LIVE DIARY--FOLLOWING THE SPREAD OF AI ON MY INVESTMENTS

 

DAIRY ON THE SPREAD OF AI

16/2/26

I am writing this as an ad hoc diary as we progress through the world of AI’s impact on the investing markets. I plan to add weekly, but it could be anytime. The impacts of AI are difficult to accurately assess, and this diary, in part, is to monitor how my views change over time. The impacts on the portfolio are already large, volatile and surprising.

Over the past few years, most of the gains in the portfolio have come from exposure to what I describe as “light” AI, being the hardware that builds the FABs and the hyperscalers that monetise the activity of token manufacture and delivery into use cases. The largest holdings have been TSMC, GOOG, AMZN, MSFT and ASML. The second driver has been specific Australian growth companies where the business is growing quite well and will certainly be much larger in 5 years, but the SP usually incorporates full value for these companies. Now and then, there is significant volatility, which allows opportunities to add or sell some. These companies include PNI, HUB, LOV, RMD, REA and to a lesser extent, NCK, CAR, ALL and TNE. The broad sectoral exposure is considered good. The rest of the portfolio has been a wash at best and a waste of time, probably more accurately.

Recently, with the software sell-off, I have added some exposure. These are the front line against AI, with the fear that software becomes redundant and all activity is generated by AI agents interacting between the client and the LLMs, leaving out the software provider. The AI fear, that is, that business models will become redundant and replaced by AI agents, has spread across various sectors. These include various forms of proprietary information providers, platforms that rely on network effects and even insurance brokers; in aggregate, this has taken a large toll on the portfolio. By its very nature, AI agents will attack high-margin information flows or services. The portfolio is, by choice, exposed to high-margin protected and growing businesses.

Market movements have also thrown up an inconsistency in logic, which is most probably due to extreme uncertainty over the outcomes. One possible, as they are all just possibilities at this stage, is that the inconsistency is in having the hyperscalers increase their capex budgets enormously. They have done this in response to immense and ongoing demand, and after seeing attractive returns on the older vintages of capex, as I described in the big tech talks. The market has been very cautious on this spend and has marked the SP down, indicating poor returns on the capex. At the same time, some of the apparent victims of this capex have been marked down as well. So everyone loses. Although that is an unlikely outcome, anything may be possible. IMO it exhibits the market's fear of uncertainty and prices that in illogically, until further evidence is shown.

There are a lot of circular discussions that can go on without end, as the outcomes at this stage are too wide to determine. So instead, I will briefly sketch out the base investment thesis and see how that changes over time.

The infrastructure builders have been the largest beneficiaries, and understandably so, as the huge capex spend ultimately ends up in their bottom lines. These include portfolio holdings, ASML and TSMC. In a broader sense, they include the chip designers, the networking manufacturers, the power suppliers and any specialist items needed in a FAB. The largest companies are NVDA, AVGO, AMD, etc. These are clearly the short-term investment plays. How long is short-term? Well, it could be quite a while if AI is a secular change across the whole economy. Ultimately, margins will probably spread across the chain, whereas at the moment they are concentrated in the hardware, but some exposure here is the lowest risk as long as capex continues to expand.

The next are the DC owners and operators, or hyperscalers, defined as businesses in storing and moving data, where more of this activity is AI-generated for themselves as well as third parties. Over time, the third-party business is expected to be a much larger part of the business. The big question here is ROI holding as the business scales up to enormous amounts of capital. The three largest are GOOG, AMZN and MSFT, all having extraordinary financial power which is required in this case. The thesis is that token usage grows enormously as AI spreads and that the returns on capex become apparent and are very lucrative. There is a good chance of an effective oligopoly forming here, although that is not in the interest of participants that dominate other parts of the chain, eg NVDA, who will actively try to undermine any competitive advantage in parts of the chain in which they do not play (e.g., fund competitors).

The next layer is the applications run by LLMs. The race for LLMs has some way to go, but it is becoming clearer that three are pulling ahead, being OpenAI, GOOG and Anthropic. The Chinese are an unknown potential, and other models such as the Musk group, Meta, and other smaller models appear behind, but this space is enormously unpredictable. The LLMSs are becoming large customers of the DCs, given their enormous usage of tokens. Enterprise, as opposed to consumer, is becoming the successful business here, but the ability to integrate into underlying businesses is critical. That is, make the LLMs useful in the real world.

The application layer is expected to be the main game over time. This is where the LLMs interact with the real operating businesses. Almost all players mentioned will have some involvement, and this is where the uncertainty of the incumbents arises.  The basic threat is the LLMs, and the agents developed from them effectively circumvent the incumbent software business, either making them redundant or effectively a low-margin facilitator. That is the risk that is being reflected in many businesses.

In theory, agents may run everything, and they become instrumental in all transactions and actions, and human involvement is nothing more than monitoring and maybe having some kind of veto. We do not know. However, theory and real business can deviate. The challenge for software businesses is to adapt and be the point of interaction with the client, and not lose that to a third party.

Below are the main issues and my current view,

Many in the market have pointed out proprietary data as a defence for the incumbents, and that is a reasonable argument. It rests on the agents not having access to the data. As long as that data stays proprietary, the software company has the upper hand. The data, however, does belong to the client, and if it opens that data, does that level the playing field? To me, this looks like a medium-term shield, maybe giving the incumbents some time to integrate and make client shifting less likely. There may be margin compression here.

Another issue is the size of loss on a wrong call by the AI agent, or the consequences of an error. AI agents work on probabilities, which in a lot of cases is close enough. If it is a low-grade administrative function, an error may be tolerable; if the activity can result in a possible business failure, even if a low risk, there is a better chance for the incumbent. I have a lot of time for this argument. Theory and business reality can differ. If something is cheaper and faster but has a slightly higher chance of error, and that error has devastating consequences, then using AI agents more readily is less likely. I would look to tax, insurance, and life-threatening work as examples. Of course, where that line is drawn, how many clients consider the risk devastating or not, is important. Im willing to bet here.

The challenge to the network effect is another issue to consider. The online marketplaces are a reasonably large part of the portfolio, and the moats here have been strong. Will clients move en masse to create another marketplace? There have been many attempts at this, and AI is another attempt. Getting volumes is critical. More eyes lead to more eyes. The moves by Amazon to limit AI crawlers on its site are instrumental. A deal with these crawlers will probably have to be met with some type of fee sharing. Following traffic will be critical here. The collapse of platforms is considered quite remote at this stage.

Another issue that is causing a lot of consternation at this stage is the seat-based models that many SaaS companies operate. As the volume switches to AI users, presumably being fewer but more active, while human seats slowly (or quickly) disappear, will promote a change in charging. This does not look like an insurmountable change to me, but it depends on execution and transferring the client base without large leakage. Probably a case-by-case outcome.  

Payments are another area that has come under some scrutiny, as AI traffic increases at a very low level. Most of these are moving to incorporate agents as customers, early days here. The uncertainty is impacting stock prices.

The portfolio is not exposed to many SaaS companies, with the larger holdings including MSFT, which has other large businesses, and TNE is the largest. The main exposure for the portfolio is the hyperscalers, which are yet to prove the returns on capex, and as noted above, the market is quite cautious here and also cautious on the “losers” of this capex generating ROI. The platform companies are in larger positions, and we can see how they keep customers.

As can be seen, the portfolio is exposed, mainly, to the second layer of companies that could be affected. Hyperscalers, platforms, insurance brokers and payment companies. The outcomes for these companies and whether they are AI “losers’ is more nebulous than the directly exposed SaaS companies. Overall, I think the portfolio is not overly exposed to the “pointy” end of potential AI disruption, but has large “secondary” exposure, which I think is reasonably defensible at this stage.

 

20/2/26

Anthropic raised $30B at a $380B valuation. Since Anthropic was the main initiator of fear in the market and was about to go into a major raise (maybe the biggest raises ever, until OpenAI comes again), it was highly incentivised to talk up use cases, interesting to see if some heat comes out of this.