portfolios where applicable. We continue to see value in Japanese Equities, Global Healthcare and insurance. As always, our principal concern is risk-management and therefore try to avoid concentration in any one place or theme. It does feel to us as though, for the reasons set out above, that there is more interest in non-software related businesses. Granted, in some sense all businesses rely on software, but mining, specialised manufacturing etc., appear to be attracting interest.
Markets have defied our expectations this year and whilst we don’t necessarily expect the returns enjoyed in 2025 to continue at such a pace, falling interest rates could act as a tailwind If timed correctly. The usual risks, like geopolitics, remain ever present and perhaps 2026 might prove to be more volatile.
[1] https://www.nasdaq.com/articles/10-artificial-intelligence-ai-infrastructure-stocks-buy-400-billion-buildout
[2] https://www.nytimes.com/2025/11/22/business/the-ai-boom-economy.html.
