Special Issue: Investing in AI From Allocation to Application

Overview

AI refers to the ability of machines to perform tasks that typically require human intelligence, such as learning, reasoning, problem-solving, and understanding language. The economic implications of AI are wide-ranging. On the productivity front, AI has the potential to unlock gains by automating routine tasks and augmenting human capabilities, particularly in sectors such as healthcare, finance, and customer service. This could contribute to GDP growth, lower costs, and reshape labor markets by raising opportunities and challenges around job displacement and skills transitions.

From a market perspective, AI has already made a profound impact on capital allocation. Investment is flowing into AI infrastructure, software, and companies applying AI at scale. This is driving sectoral shifts, with tech, data-rich industries, and AI-enablers poised to benefit. Yet progress remains uneven: Only about 5% of AI pilots are delivering meaningful returns, leaving many companies and industries stalled on adoption. Furthermore, AI also introduces new risks, such as regulatory uncertainty, ethical concerns, and potential disruption to business models.

The labor market remains resilient, supporting the Fed’s preference to hold off on rate cuts. Over the past three months, the U.S. economy added 449,000 new jobs, and the unemployment rate remains near historic lows at 4.1%. Wage growth remains above pre-pandemic levels, increasing by 3.7% year-over-year in June. The most recent JOLTS (Job Openings and Labor Turnover Survey) showed that vacancies in May rose to the highest level since November 2024 and well above pre-pandemic levels.

Background and AI Models

The idea of AI has been around since the mid-1950s. In 1956, Arthur Turing proposed a test to determine if a machine could exhibit human-like intelligence. The term “artificial intelligence” was coined at Dartmouth College in 1956 by computer scientist John McCarthy. AI’s evolution has been rapid, from early chatbots like ELIZA to today’s multimodal large language models (LLMs).

There are 10 core components of AI:

1. Machine Learning: Enables AI systems to learn patterns from data and improve performance over time without being explicitly programmed. An example is Netflix recommending shows based on your viewing history.
2. Robotics: Combines AI with mechanical components to create systems that can act and interact in the physical world, such as self-driving cars or warehouse robots.
3. Deep Learning: A subset of machine learning that uses multi-layered neural networks to model complex patterns and decisions. For example, Google Photos can automatically recognize people and objects in images.
4. Expert Systems: AI programs that mimic human expert decision-making using a set of rules and logic, such as medical diagnosis tools that assist doctors in identifying diseases.
5. Reinforcement Learning: A type of machine learning where an agent learns by interacting with an environment and receiving rewards or penalties. An example of this is AlphaGo, the AI that learned to play and master the game of Go.
6. Neural Networks: Algorithms inspired by the structure of the human brain, designed to process and interpret data through interconnected layers, such as speech-to-text systems like those used in voice assistants.
7. Natural Language Processing (NLP): Allows machines to understand, interpret, and generate human language. ChatGPT and translation tools like Google Translate are good examples of NLP.
8. Computer Vision: Enables AI to interpret and analyze visual data from the world, such as images and videos that are used in the facial recognition systems in smartphones and at airports.
9. Speech Recognition: Converts spoken language into text so that machines can understand and respond to voice input. Examples include Siri, Google Assistant, and automated customer service systems.
10. Planning and Optimization: Allows AI to set goals and determine the best way to achieve them under constraints, such as route optimization in apps like Google Maps and Uber.

LLMs are built using deep learning and trained on massive datasets to predict and generate text, answer questions, summarize information, and more. Models vary widely in size, from billions to hundreds of billions of parameters (the internal values LLMs adjust to learn patterns in data), impacting their power and computing demands. Below is a summary table of the top 10 largest and most well-known LLMs.
AI hyperscalers are the largest cloud and infrastructure providers with the scale, computing capacity, and capital to build, train, and deploy these cutting-edge AI models. Amazon Web Services (AWS), Alphabet’s Google, Microsoft’s Azure, and Meta are the largest U.S. AI hyperscalers. AWS is the largest hyperscaler with a 19% market share. Overall, AI hyperscaler capital spending increased by nearly 60% between 2023 and 2024 and is estimated to rise to $340 billion by the end of 2025.

Economic Implications

Labor and Productivity: AI can significantly boost output per hour, with generative AI already reducing time spent on tasks like writing, analysis, and programming by 60%–80%. However, job displacement pressures are visible in service operations and entry-level roles, while demand is rising for AI-specific skills. The net effect will likely be labor reallocation rather than wholesale elimination. In the words of Nvidia CEO Jensen Huang, “You’re not going to lose your job to AI, but you’re going to lose your job to someone who uses AI.”

AI-specific jobs have surged to nearly 3% of total U.S. job postings. According to a McKinsey survey, generative AI is expected to see notable increases in technology and product and service development jobs, while the service operations workforce size is expected to shrink. Of the survey’s respondents, 46%believe that more than 20% of the total workforce will need to be reskilled within the next three years.

According to the AI Labor Index (see https://www.ailaborindex.com/), 45% of tasks have the potential for complete automation, but no job is entirely automatable, although customer service representatives have the highest automation potential of all jobs. The table below is a list of the most to least automatable occupations. Note that “Automation Potential” is a score indicating how automatable the nature of the job is. This ranges from 1 (automation is infeasible and humans are essential) to 5 (job is fully or nearly fully automatable). “Adoption” refers to the current real-world usage of AI tools per job field. This is based on data from the Anthropic Economic Index. Lastly, “Economic Impact” represents the potential value that could be unlocked through automation. This value is calculated based on total wages and the percentage of tasks with high automation potential.

Energy and Power Demand: Data centers and AI workloads are straining power grids. U.S. demand is set to double by 2030, and given their need for reliable baseload, AI data centers are leaning on natural gas and nuclear. This elevates utilities, nuclear, and grid modernization as indirect beneficiaries of AI.

The EIA’s latest Short-Term Energy Outlook projects U.S. electricity consumption to grow by 1.7% annually through 2026, reversing nearly a decade of flat demand. Longer term, a study from the National Electrical Manufacturers Association (NEMA) forecasts a 50% increase in consumption over the next 25 years, driven largely by data center expansion and the rise of e-mobility. With data centers requiring constant, reliable output, natural gas and nuclear remain the most viable sources to meet this demand. Reflecting this need, the White House issued four executive orders on May 23, 2025, to overhaul and accelerate the nuclear sector with the aim to quadruple capacity by 2050, streamline regulations, and strengthen supply chains. Together, these steps position nuclear power at the center of U.S. energy, economic, and security policies. According to leading AI companies, power constraints are the key to future AI development. Sam Altman, the CEO of OpenAI (who created ChatGPT) in July noted, “The limit won’t be the algorithms and the research, but it’ll increasingly become the physical instantiation that it takes to make this work. Chips, cables, servers, energy, everything that you need to power this brain. And the more of it, the better.”

In their second-quarter earnings call, Amazon CEO Andy Jassy said, “You see some of the constraints and they exist in multiple places…the single biggest constraint is power.”

Market Implications

Investors are seeking to determine how best to position themselves for the growth of AI, not only through established leaders such as Nvidia and Microsoft, but across the broader AI investment ecosystem. Investment opportunities can be arranged into six layers: energy, chips, data infrastructure, models, software frameworks, and applications. Each layer has distinct risk/return. For example, energy and chips offer scale and visibility, whereas applications offer growth but face adoption hurdles.

Below are some details of the AI investment stack:

1. Energy & Infrastructure: Includes power generation, transmission and distribution companies, such as utilities companies, uranium and nuclear companies, and oil production and exploration companies.
2. Chips: This part of the AI investment stack includes high-performance semiconductors, such as GPUs and ASICs.
3. Data Infrastructure: Includes the physical and cloud-based architectures for data storage, networking, and processing. Data center REITs and digital infrastructure companies fall into this part of the AI investment stack.
4. AI Model Companies: This part of the AI investment stack includes organizations developing and commercializing foundational AI models, including LLMs, such as ChatGPT. Companies in this part of the AI investment stack include both public companies (such as Meta and Alphabet) and private companies, including OpenAI, Anthropic, and xAI. Notably, the most valuable AI-native companies are privately held.
5. Software Infrastructure: Includes middleware and development frameworks that enable AI integration. Snowflake and Gitlab are examples of these types of companies.
6. Apps & Services: Represents the applied layer where AI capabilities are embedded into products, platforms, and workflows. Palantir, C3.ai, and SentinelOne are some of the companies that fit into this part of the AI investment stack.

Hyperscaler capex is running at record highs, with Microsoft, Google, and Meta signaling further investment in 2026. For 2025, Meta has raised 2025 capex spending from $65 billion to $72 billion, while Alphabet increased its planned capex spending from $75 billion to $85 billion for the year. S&P 500 earnings growth is increasingly concentrated in technology and communications, reflecting AI’s pull on capital allocation.
Private equity and venture capital-backed investments in AI technology have increased, indicating strong interest and opportunity from within private markets to capitalize on AI’s growth and innovation potential. Venture and private equity continue to pour into AI, with OpenAI, Anthropic, and xAI commanding valuations north of $70 billion to $300 billion. Meanwhile, smaller players are innovating in vertical applications and developer tooling. Globally, private AI investment increased 45% between 2023 and 2024. Despite recent fluctuations, private AI investment globally has grown substantially over the past decade.

Using AI

The following is a brief overview of practical ways AI can be leveraged today in business and the risks to using it from client and business perspectives.

Real-world AI deployments span from call centers to legal drafting and compliance. In healthcare, AI is speeding up drug discovery. For example, Moderna recently credited AI for cutting vaccine design time from months to weeks. In law, junior associates are using AI to summarize case law, freeing them to focus on strategy. In finance, asset managers are experimenting with AI-driven research, portfolio management, and compliance monitoring. In advisory work, AI is drafting newsletters, analyzing client portfolios, and improving reporting efficiency. For firms, the best approach is to use AI as a force multiplier—augmenting human advisors rather than replacing them. The opportunity is scale, but the risk is over-reliance as compliance, accuracy, and bias must be actively managed.

Below is a table of the use cases, benefits, and risks of implementing AI into business practices:

Case Studies

A recent MIT study found that industry-level generative AI transformation remains limited as only tech, media and telecommunication, and professional services industries reported signs of structural disruption.

Further, only 5% of custom enterprise AI tools reach production. The MIT study notes that chatbots tend to gain traction because they’re simple to deploy and broadly adaptable, but they often fall short in mission-critical workflows where memory and customization matter. Enthusiasm and budgets are usually enough to launch pilots but turning them into workflow-integrated systems that deliver lasting value is far less common. Key takeaway: adoption is high, but transformation is rare, with poor enterprise-level AI tools a key reason for slow adoption.

The MIT study also found five myths about generative AI in businesses:

1. AI will replace most jobs in the next few years. Research found limited layoffs from generative AI, and only in industries that are already significantly impacted by AI. There is no consensus among executives as to hiring levels over the next 3–5 years.
2. Generative AI is transforming business. Adoption is high, but transformation is rare. Only 5% of enterprises have AI tools integrated in workflows at scale and 7 of 9 sectors show no real structural change.
3. Enterprises are slow in adopting new tech. Enterprises are extremely eager to adopt AI and 90% have seriously explored buying an AI solution.
4. The biggest thing holding back AI is model quality, legal, data, and risk. What’s really holding it back is that most AI tools don’t learn and don’t integrate well into workflows.
5. The best enterprises are building their own tools. Internal builds fail twice as often.

An Apple study found that as task complexity increases, LLMs initially allocate more tokens (or thinking effort) but then reduce reasoning effort, which leads to a collapse in accuracy. The study found that for simple tasks, LLMs often get the right answer quickly but keep exploring wrong options, wasting tokens. Even when given the right step-by-step method, they often don’t follow it correctly, showing weak problem-solving with symbols and rules. Key takeaway: AI is not perfect; it cannot replicate human experience and often fails at efficiently solving complex yet logical tasks.

When using AI in business, always consider both the abilities and limitations of different LLMs. For example, LLMs (ChatGPT and Gemini, for example) are unable to play chess. However, IBM’s DeepBlue AI model defeated World Chess Champion Garry Kasparov in 1997 because that was its purpose! LLMs (like ChatGPT) are trained to predict the next word in text, not to evaluate millions of board positions or calculate optimal moves. LLMs may produce plausible chess moves in conversation, but they lack the structured reasoning and rules enforcement that chess engines have. Key takeaway: Use the right tool for the right task.

The reliability of LLM responses often comes into question. Reddit and Wikipedia are the top-cited sources for several major LLMs, including Google’s AI Overview, ChatGPT, and Perplexity. A McKinsey study found that people are just as likely to fully review AI outputs as they are to barely check them. The study found that 30% of respondents said that they reviewed a maximum of up to 20% of all AI outputs, and only 27% of survey respondents said that they reviewed 100% of AI outputs. Furthermore, generative AI models can produce outputs that are plausible but incorrect, and these are known as ‘hallucinations.’ Common causes of hallucinations include insufficient data, faulty architecture, overfitting, or the method of generation for the AI model. Key takeaway: Cross-check facts as AI can fabricate or miss nuance, so use it as a draft and not a source of record.