In 2017, a team of researchers at Google published a paper that may have altered the course of human history. The innocuous-looking 15-page document was titled “Attention Is All You Need”, and it studied process improvements in machine learning techniques applied to a rather uninspiring task: German/French to English translation, yet the concepts it introduced would be largely responsible for the subsequent development of generative AI in its current form. Today’s Large Language Models (LLMs), such as ChatGPT, are in large part implementations of the “attention” mechanism and “transformer” neural network architecture described in the 2017 paper.
Fast-forward to 2024. OpenAI (the creator of ChatGPT) went from being a relatively unknown AI project to raising 6.6 billion dollars at a valuation of $157 billion, while Nvidia, the leading manufacturer of GPU cards (the computing hardware used in generative AI), has grown 3690% to become one of the largest companies in the world.
While my limited understanding of machine learning prevents me from fully appreciating the technical details of “Attention Is All You Need”, one notable feature of the work is its elegant simplicity. For such a paradigm-shifting piece of research, you might expect some sort of wild technological breakthrough, or obscure mathematical prowess. Instead, “Attention Is All You Need” innovated by simplifying an existing neural network methodology, making it more efficient, removing layers, and focusing on what mattered. As the paper itself states, “[we] propose a new simple network architecture (…) based solely on attention mechanisms, dispensing with recurrence and convolutions entirely”, doing more with less.
The Importance of Growth in Technology Infrastructure
But AI today is not just about fancy software and academic research. Training and using transformer-based models requires a lot of specialized computing resources. At the core, AI needs huge, industrial-scale investments in computing hardware and energy infrastructure: endless rows of servers stacked in nondescript warehouses, guzzling megawatts of power. At ground level, the AI revolution has mainly taken the form of a massive increase in data center growth. Huge investments are being made in the computing power used to deliver generative Artificial Intelligence tools in various forms.
These investments have real economic impact, transforming entire towns and regions, and providing high-quality jobs across the country. Here in Wisconsin, Microsoft is planning to invest over 3.3 billion dollars in its new data center in Mount Pleasant. The plan also involves partnerships with local universities and a training program for over 1,000 workers, a development heralded by Governor Evers as a “watershed moment for Wisconsin”.
For the end-users, the hope is to embed AI tools in just about every business process, dispensing with the vagaries of employee performance, improving speed and enhancing reliability. Potential applications exist in every sector of the economy, even in the most advanced scientific domains. Demis Hassabis and John Jumper at DeePMinds were recently awarded a Nobel Prize in Chemistry for their work using an AI model known as AlphaFold, that can predict the three-dimensional structure of a protein from a sequence of amino acids, successfully solving a problem that chemists wrestled with for over 50 years. If you are wondering, AlphaFold is also rooted in the transformer mechanism introduced in “Attention Is All You Need”.
In the business world, early results are also promising. Honeywell, a large industrial conglomerate, has reported achieving productivity gains equating to adding 187 full-time employees by incorporating Microsoft’s Copilot AI technology. Microsoft itself reported a 42% increase in employee HR query resolution accuracy and a 60% decrease in cash collection delays in early results of its AI powered autonomous agents. Jensen Huang, the CEO of Nvidia, said that building the firm’s market-leading products “is no longer possible without AI”. In other words, today’s AI is already building tomorrow’s AI computing tools.
With the potential for such incredible breakthroughs, the stakes could not be higher, the goal is to replicate the success of Alpha Fold in every domain of business, science and life. Considering potential knock on effects in fields like automation and robotics, the possibilities offered by AI are endless, a revolution potentially even greater than the internet, perhaps on par with something resembling the industrial revolution, but with a sprinkling of sci-fi.
In support of this cause, Sam Altman, the CEO of OpenAI, recently attended a White House meeting and presented a plan titled “Infrastructure Is Destiny”. Standing next to a portrait of Franklin D. Roosevelt, he pitched high-profile attendees (including Commerce Secretary Gina M. Raimondo and Jake Sullivan, the national security adviser) on the idea that AI data centers would be a catalyst for the re-industrialization of America, creating over half a million jobs, and called for investments in a new generation of data centers costing 100 billion dollars each, dwarfing Microsoft’s suddenly insignificant Wisconsin project.
As AI continues to grow, it will need more, a lot more: more energy, more compute, more data. Today, with an incoming Trump administration promising a new wave of de-regulation, AI enthusiasts may feel that the glory days of AI are upon us. Sam Altman recently announced a $1 million donation to Trump’s inaugural fund, declaring that “President Trump will lead our country into the age of AI”.
The Year of Financial Market Momentum
Not that everything is necessarily all rosy on the AI front. The race to scale generative AI by pouring more and more resources into it is far from unanimously popular. Concerns from critics range from environmental impact, wasteful energy expenditures and even concerns that this is simply not the right path to AI improvement and that new areas of research are needed.
Trivia: What is Warren Buffet’s current largest portfolio position?
While AI has been growing at breakneck speed, there are already signs that the pace of improvement could be slowing. The jump in intelligence from ChatGPT 2 to version 3 and then version 4 has been mesmerizing, but the much anticipated “GPT 5” model has failed to materialize. Instead, OpenAI seems increasingly focused on smaller incremental changes to its existing model. In a recent interview, Matai Zahira, the CTO of Databricks, reports instances where “you double your training costs, you increase quality by only around 1% or something like that”. Echoing this sentiment, Sundar Pichai recently mentioned that all the “low hanging fruit is gone” in AI development and “the hill is steeper”.
None of these concerns have deterred investor enthusiasm. In recent years, investing only in a handful of the world’s largest stocks, and particularly those in the tech sector most connected with generative AI, would have been the winning strategy. While sector outperformance in itself is nothing unusual, the extent and scale of this particular episode has been truly remarkable. In the two years since ChatGPT was released, a portfolio composed only of an equally weighted of Nvidia, Meta, Microsoft and Google would have returned 266%, compared to 46% for global stocks; 34% for a 60/40 portfolio; and 60% in S&P 500 stocks and the rest in bonds.
Risk-adjusted measures of performance, such as Sharpe ratio (a measure of the amount of returns per unit of risk), further illustrate the staggering price momentum behind these stocks.
An equally weighted “AI portfolio” composed only of 4 stocks previously mentioned would have produced a Sharpe ratio of 3.4, compared to 1.5 and 1.7 for our two diversified benchmarks. For reference, even hyper-sophisticated investors like hedge fund managers would be thrilled with any Sharpe ratio greater than one. A Sharpe ratio over three would place you in the rarified air of the most efficient forms of investments our financial markets have been known to achieve, ever, and required no sophisticated trading algorithm, you simply had to follow the trend, ride the momentum.
The Impact of Technology on Investment Portfolios
In the two years since ChatGPT was released, a portfolio composed only of an equally weighted of Nvidia, Meta, Microsoft and Google would have returned 266%, compared to 46% for global stocks; 34% for a 60/40 portfolio; and 60% in S&P 500 stocks and the rest in bonds.
In the financial market, just like in day-to-day speech, momentum means continued strength: the stocks that went up the most in the recent past continue to outperform. In finance, it’s also an investment factor that can be measured and quantified. According to data collected by Piper Sandler, 2024 has been the year of momentum investing, with momentum dominating factors like earnings growth or return on equity. Momentum has been the only game in town.
The beauty of momentum is that it tells a simple, incredibly convincing story. It’s self-validating and re-enforcing. If in doubt about what the best investment is, one can simply point to a chart of recent performance to make clear exactly who the winners are. For the most part, investors got the message loud and clear; with markets at all time highs, retail investors are staying long and strong. According to data from Goldman Sachs, equities now represent the largest percentage of US household’s financial assets ever, surpassing even the late 90’s internet bubble.
But it isn’t only retail that is on board, according to Bank of America’s global mutual fund manager survey, levels of cash holdings (usually held as a defensive position to protect against volatility) are now at some of the lowest levels ever recorded.
U.S. Stock Market’s Performance Outpaces GDP Growth
In reality, the latest bout of momentum driven frenzy is only the latest in a long series of events that have benefited growth stocks, tech companies in particular. Tech-lead outperformance really traces its roots all the way back to 2011, the beginning of the Quantitative Easing. Then came the COVID lockdown that pushed millions into forced technological adoption to communicate and work, and now, AI promises the dawn of a new economic era altogether.
“What surprised me during my research was realizing just how much the narrative had shifted—just a few years ago, everyone was incredibly bearish, but now, mutual fund cash levels are at historic lows, showing dramatic optimism.”
SYL MICHELIN, Walkner condon – U.S. Expat financial advisor
The decade plus long of barely interrupted rally in stocks has lifted the market to some of the highest valuation levels ever recorded. In 2018, when Apple became the first company to reach a market capitalization of $1 trillion, it felt like it was perhaps the destiny of a uniquely successful company and brand, to be matched perhaps only by Amazon. Today, Apple is no longer alone, and the trillion dollar companies are more a club of sorts, a very tech oriented one at that.
The decade plus long of barely interrupted rally in stocks has lifted the market to some of the highest valuation levels ever recorded. In 2018, when Apple became the first company to reach a market capitalization of $1 trillion, it felt like it was perhaps the destiny of a uniquely successful company and brand, to be matched perhaps only by Amazon. Today, Apple is no longer alone, and the trillion dollar companies are more a club of sorts, a very tech oriented one at that.
Other than Saudi Aramco (a different topic altogether), the one member of the trillion dollar club that stands out the most is perhaps its latest entrant: Warren Buffett’s company, Berkshire Hathaway. What stands out is that Berkshire has a decidedly non-techy feel as a company. Warren, known for his old-fashioned and no non-sense approach to investing, prefers simple businesses and avoids technologies he doesn’t fully understand. He has historically been very shy about making bets on technology. There is obviously more nuance to Warren Buffet’s investment philosophy. After all, his largest holding by far is still Apple, but even then, Tim Cook himself largely attributed Warren’s Apple investment to his understanding that Apple is a consumer business at its core.
– Tim Cook, APPLE – CEO
It’s notable that Berkshire, which operates as a holding company for Warren’s investments, does not own any shares of Nvidia, Google or Microsoft (other than nominal amounts in S&P 500 tracker ETFs). Even his enthusiasm for Apple seems to have been waning as of late, as he sold over 600 million shares in 2024.
If we are truly at the dawn of a new industrial revolution, and if following technology-led momentum is the only winning strategy, Warren, it would appear, didn’t get the memo. In fact Warren’s current investments interests seem very focused on a firmly non-momentum asset: cash. In contrast to mutual fund managers, Warren is now holding more cash than ever before, and his war chest recently reached over 300 billion dollars, representing his larger individual holdings.

Perhaps, the simplest explanation for Warren’s current cash binge is that stocks are simply too expensive, that the cost of entry into this new AI world is simply too high, skewing the risk/return dynamics unfavorably. One example of today’s exorbitant prices is found in one of Warren’s favorite valuation metrics: the ratio of stock price relative to the size of the U.S. Gross Domestic Product. By that measure, stocks are grossly overvalued, more so than even in the late 90’s bubble.
Simply put, the U.S. stock market’s performance is racing ahead of GDP growth, and increasingly it has also been racing ahead of forward looking earnings estimates, as shown by the widening gap between the two lines in the below chart.
Momentum, as we see it in today’s markets, is largely unburdened by immediately observable metrics like earnings or GDP, and instead follows its own internal logic. In a way, momentum is a bet on the unseen, on horizons bigger and broader than can be captured in data, perhaps, on the dawn of a new era of technological revolution that the likes of Warren Buffets simply fail to recognize.
How to Navigate a Shifting Market – Momentum vs. Valuation
In a momentum market, valuations take the back seat. To put it bluntly, valuations don’t matter. For the last three years, I’ve been fairly consistently positive about the market, both in client discussion and in our firm’s annual outlook report. Yet my bullish inclination never truly hinged on valuation, which has been elevated for a long time, but rather on the fact that a strong economy can support elevated stock prices, at least in the short term. As I highlighted in both of my 2023 and 2024 Outlook Reports, the resilience of the U.S. economy, in particular the strength of our consumer and job market, has underpinned a lot of the markets’ strong performance, long before momentum seemingly took over.
Over the course of 2024, retail investor enthusiasm about the stock market has been accompanied by a steadily declining confidence in their employment income. According to University of Michigan survey data, Americans now only see a 43% chance that their income will rise in the year ahead, a level on par with some of the lowest readings ever recorded in this series, lower even than during COVID lockdowns.
In contrast, according to Bloomberg’s long-running conference board survey, U.S. consumers have never been so confident about the stock price increase. Since the survey started in 1987, confidence rates have averaged about 36%, today they stand at a record breaking 56%.
This interesting dichotomy challenges the old investment adage that “the consumer is the economy”. Despite markets soaring to new heights, and retail enthusiasm, 2024 also brought signs of deterioration in the employment picture. The unemployment rate in particular, which had been declining consistently since COVID (or, ignoring the COVID blip, since 2009) is now rising again.
While the trend is still relatively muted, it may be enough to explain the consumers lack of confidence in their earnings power.
Credit card delinquencies also appear to validate consumer concerns about their finances. Despite endless self-congratulation from central bankers about slowing inflation rates, consumers are still feeling the pinch, with delinquency rates on credit cards rising to decade highs.
Some may be tempted to see a connection with AI here too. After all, one of the often-cited risks of AI is rising unemployment: as AI adoption accelerates, jobs could be replaced, leading to mass unemployment. In that scenario, Wall Street gains could go hand in hand with main street pain. Sam Altman has long been a proponent of UBI (Universal Basic Income) for this exact reason.
If this was happening, you’d expect to see the trend capture in one specific piece of data: labor productivity. Labor productivity measures the amount of economic growth output attributable to the efforts of workers. If AI was replacing jobs efficiently, you’d expect the same (or more) amount of growth to be achieved by a shrinking number of workers, leading to a rise in labor productivity. While there have been sporadic signs of strong U.S. labor productivity in recent years (I actually highlighted productivity rebonds as a strong disinflation contributor in my 2024 market outlook), recent numbers have failed to really validate the encouraging 2023 data.
If AI led productivity growth, you’d also expect the more technology oriented sectors to lead productivity growth. Instead, reality on the ground tells a very different story. Productivity in the U.S. has been largely led by the oil and gas industry for at least a decade now and the trend shows no signs of stopping.
Instead of AI, it appears that a very different form of innovation is driving the US labor productivity story: horizontal drilling and fracking.
In a recent policy brief, the Economic Strategy Group of the Aspen institute concluded that “[generative] artificial intelligence (GenAI) is likely not a driver of this recent upswing in productivity growth”, crediting overall business dynamism and business creation instead. While isolated success stories have been reported, the overall impact of AI on the economy as a whole remains small. A study of Fortune 500 companies by the Census Bureau’s Business Trends and Outlook Survey found that only 3.8% of businesses actively used generative AI to produce goods and services. Other studies have found much higher rates of adoption, but often relied on much looser definitions of “adoption” and may have captured even casual usage, such as using ChatGPT to review and email.
While somewhat encouraging in terms of adoptions, these numbers appear very low when you consider that a recent study of the same group of companies found that AI was mentioned in 80% of earning calls, suggesting that talking about AI is growing much faster than actually using AI.
Despite reports of jobs lost to AI, it therefore appears that the much feared AI-led job armageddon is simply not happening. Instead, what we are seeing is likely to be a rather typical form of economic slowdown. The much talked about IA adoption is surely happening, but much more slowly than you might infer from media headlines, or CEO chatter.
AI Models – Probabilistic Hallucinations vs. Financial Market Reality
Part of the reason behind low adoption undoubtedly comes down to limitations of existing AI models. Current models remain plagued with nagging issues like “AI hallucinations” (generating wildly incorrect answers) which makes their widespread use difficult in real life situations.
While new models and techniques have been lowering hallucination rates, some researchers believe it may be impossible to eliminate them completely, for the simple reason that hallucinations are part of the fabric of these models. This point represents perhaps the key difference between generative AI and traditional software: most of today’s software is deterministic in nature, in the sense that the same inputs will reliably produce the same outputs. The transformer architecture introduced by “Attention Is All You Need” is fundamentally a probabilistic system, one that assigns probabilities to bits of data and samples them to generate outputs. With this process comes a great deal of uncertainty and variability, but this is also what allows AI models to be so creative and answer even the most elaborate or obscure questions, in ways that at times feel so eerily human.
But no matter how big AI gets, there is still probably no bigger probabilistic system than global financial markets. The stock market is, fundamentally , a giant worldwide network of people, computers and algorithms, all working to assign probabilities to future cash flows. Despite the endless amount of resources poured into this exercise over decades, markets remain largely unpredictable and have resisted all attempts to attribute deterministic properties to their behavior. While most issues with AI hallucinations can be addressed by establishing guardrails and double checking outputs, market investors are playing with live bullets, and by the time we realize our mistakes, the damage is already done.
As we enter 2025, the market finds itself at an interesting juncture, marked by the confluence of frenzied stock market enthusiasm, all time high prices, stratospheric valuations, coupled with rapid disinflation, a slowing labor market, and declining federal interest rates. In the midst of all this, so much of our future hopes are tied to an immensely promising technology, which will almost certainly transform our lives, but which has so far engulfed vast amounts of resources for relatively little palpable benefit to the economy as a whole.
As the year unfolds, prudent and value-conscious investors may ultimately be glad not to have abandoned long established principles of diversification, risk mitigation, and asset allocation. For the time being, however, prices speak for themselves, and momentum is well truly all you need.
Key Takeaways
- Simplification Drives Innovation
The 2017 paper, “Attention Is All You Need” revolutionized AI by simplifying neural networks, enabling breakthroughs like ChatGPT.
- Infrastructure Fuels AI Growth
Massive investments in data centers and computing power are transforming economies and enabling applications like AlphaFold, which solved a decades-old scientific challenge. - Momentum Dominates Markets
Tech stocks surged in 2024, driven by momentum investing, but valuations and fundamentals took a back seat, raising concerns about sustainability.
- AI Progress Slows
Diminishing returns on AI advancements and slower adoption suggest its real economic impact is still limited, despite high expectations.
- Caution Amid Optimism
AI’s potential is vast, risks like overvaluation and economic imbalance highlight the need for prudent, diversified strategies.
paragraph 1 text content
