AI Investing Framework: Six Layers, Where Value Accumulates

Most investors chasing AI right now are asking the same question: what is the next NVIDIA?

I understand why. When a major innovation wave arrives, everyone wants to find the one company that captures all of it. The stock market loves a clean story, and AI has become one of the biggest stories in decades. But that question is almost certainly the wrong one. Investors who chase it are setting themselves up for disappointment, even if the technology itself turns out to be real.

The better question is: where does value accumulate, and who has the moat?

Those are different questions. They lead to a completely different way of analyzing the opportunity.

The internet changed everything. It also created infrastructure winners, software winners, marketplace winners, advertising winners, payment winners, and a long list of companies that disappeared. Smartphones did the same thing. Cloud computing did the same thing. The theme was real in every case, but not every company attached to the theme became a great investment.

AI is likely to follow the same pattern.

Some parts of the AI ecosystem will become incredibly valuable. Some will become commoditized. Some will grow quickly but struggle with margins. Some require enormous capital investment and produce only regulated or modest returns. Some will look exciting today and be forgotten in five years.

That is why I think in layers. When I look at AI investing, I break it into six parts:

→ Utility Infrastructure,
→ Chips and Compute,
→ Cloud Providers,
→ AI Models,
→ software applications, and
→ existing businesses that use AI to improve productivity.

Each layer has different economics, different moats, and different long-term potential.

In July 2025, Blackstone announced it would invest more than $25 billion in Pennsylvania digital and energy infrastructure and help catalyze another $60 billion in additional investment. That is the type of announcement that shows AI is not just a software story. It is a power, infrastructure, financing, and real-assets story.

AI requires enormous amounts of power. Data centres need electricity, transmission, substations, transformers, cooling, and backup systems. That part is not in doubt.

The part I am skeptical about is the assumption that utilities automatically become great long-term investments because of it.

Utilities are regulated, capital-intensive, and slow-moving by design. A utility can benefit from higher demand, but the returns are often negotiated through regulators and shared with customers over time. Essential does not always mean high-return.

Green energy was a useful lesson here. The theme was real. The world needed more renewable energy. Capital flooded into the sector. But many green energy stocks did not produce the investor returns people expected. A real-world need does not automatically translate into strong shareholder returns.

The same caution applies to AI power demand. Utilities will build, but they will not build on pure speculation. A utility needs a capital recovery path before committing to a major substation, transmission upgrade, or new generation resource. That is where large-load tariffs matter: they are designed to make large customers commit for long periods, pay minimum charges, provide collateral, and absorb part of the infrastructure risk.

The timelines also matter. Even when projects are approved, power infrastructure takes time. Approximate timelines to capture effort and investment impact.

Slow, regulated, capital-intensive, and shared with customers, regulators, and governments. That is very different from owning a software company with high margins and global scalability.

Where I prefer to look in this layer is companies that finance, own, and operate physical infrastructure rather than traditional utilities themselves. Brookfield Corporation is the name that comes to mind. Not because it is a pure AI stock, but because it understands long-duration infrastructure, renewable power, private capital, and large-scale financing. AI may become another demand driver for assets Brookfield already knows how to manage.

I already own and invest in a number of companies that participate in the AI transformation. Do not take any mention as investment recommendations. Please do your own due diligence.

The chip layer is where barriers to entry are genuinely high. Only a small number of companies can compete at the leading edge, and the supply chain is extremely concentrated. That creates pricing power when demand exceeds supply.

But the AI chip story is more complex than it looks. A GPU is not alone. It needs high-bandwidth memory. It needs advanced packaging to connect the processor and memory efficiently. It needs specialized servers. It needs networking to connect clusters. Every bottleneck in that chain can become an investment opportunity.

I want to be honest about my own track record here. NVIDIA would have made me far more money than my actual position. I bought a small position because I knew part of my motivation was FOMO. That is a real investor lesson. Sometimes you see the right trend and do not size the position aggressively enough. Sometimes that caution protects you, and sometimes it costs you upside.

Broadcom was different. I bought it because networking demand was going to keep growing regardless of the exact shape of AI. More data, more cloud usage, more connectivity, more infrastructure. AI did not create that thesis; it accelerated it. That is the kind of investment I prefer. A company that benefits from a major theme without requiring me to predict every detail of how it unfolds.

Valuation discipline still matters here. Semiconductors can be cyclical. Demand looks endless during a boom, but supply eventually catches up, customers optimize spending, and competition increases. Position sizing matters more when expectations are already elevated.

Cloud providers are not simply renting servers. They are becoming the infrastructure that software companies are built on top of. Once a business builds its workflows into AWS, Azure, or Google Cloud, switching becomes extremely difficult. Applications, databases, security models, identity systems, compliance processes, data pipelines, and internal developer habits are all tied together. A company does not usually walk away from all of that just because another provider is slightly cheaper.

That stickiness creates durable revenue. It also gives cloud providers multiple paths to monetize AI. They can rent compute, host models, sell AI tools, provide databases, offer analytics, and embed AI into developer workflows. They can also build their own custom chips over time to reduce dependence on external suppliers.

The other advantage cloud providers have is the ability to spend. AI infrastructure is expensive. GPUs are expensive. Data centres are expensive. Power is expensive. Networking is expensive. Talent is expensive. Only a small number of organizations can fund that capital spending year after year.

Microsoft, Amazon, Alphabet, and Oracle are not pure AI plays, and that is part of the appeal. They already have customers, cash flow, distribution, and infrastructure built over decades. AI becomes another layer of monetization rather than the entire investment thesis. If a startup builds an AI product and succeeds, it may still run on cloud infrastructure. If an enterprise builds its own AI workflow, it may still use cloud tools. If a model company grows, it still needs compute. The cloud layer wins across multiple outcomes.

The model layer gets the most attention because it is what people interact with directly. ChatGPT, Claude, Gemini, and others feel like the product. But visibility does not always equal durable profitability.

Training models requires enormous compute. Inference costs money every time a user interacts. Talent is expensive. Safety and compliance are expensive. The product improves quickly, but the cost of staying at the front is very high.

Model performance can also shift fast. A model that feels far ahead today may be caught several months later. If users can switch easily, the leader has to keep earning its position every cycle. That makes the moat harder to assess.

There are moats in this layer, but they are not always obvious. Performance matters. Cost matters. Enterprise trust matters. Developer adoption matters. Proprietary data matters. Distribution matters. Distribution may matter most of all, because a powerful model still needs to reach users in a natural way.

That is why partnerships are critical. OpenAI has Microsoft. Anthropic has deep relationships with Amazon and Google. Gemini is embedded inside Google's ecosystem. Meta has open-source models and massive consumer distribution. Apple has the device layer. The model alone is not enough.

For most public market investors, the problem with this layer is that many pure-play model companies are private. Most investors get exposure indirectly through Microsoft, Amazon, Alphabet, or Meta. That may actually be the cleaner approach. Instead of betting on which model wins, you own companies that benefit from AI adoption across multiple paths, regardless of which model ends up on top.

AI lowers the cost of building software, which means more products get launched by more people in less time. That sounds like opportunity. It also means competition can explode.

This is where I am most cautious. There will be application winners. But many AI applications may end up looking more like features than standalone companies.

Consumers do not want twenty-five AI subscriptions. They want useful features inside the tools they already use. Microsoft can embed AI into Office. Google can embed it into Search, Gmail, and Docs. Adobe can embed it into creative workflows. Salesforce can embed it into CRM. Established platforms have a structural advantage because the users are already there and the switching costs are already in place.

The pricing economics are also different from traditional software. Traditional SaaS often charged per user with near-zero incremental costs on usage. AI software has usage-based costs because every model call requires compute. Tokens work more like kilowatts: the more you use, the more it costs someone. Companies that do not design pricing with that in mind will find margins compressed as usage grows.

The application layer winners will survive because they solve genuinely expensive problems, integrate deeply into workflows, own proprietary data, or create real switching costs. The weaker products will be copied, bundled, or ignored.

China did this to manufacturing. Shenzhen turned product development into a turnkey process where almost anyone with an idea and a modest budget could get a physical product designed, prototyped, and manufactured at scale. The barrier to entry for making things collapsed. What followed was not a small number of dominant manufacturers. It was an avalanche of competitors, price compression, and razor-thin margins for anyone without a real distribution advantage.

AI is doing the same thing to software. The cost to build a functional application is falling fast. Research and development no longer requires years or large teams. That is not just an opportunity. For most application-layer companies, it is the threat.

Not every AI winner needs to be an AI company.

Some of the best long-term investments may be existing businesses that use AI to improve productivity, margins, customer service, logistics, software development, fraud detection, and decision-making. This kind of benefit is harder to see because it does not show up as an AI revenue line. It shows up as operating leverage. A company may grow revenue without growing headcount at the same pace. It may reduce support costs. It may improve inventory management. It may automate back-office work. Only time will tell.

A bank, insurer, railway, retailer, or payment network that uses AI to improve returns on capital is not an AI investment in any obvious sense. The business may not be described as an AI story. But AI can still support its long-term compounding.

Costco using AI better does not make Costco an AI company. But if AI helps with inventory, logistics, pricing, forecasting, or fraud prevention, it supports the business. The same idea applies to many companies that already have strong moats.

This layer interests me because it is the least speculative. You are not betting only on the technology. You are betting on a business that already works and may become more efficient because of it.

AI is not a reason to abandon diversification. It is a reason to understand what you already own.

If you hold a broad S&P 500 ETF, a Nasdaq 100 ETF, or a global equity ETF, you probably already own Microsoft, Apple, NVIDIA, Amazon, Alphabet, Meta, and Broadcom. You may have more AI exposure than you think. For Canadian investors especially, this matters. Many Canadian portfolios already hold U.S. index ETFs with heavy technology weighting. Before buying another AI stock, it is worth checking what your existing holdings already give you.

If you have started index investing, the AI exposure is already there through market-cap weighting. You do not necessarily need a separate AI portfolio on top of it.

In my Layer Cake Portfolio Strategy, AI does not belong in the Foundation layer. It is too uncertain, too valuation-sensitive, and too concentrated for that role. The Foundation layer should remain broad, diversified, and stable enough to hold through cycles. AI fits better in the Engine, Compounders, and Accelerators layers. Large technology platforms fit in Compounders when business quality and valuation make sense. Semiconductor leaders and infrastructure suppliers fit in the Engine layer. Smaller or more speculative AI names belong in the Accelerators layer, if at all.

The key is sizing. A good theme can still become a bad portfolio decision if it becomes too large. I do not want my portfolio to depend on predicting the exact AI winner. I want exposure to the trend, but in a way that respects uncertainty.

When I look at a company in the AI space, I ask a few basic questions.

→ Where does the company sit in the AI stack?
→ Does it control a bottleneck?
→ Does it have pricing power?
→ Can customers switch easily?
→ Is it improving margins, or just spending more capital?
→ Does it have durable cash flow outside AI?
→ Is the valuation already assuming perfection?

That last question matters most. A lot of investors focus only on the story, and the story may be true. But the stock can still disappoint investors if the price already assumes too much success. This is one of the biggest investing mistakes that quietly kills long-term wealth.

It happened during the internet bubble. The internet changed everything, but many internet stocks failed. The theme was real, but the market overpaid for many of the early stories. Not every early winner survived.

Separate the theme from the investment. A theme can be obvious and still be difficult to invest in. A company can be important and still be too expensive. A product can be useful and still lack a moat.

I do not think AI is hype. I believe it is real. The infrastructure layer is real, but capital-intensive and slow. The chip layer is powerful, but valuation-sensitive. The cloud layer looks durable because it controls infrastructure, deployment, and ecosystems. The model layer is exciting, but expensive and uncertain. The application layer is innovative, but likely crowded. And the most overlooked winners may be existing businesses that quietly become more efficient because of the technology.

That is the full picture. Before chasing the next AI headline, look at your ETFs. Look at your technology exposure. Look at your concentration. You may already own more AI than you realize.

Smarter investing does not start with the headline.

It starts with the portfolio.

Remember why you invest, it’s to enjoy life and everything it has to offer. Don’t forget to treat yourself once in a while.

Disclaimer: This product contains affiliate links. If you click and make a purchase, I may earn a small commission at no extra cost to you.