Investment Insight: Decoding A.I. and Generating Investment Idea
[social-media-table2]
Artificial Intelligence (A.I.) refers to systems or machines designed to simulate human intelligence, enabling tasks like learning, reasoning, problem-solving and decision-making to automate processes, enhance efficiency and unlock data-driven insights.
As we were pondering what to write for this paper’s lede, we decided it would be better if we didn’t write one at all. Instead, we punted the prompt to Claude, a large language model developed by Anthropic. Our query would be answered by Claude 3.5, a sixth-generation A.I. model that balances speed and performance.
Our request to the chatbot, which was answered in less than 10 seconds, was: “Explain what A.I. is to my 81-year-old father in one paragraph.” Claude’s response was clear, reasonably accurate and understandable (though not exactly succinct):
Think of artificial intelligence, or A.I., as being like a very sophisticated calculator that can learn from examples, similar to how you learned things throughout your life. Just as you learned to recognize cats by seeing many cats over the years, A.I. systems learn by being shown millions of examples by humans. These systems can then help with tasks like having conversations through text, helping doctors spot things in X-rays, or making it easier to find specific photos on your phone. It’s not magic–it’s just another step in the evolution of technology–like how we went from rotary phones to smartphones, and like any tool, it can make mistakes and needs human guidance to work properly.
If A.I. is simply an advanced calculator, what is all the fuss about?
First, What is A.I.?
A.I. is an abbreviation for artificial intelligence, which is a field of computer-oriented research and development that started to gain traction in the 1980’s and 1990’s. The goal was to arrive at conclusions when there is incomplete information. As humans, we can pull in all sorts of information from our life experiences and use them to arrive at an answer.
For example, if we asked someone for a good place to eat, a human answerer of that question would make certain assumptions, including location, time of day and quality minimums. With a lifetime of experiences, it’s a question that’s easily answerable, even for a stranger.
A simple Google or DuckDuckGo search would easily answer such a question, too, but the inputted text string would need to be fairly precise to end up with a quality result. Such a string might be, “What is a good sushi restaurant for lunch in Santa Barbara, California?”
A.I. attempts to make things easier on the user. “Where should I go to lunch today?” is all it might require. It endeavors to accurately fill in many blanks that a human would not consciously consider, then use that human-like knowledge to answer the question. The thing is, A.I. needs to have built a dataset of knowledge first before it’s ready to be deployed to the world.
As a category, this process is called Machine Learning (ML). The premise is that computers can be setup to make decisions without being explicitly programmed for each step. With flexible rules, algorithms use feedback loops to process data, refine models and improve performance.
To get ML started, a base dataset is fed to algorithms, which attempt to uncover patterns. After some augmenting to the algorithm, the training data set is swapped out and replaced with new, previously unseen data. This step ensures that the initial learning was robust, and the detected patterns weren’t anomalies. The process is rinsed and repeated until the model reaches a rate of accuracy sufficient for deployment.
There are three main types of ML. Supervised Learning is based on labeled data, meaning one could use an algorithm to predict a home price based on size and location. Unsupervised Learning means the algorithm is finding hidden patterns in the data, such as bucketizing customers into categories the model determines. Reinforcement Learning has been the most difficult to implement, which is learning by trial-and-error that’s guided by rewards and penalties.
The DeepSeek News
The goal of artificial intelligence (A.I.) is to arrive at conclusions when there is incomplete information.
In late January, DeepSeek, a Chinese A.I. laboratory, presented a new A.I. reasoning model called R1 that is fast, reasonably accurate and uses novel logic shortcuts that purportedly result in a 90% to 95% reduction in computing power consumption. R1 uses reinforcement learning to generate chains of complex thought autonomously and self-verifies the information, resulting in a scalable, ultra-efficient model like none before it.
The release on GitHub, a software development collaboration platform, rocked stocks in the A.I. space, especially after software engineer and early-stage investor Marc Andreessen called attention to it in a tweet saying DeepSeek is “a profound gift to the world.” In subsequent trading, investors quickly dumped A.I. stocks, fearing that massive capital investments could be worth a whole lot less if model costs drop 90% or more.
We thought knee-jerk dumping of many A.I.-related stocks missed the mark and wrote about it in the 700th issue of The Prudent Speculator on February 4th. Despite the quick downward snap, stocks in the Information Technology sector have bounced off their January lows and continue to maintain sizable upside potential.
Is it Not a Big Deal?
We want to emphasize that we are not discounting the DeepSeek news. In fact, we believe it’s a major step forward and appreciate that its open-source access creates a far bigger splash than it would if it lived inside the walls of a single company. In the same breath, the billions of dollars being spent on A.I. investments, especially those related to hardware, are still necessary and needed.
But how could one dramatically reduce computing power needs and still not exhaust demand? That’s because we, in the global sense, were never going to satisfy foreseeable demand for A.I. hardware and software in the first place. Manufacturing, energy and software A.I. demands far exceed available capacity. Supply is so scarce that some private equity firms have been using stockpiles of Nvidia chips as carrots to dangle in front of startups needing more than capital injections. The financial terms often end up being worse for founders. We suppose having less-than-ideal terms is better than not having the chips to make everything work.
Beyond physical constraints, there isn’t going to be one A.I. model that suits all use cases. We asked DeepSeek, “Why should I use different A.I. models?” It offered 8 bullet points: 1) Task-specific performance, 2) Efficiency vs. Accuracy Trade-offs, 3) Data Constraints, 4) Robustness and Generalization, 5) Ethical and Bias Mitigation, 6) Innovation and Experimentation, 7) Cost and Scalability and 8) Domain-Specific Needs.
The next suggestion was unexpected but helpful in proving our point that one model does not fit all use cases. It suggested mixing models for complex workflows such as recommendation systems (Collaborative filtering + Natural Language Processing for personalized suggestions) and self-driving cars (Convolutional Neural Networks for vision + Long Short-Term Memory for motion prediction + Reinforcement Learning for decision-making).
And then there’s a question about what should be considered a resource-intensive component of artificial intelligence and what is not. On Threads, Chief Artificial Intelligence. Scientist at Meta Platforms (META) Yann LeCun offered some insight, “Much of those billions are going into infrastructure for *inference*, not training…Once you put video understanding, reasoning, large-scale memory, and other capabilities in A.I. systems, inference costs are going to increase. The only real question is whether users will be willing to pay enough (directly or not) to justify the [expenses].”
Training is like studying for an exam and practicing questions. It involves enormous up-front costs. ChatGPT’s GPT-4 cost more than $100 million in the training phase (it told us). Inference is like taking the actual exam with new questions the test-taker hasn’t seen before. It appears much less expensive, costing about three tenths of a penny ($0.003) per request, but the number of hits can rise rapidly.
Upfront costs during the training phase are enormous with specialized applications like vehicles costing billions. Those costs may end up paling in comparison to the inference phase, which could be considered the equivalent of operating a factory (vs. ‘building’ it during the training phase). The good news is that per-query costs can generally be passed to users, so companies are not destined to give away any goods for free.
Thing is, the models have to be good enough for users and businesses to justify potentially high sums to pay for them. That means it’ll be important for the right models to be available at the right prices.
When Does the A.I. Music Stop?
We can’t be sure, but our thinking is that we are in the early days of A.I. computing, mainly because the technology is still mostly in the development phase and our expectations for A.I.-related benefits reach well beyond a small group of large companies.
Microsoft (MSFT) has rolled out Copilot, an A.I. assistant that is integrated into its Office 365 suite. Apple (AAPL) was somewhat late to the party, but it now includes Apple Intelligence in recently released devices like MacBooks and iPhones. Alphabet’s (GOOG) ads platform and Google search engine are using artificial intelligence to deliver more precise results, allowing searchers to spend more time on Google’s page, rather than clicking to a suggested link or navigating away. Meta Platforms has had great success deploying A.I. into its ad-targeting products, despite being kneecapped by Apple privacy protections a few years earlier.
The infrastructure buildout for A.I. has been massive and efficiency gains won’t curtail a need for hardware anytime soon. Cisco Systems (CSCO), a networking and digital security company, recently reached a share price not seen since the Tech Boom in 2000. Revenue and earnings growth began to accelerate on account of computing needs nearly a decade ago and have surged recently as the A.I. needs have triggered another network spending super cycle. The company also acquired Splunk, a software platform to analyze large quantities of data, which further adds to revenue growth potential.
After experiencing more than a decade of falling net income and market share losses, International Business Machines (IBM) has come roaring back lately. IBM was arguably one of the first A.I. companies with its Watson natural language Q&A platform (made famous by its appearance on Jeopardy in 2011), but the system was riddled with problems and generally had a disastrous rollout. Watsonx replaced the original platform in 2023 and is offered with three flavors of Services: A.I. (for large language models), Data (scaling workloads and analyzing data) and Governance (compliance and regulation). Analysts project 4% to 8% EPS growth in each of the next three years with similar rates of sustainable top-line growth.
Other purer-plays for A.I. include NetApp (NTAP), a data infrastructure firm that plays nicely with a wide array of major A.I. players. Oracle (ORCL) provides cloud services and database management for A.I. applications. Oracle’s success has been enhanced by a painful shift to a recurring revenue model a decade ago. Lumentum Holdings (LITE) manufactures optical, laser and photonics components that are in high demand as A.I. use accelerates. HP Enterprise (HPE), which is working on a purchase of Juniper Networks (JNPR), offers enterprise-scale A.I. solutions to support business growth for a diversified client base.
Who Else Benefits from A.I.?
There are less-direct beneficiaries of A.I. that may come as part of a second wave, meaning that the benefits of A.I. aren’t going to be apparent for a while longer. We like PayPal (PYPL) for its ability to improve customer experiences with A.I., streamline checkout and data entry processes, improve security and combat fraud. The payments platform is planning to widen the use of A.I. to make suggestions to shoppers and keep them coming back, helping expected earnings per share to grow by more than 10% per annum for the foreseeable future.
Financial Services titans like Citigroup (C), Goldman Sachs (GS) and Morgan Stanley (MS) have started pushing to modernize their infrastructure in order to revolutionize the way the companies function from infrastructure to customer interfaces to investments. Citi announced in January that it spent $11.8 billion in 2024 to deploy A.I. tools to 30,000 developers in an effort improve workflows for 143,000 employees. Morgan Stanley says A.I. will enable innovation and optimize processes. Goldman CEO David Solomon told investors in January, “We are leveraging A.I. solutions to scale and transform our engineering capabilities, simplify and modernize our technology stack and drive productivity.”
Digital Realty Trust (DLR) is a Real Estate Investment Trust (REIT) that specializes in data centers. Management says supply remains tight and “data center infrastructure remains a critical resource to support A.I. innovation.” We see opportunity in the data center real estate space for years to come.
Health Care is another sector that stands to benefit, though we thought we should ask A.I. (through Google’s Search Labs) how it can help with drug development. The list of opportunities it offered was long: target identification, drug design, virtual screening, ADMET (absorption, distribution, metabolism, excretion and toxicity) screening, trial design, drug repurposing, faster drug discovery, improved efficacy, reduced costs and drug personalization. It’s hard to tell which companies will see general improvements and which will have blockbusters thanks to A.I. because the drug pipelines still need to have robust outcomes. On the list of potential winners in this area are Amgen (AMGN), Bristol-Myers Squibb (BMY), Gilead Sciences (GILD), Merck (MRK) and Pfizer (PFE).
In the Industrials space, Eaton (ETN), a manufacturer of engineered products, can benefit from a jump in demand for the company’s electrical equipment. Engine maker, Cummins (CMI), is likely to see fuel economy and emissions improvements as A.I. is deployed (through its C3.ai software), making it more attractive to upgrade or use CMI powerplants in vehicles. In addition, CMI’s Power Systems segment could see new orders for power generation applications. Siemens (SIEGY), a global conglomerate, can help its enterprise and government customers in a broad set of A.I.-related areas of need, including energy consumption reductions, productivity enhancements, reliability upgrades and resource allocation optimization.
A.I. Benefits Are Far and Wide
Our thinking is that artificial intelligence will offer society (and our stocks!) benefits that substantially outweigh the downside.
In the preceding text, we offered a few main areas, along with more than a dozen A.I. stock pick ideas, that we believe will see significant benefits from A.I. technology. We think the ultimate outcome is that A.I. will reach into virtually every aspect of life. Perhaps it has already if you’re even reading this paper from a iPhone that has Apple Intelligence or if you found us through an A.I.-powered Google search.
Broadly speaking, our thinking is that artificial intelligence will offer society (and our stocks!) benefits that substantially outweigh the downside, though we aren’t closing our eyes to the idea that some things will not work out or are potentially scary.
For our broadly diversified baskets of Value stocks, we still need to see companies monetize the budding technology in a way that trickles down to earnings. As we alluded to in our coverage outside of the Tech sector, this might not be evident in a way that allows us to draw a straight line between A.I. and cash flow, but the gains should still be meaningful over time, even if near-term costs to get things going are large.
PARTNER WITH US
For more than 47 years, we have collaborated with our clients in their investment decision making process as they pursue their long-term financial goals. We are committed to keeping your goals, concerns and attitude about investing at the heart of your plan. If you’re ready to experience our personalized investment approach and exceptional client service, contact Jason R. Clark, CFA at 949.424.1013 or jclark@kovitz.com.
TPS Investment Insights - Decoding AI, February 2025