Current thoughts on the markets

In prior letters, we have explained why economic and market narratives will continue to oscillate much more frequently than before. It doesn’t always mean that markets will go up first and go down next, it mostly means that the narratives and factors that drive markets will dramatically change from one period to the next. The biggest change in market narrative over the last 18 months has been AI, specifically AI-related capital investments, which has for the most part kept an otherwise fragile economy going.

In our opinion, the AI capex narrative is currently reaching its high point, and we think it is unlikely to live up to its expectations in 2025. We have previously written about that the missing ROI link from most of the AI investments. A few months later, the ROI debate is starting to pick up steam. Commentary from Google’s CEO (that their framework has been that cost of missing out on AI investment is higher than the cost of overspending) and Microsoft’s CFO (AI infrastructure ROI will play out over at 15-year+ period) resonate with us and further provides evidence that the current level of infrastructure spending is unlikely to accelerate further.

It is important to recall that capital spending is cyclical by nature and not meant to recur every year. In current analyst models of AI beneficiaries, forecasts largely anticipate that these investments are in fact recurring. As we have highlighted in our past letters, this high level of spending can only be made by the Mag-7 companies and we believe it is highly dependent upon their stock prices outperforming.

Looking at the chart below, which highlights the Mag-7 as a % of the S&P 500, we do not think the historical pace of increase can continue. For example, between 2012-2016, the percent growth – of the Mag-7’s S&P ‘market share’ gain – was 1 ppt/year, however, between 2016-2024, the percent gains accelerated to about 3 ppts/year. To us, this new rate looks unsustainable as it would imply the Mag-7 will become two-thirds of the S&P within the next 10 years.

There are societal implications to these concentrations as well. Consider consumer choice, security of data and information, and the ability to sway populations through misinformation in any one direction. We don’t think that society can afford to have the Mag-7 control most of the economic benefit of digital transformation and impact social changes at such large scale they are operating at. Market forces will likely not allow this to happen.

AI Will Not Kill Software

In recent discussions, there has been a prevalent narrative among well-known VCs and industry pundits regarding the perceived obsolescence of software companies due to the advent of AI. The argument posits that software lacks a defensible moat as AI possesses the capability to effortlessly generate code. We disagree.

Historically, similar prognostications were made with the rise of open-source software, yet the outcome was contrary to the predictions. Today, it is commonplace for companies to offer an open-source version of their commercially available software, and there are widely available libraries and environments to download and leverage free software that a customer wants. However, the intrinsic value of a software enterprise extends beyond mere code generation. In fact, the advent of open-source served to amplify this value. The needs for interoperability, support, and service will persist irrespective of the ease with which one can instantiate a C++ or Java module.

The essence of a software company's worth is anchored in its integrations with the existing enterprise tech stack, the interoperability between various workflows, and the capacity to provide timely support to customers. Additionally, the customization of software to meet the unique requirements of each client is a significant value driver.

In the near term, we do see the seat-based software vendors getting marginally hurt as AI helps improve productivity and reduce hiring of new software programmers. However, as many of the current AI related trials come towards full scale adoption, the needs for hiring new engineers will rise again. Over the last decade, the emergence of cloud platforms has been a pivotal force in accelerating software adoption, and similarly, we believe the emergence of AI platforms could serve as a catalyst for further software adoption within enterprises. This implies that software companies stand to gain during the next phase of AI adoption, once the foundational models and infrastructure of AI reach maturity during the initial wave.

The most transformative aspect of AI lies in its democratization of computer interaction, making it accessible to the average individual. As interactions between users and computers proliferate, there will be increased reliance on new software tools, along with the demand for sophisticated workflows, interoperability, and customization, which is poised to escalate.

We are confident that the trajectory of the software industry is not one of decline but of renewed importance and integration within the fabric of the IT ecosystem.

Incoming Slowdown of AI Capex Will Provide Room for Software Investments

Mag-7 executives and many prominent VCs have done a disservice to society by creating so much hype around AI. Though exciting, technologies like these need time to mature and evolve. Of course, it will only be known in the future, but we feel that you cannot accelerate tech adoption by brute force (in this case, massive capex investments). In fact, in this case, the entire market and economy is unlikely to make the fast leap forward as the constant discussion around NVDA’s stock may have everyone believe.

As we look out in the near term, the two elements that can anticipate a slowdown in AI capex investment are going to come either from the large model companies or from the enterprise customers. The early adopters and the model companies (who have an inherent interest in selling the latest and greatest) will likely continue to push the envelope but our discussions with the mainstream companies are highlighting that most have not crossed the chasm yet, beyond a mandatory trial/concept phase. As the noise gets louder around ROI, and datapoints scarce, there will be less airwaves/momentum to sustain the hype.

All Roads Lead to US Elections and Path Forward on AI

First, let’s talk briefly about the upcoming US elections. We continue to believe that the outcome of US elections has the potential to, and is likely to, alter the current market narrative. We do feel that conditions are in place for a factor rotation and the elections can provide the much-needed catalyst for that rotation. The rapid rise of the FAANG/Mag-7 (and broadly large cap thematic investing) has almost become sacrosanct investment wisdom, and this market narrative has accelerated over the last 8 years driving an uncomfortably high share of the S&P and the economy. In our opinion, it would be very unlikely that this acceleration sustains at the same rate in the future because many critical factors are aligning: impact, valuations, positioning, and catalyst.

Secondly, AI, of course, has been one of the most popular debates of our times, causing constant excitement across all dimensions. Never before have we seen such engagement from all levels including businesses, consumers, education institutions, government and more importantly everyone in the investor community. AI is important for us to discuss because in our view it has single handedly (or at least has been one of the major reasons besides government stimulus) helped prevent the US economy going into a recession in the last 18-months and driven the animal spirits. There are also trillions of dollars that have been put on the line for AI (either in terms of capex, or investments, or assigned valuations). The trajectory of AI’s future has important consequences to the much broader ecosystem than merely an interesting technology discussion, and hence we are choosing to spend substantial part of this letter on discussing AI.

These two important market themes (Factor Rotations and AI) and their dramatic ascent in importance, point to one of our important theories (frameworks) that we have outlined in prior letters, which is, that over time, market narratives will frequently change but will also exhibit higher and higher amplitudes given the fundamental shift in the increased connectivity/communication of information/data, overlapping with the breaking down of the traditional ‘capital pyramid’ (that we discuss next).

Blurring of Lines in the Capital Lanes

Our frame of reference is that there should be a clear separation between high level thematic or opinion layer vs. data or financial driven operating execution layer especially as it relates to capital and investment. We can frame this as a ‘capital pyramid’ – where the largest amount of capital should typically be invested in the most concrete opportunities with clear financial metrics and high certainty of successful outcomes, while relatively smaller amounts of capital should typically be invested in high level opinions and themes which are much higher up the risk curve and can have a vast variability in outcomes.

Venture capital (typically a small percentage of total system’s deployed capital) was clearly assigned as the “asset class” to bet on 10+ year far-away risky ideas driven by themes, TAMs, and opinions. Most other investing asset classes (which generally have been a larger percentage of total capital deployed) would typically focus on tangible results-driven or demonstrated financials driven projects (recall ‘cash flow-based value investing’). This should be simple stuff, nothing earth shattering, so let’s discuss why this has now become an issue.

Our observation is that the above paradigm has changed dramatically over the last several years. Given zero barriers in information flow, increasing difficulty in assessing facts from opinions, and tremendous amount of availability liquidity, all of these aspects together have now created a complex situation where the relative separation of the asset classes in their supposed investing lanes is going away, which, to a certain degree, is creating a huge misallocation and changing the shape of the ‘capital pyramid’. We have definitely seen that in the debt markets (decline in mezzanine-type funds), rise of private credit taking over tranches what used to be underwritten by banks, the rapid rise of crossover vehicles (publics/privates), VCs becoming late-stage asset managers (Sequoia is a good example), family offices doing direct deals, the rise of co-investments, and the list goes on. These are just a few signs to show that traditional barriers of separations are dissolving; hence, gigantic pools of capital are sloshing in areas that are structurally not prepared (or suitable) for the 10x capital infusion. Add to this dynamic the unprecedented fiscal stimulus as well as the government involvement, and it can help explain somewhat the pockets of craziness in capital markets as well as puzzling valuations of private assets. We tend to believe that AI is one such theme which ended up receiving a massively disproportionate capital flow due to the above dynamic. This mismatch of supply-demand driven by misallocated large capital pools artificially creates hypes and parabolic moves of large amplitudes, that inevitably leads to eventual disappointment. We saw that during the 2019-2021 era, when drastic amount of crossover capital led to dramatic growth rates in growth/later-stage tech companies that shouldn’t have been growing 30-80%, or the number of SPACs that went IPO during the 2021 timeframe (most of which have now destroyed returns for anybody investing in ‘here and now’ craze).

It is important to understand these types of changes in market structure and the capital vs. information flow, and how these structural changes helped create this environment where rapid rise of narratives and themes can somehow mobilize enormous public opinions as well as enormous dollars at a scale never heard before. When we are discussing AI or flying cars or space travel, it is with this lens that we are viewing and understanding these developments (and pontificate on what might be in store in the future). When these hype cycles are running, their amplitudes will convert the biggest of the naysayers and non-believers. However, these narratives will turn just as fast because the underlying fundamentals are almost always unlikely to match the parabolic increase in expectations. As we described in a recent podcast interview, “capitalism does not like parabolic moves”. If you recall the rise of Web3, or blockchain, or ESG, or NFTs, or EVs, they all would have made you believe that these narratives were truly going to take over the world. However, eventually, these narratives encounter the harsh realities (usage, adoption, ROI, workflows, habits, etc.) that bring about the reversals while also creating massive destruction of speculative capital.

So, what’s our key takeaway? We encourage investors to not believe or invest in everything exciting they hear and see, especially from the very people who are supposed to benefit from the hype. More capital typically gets destroyed rather than created in parabolic moves because the largest dollars typically get invested at the peak of the hype. For e.g., there have been reports that ARKK ETF funds have destroyed $14B+ of capital over the past decade despite a parabolic move during 2020-2021. Remember that patience is key because the rapid rise and fall of thematic narratives will continue to happen given what we’ve outlined above.

The Artificiality of AI

While there is no shortage of content, opinions, or reports on the topic, we encourage you to listen to the podcast interview below, where Scalar Gauge dives deeper into the markets and our views on AI. Please see the link to episode online, Apple, or Spotify.

We know that AI has captured the imaginations of people given the shock value many experienced during their first interaction with ChatGPT. The generation of videos, images, and well laid out text summaries appear magical! It’s not hard to start dreaming how it can solve all the problems. It appears AI has all the answers, and its very nature of effortlessness speaks so well to the inherent lazy nature of human beings. If there was just a simple way for us to say the question and get all the answers from an agentic AI, we can all usher into a new world of human masters running AI agents everywhere to do all our mundane and dirty work for us, which can drive insane productivity gains and accelerate GDP growth massively…right? Not so fast!

Our slight dis-belief against the AI super-believers is not about the technology aspect of AI (which is amazing), but more about the full-on dream mode where AI is expected to be driving significant and measurable improvement in GDP growth. We don’t believe that is an easy task based on our study of past hyped-up parabolic behaviors and understanding of investor psychology. Our issue is not that AI isn’t going to help improve productivity, but it’s the constant drumbeat that AI is the most important discovery and game changer that’s bigger than the internet, giving rise to a parabolic hype curve which we are worried can drive large scale capital destruction in the future, if the future results don’t match the heightened expectations (which is likely).

AI is just a good tool, no more no less, just like Microsoft Excel has been an excellent tool for human beings to do our tasks better. But we don’t call Excel to be the next industrial revolution. When people start comparing AI to be much bigger than the internet or semiconductors (which perhaps have been the most groundbreaking innovation in our lifetimes!), our BS meter perks up. Yes, of course, Jensen has to go around the world and meet all the governments to show them the 20-year vision of everything run by AI so they can make massive investment in GPUs and help sustain Nvidia’s valuation, but that doesn’t mean his vision will indeed be correct (or at the very least will not face major setbacks). Let us make a few simple points below:

We believe that a business has to offer a certainty of outcome (otherwise, it is a hobby). The notion of ‘get it done’, execution, and 100% delivery all arise from the fact that even if your outcomes are 90% or 95% correct, it is perceived as not reliable and very few people will pay for that product or service. When you turn on a TV or a phone or car, it turns on not 95%, but 100% of the time! When you search something on Google, it gives you the right answer in the first couple links almost 100% of the time. That’s why users have trust and reliance in these products. If your car won’t turn on even 1 day a month (3% error rate) in the morning when you have to drop kids to school, it will be a nightmare. The point being that for a business to be paid consistently, it has to deliver its output close to 100%.

It is the very nature of AI tools to not be accurate or without exorbitant costs/manipulation there is no way to push them to be 100% accurate. It is great for media generation or creative outlets where accuracy isn’t needed, but for the majority of enterprise use cases, 100% reliability is important, and we don’t think AI can get there without the underlying uneconomic cost.

The biggest flaw in the argument that AI will overtake human jobs is that somehow whatever the question is, the argument assumes that the data already exists, or the answers can be easily inferenced. All what AI has to do is just go and look through in a big pile of transformed words, dig out that answer, and articulate it in a clear concise fashion. The key underlying flaw is that the true answers (or data to give the answers) to most of the real world questions don’t exist at all (or at least not in digital formats). Either the data is simple enough to exist within the corporation (e.g. financials of a company in 10-K) where existing tools can just grab the data (and AI layer mostly is a thin UI interface layer), or the data is complex enough that AI can’t really access what is needed hence answers continue to be very generic. Although we can understand that there might still be plenty of jobs where simple automation may be a value add from AI (e.g. in customer calls through AI bots can skip the “press 1 for …” but we don’t think the premise of Agentic AI gets realized by just offering a thin UI veneer.

The main job human beings do in any organization – is to come up with answers/execution by going across various mediums, variety of sources, inside/outside the corporate walls, resolve conflicts within different people/parts of the organization and demonstrate leadership at their respective levels. We do understand there may still be plenty of simple sub-categories of jobs that might benefit from some AI automation (read tools) however, that description does not equate to “the next industrial revolution”.

We also believe that throwing more compute isn’t going to improve AI models beyond a point, because the underlying quality of data becomes the biggest bottleneck.

AI has delivered output at a breathtaking pace, because the first step of data collection was easy (freely available data on world wide web, especially in the fields of education, text, images etc.). First came quantity, which delivered the wow factor, but the second step now needs quality, which is a different kind of battle.
The next set of data which really drives the output and value (enterprise data, and data embedded in the workflows for people/organizations is scarce and not even digitally captured in majority of cases). Hence, for Agentic AI to learn those workflows on the same scale (as the internet) is highly demanding and will be a very slow process.

AI will continue to have Inaccuracies and hallucinations. Much has already been written on this topic. Just as an example, there has been a recent UMass-Amherst study highlighting how dangerous the results could be:

Hallucinations in AI-generated medical summaries remain a grave concern (clinicaltrialsarena.com)

AI was asked to create 500-word summaries of 50 detailed medical notes. Research found that GPT-4o had 21 summaries with incorrect information while Llama-3 had 19. That is a staggering rate of inaccuracy in a field that can’t afford to have inaccuracies.

Another important aspect here which is much less discussed is the fact that engineers typically focused intentionally for these models to do well in respective benchmarks but when used in real-world, the observed performance is much lower than discussed in benchmarks.

In the near term, AI will add value to internet information summarization, content creation, education, consumer centric use cases and coding productivity enhancement. Our conjecture is that horizontal AI does well in consumer/entertainment centric use-cases, while vertical AI will do well in select enterprise use cases. We do see AI benefitting large scale consumer oriented internet companies (ability to add a fun feature and amortize over billions of consumers who are not looking for a critical ROI on their time spent with AI), as opposed to enterprise oriented workflows. We do believe that in the long-term, in an enterprise environment, the value is more likely to be accrued to the application/software layer within select verticals, rather than at the capex/data center/horizontal model layer (which is where most of the value accretion has happened so far but will likely diminish in the future). Applications will have differentiation around compliance, as well as security. Application logic can be customized to a use case which can deliver powerful ROI. Applications can have sophisticated cognition and routing, incorporate RAG databases and most importantly organizational workflow. All of these are required to deliver end results and value to the end user be it consumer or enterprise. The value will migrate from the current world of horizontal models and infrastructure towards end-applications for specific verticals that will do problem solving through the workflow of the customer.

AI over time will be relegated to become a tool to help improve productivity of many jobs, however, the impact will not be close to the gigantic revolutionary platform shifts in the past (e.g., semiconductors, internet, wireless etc.). If that is the key premise, we can look at the GDP growth over the last 30 years and see if any of these gigantic waves of innovation ever really drove the GDP growth beyond the normal rate.

Below is the GDP growth rate and we don’t think it is easy to point out any timeframe (outside of COVID bounce) where you can see disproportionate impact from technology waves or technology revolutions of the past.
Our point is that if AI is unlikely to achieve the level of impact that past groundbreaking technologies like the internet or semiconductors did, and even those tectonic shifts couldn’t really accelerate GDP growth rates, then, the idea that somehow AI will drive much higher GDP growth rates through massive productivity gains is a very low probability scenario. (US GDP growth rates below)