.png){kind=logo}
Calling something an "AI investment" doesn't tell investors very much anymore. The label can apply to semiconductor designers, cloud platforms, data center developers, foundation model companies, enterprise software vendors, and consumer applications. Those businesses may all benefit from the growth of artificial intelligence, but they don't have much else in common when it comes to capital requirements, competition, or pricing power.
As Neil Druker, Managing Member of Melanie Lane Capital Management in Boston, explains, AI is better understood as a series of interconnected layers rather than a single market. Each layer has different economics, and the fact that a business is essential to the technology doesn't necessarily mean it will capture the most value from its growth.
That's an important distinction for investors.
The more useful questions are: where scarcity exists, how long it is likely to last, and whether the companies benefiting from it can turn that advantage into durable returns.
Advanced computing sits at the foundation of today's AI systems, and the most visible beneficiaries have been companies supplying accelerators and the components that support them. But the hardware layer extends well beyond the processors themselves.
Memory, high-speed networking, advanced packaging, manufacturing equipment, and power management all play a role. A shortage in any one of those areas can become the constraint that limits the rest of the system.
The investment challenge is separating a temporary shortage from a durable competitive advantage. Scarcity can support strong pricing and margins, but the same economics can attract new capital, encourage competitors, and push customers to seek alternatives.
For Druker, near-term demand is only part of the picture. Investors also need to ask how durable a company's technical lead is, how deeply its software ecosystem is embedded, what switching costs customers face, and how much of the broader supply chain it actually controls.
A company can be selling something the market desperately needs today without having the same bargaining power several years from now.
Cloud providers play an important role in the AI stack by providing computing capacity, financing, software, and customer access. They also take on much of the capital burden required to build the infrastructure that supports AI workloads.
That can look like an attractive place for economic value to settle, but the economics aren't automatic.
AI workloads require significant capital, and the cost of building that capacity must be considered alongside utilization, energy costs, depreciation, and the bargaining power of very large customers.
Revenue growth alone doesn't answer the question.
Investors need to know whether the infrastructure is being used efficiently and whether providers can earn returns above their cost of capital after accounting for the full cost of building and operating it. Rapid growth can produce an impressive revenue number while creating less economic value than the headline suggests if utilization or pricing doesn't keep pace.
Foundation-model companies face a different challenge. They can create enormous technological value without necessarily creating an equally durable business.
Training costs matter. So do inference costs, proprietary data, distribution, developer adoption, and customer retention. And those factors don't always move together.
One of the biggest questions is whether model performance will continue to provide meaningful differentiation. As models become more capable and performance differences narrow, competitive advantages may shift toward distribution, workflow integration, security, and specialized data.
That means technical leadership and business durability have to be evaluated separately.
The strongest model today may not belong to the company with the strongest customer relationships tomorrow. In many markets, those relationships are what ultimately determine pricing power.
At the application layer, AI companies are trying to become part of the work customers already do. The value can come from reducing labor costs, increasing productivity, improving output, or making services that weren't economical to deliver.
The strongest businesses tend to have something beyond the underlying model. They may control distribution, understand a specialized workflow unusually well, or have proprietary data that produces a measurable advantage.
The risk is that the application itself becomes easier to replicate as model capabilities improve and development costs fall.
Three questions can help distinguish the stronger businesses from the rest.
Does the product become essential to a customer's workflow, or is it simply useful? Can customers clearly measure the benefit they're getting from it? And can the company maintain its pricing as similar model capabilities become more widely available?
A company that can answer yes to all three has a very different economic position from one that depends mostly on being early.
AI is often discussed as a software story, but scaling it depends on a surprisingly physical set of resources.
Electricity generation, transmission capacity, cooling, land, permitting, and construction capacity can all limit how quickly new data centers come online. Those constraints create opportunities for businesses that can help relieve them.
They also create plenty of reasons for investors to be cautious.
Infrastructure projects can take years to develop and require substantial financing. According to Neil Druker, regulation, permitting delays, local opposition, construction costs, and contract terms can all affect the eventual economics.
That's why contract terms and capital structure deserve as much attention as the headline demand story. A real bottleneck can still produce a poor investment if the company needs too much capital to address it or accepts terms that leave too little of the resulting value with shareholders.
The companies with the most interesting technology aren't necessarily the ones that capture the most economic value.
Bargaining power can come from controlling a scarce resource, owning the customer relationship, creating meaningful switching costs, or growing without incurring disproportionate capital investments.
The problem is that bargaining power changes.
Customers willing to pay almost anything for scarce capacity today may negotiate more aggressively once supply expands. Cloud providers may support model developers when doing so drives infrastructure demand and pull back when the economics change. Application companies may switch underlying models as performance converges.
That makes who owns the customer relationship an ongoing question rather than something investors can answer once and leave alone.
The economic balance between the layers can shift much faster than the technology itself.
Investors don't need a perfect forecast of how large AI will become to evaluate these businesses. A consistent framework can be more useful.
What scarce resource or capability does the company control? How long is that scarcity likely to last, and what could weaken it? How much capital does the company need to maintain its position? Who owns the customer relationship and, with it, the pricing decision? And how much of that future success is already reflected in the valuation?
That last question can be the easiest to overlook during a period of enthusiasm.
A company can have a scarce resource, strong customer relationships, reasonable capital requirements, and attractive competitive dynamics and still be a poor investment if the stock price already assumes everything will go right.
Artificial intelligence is likely to create significant economic value. That doesn't mean the value will be distributed evenly across the companies building and supporting it.
That's where the layered view becomes useful.
Instead of treating all AI-related growth as equally durable, investors can look at where technology, scarcity, market structure, and capital discipline reinforce one another. Those are the areas where economic value is more likely to persist.
For institutional investors, this offers a more practical way to think about the AI infrastructure opportunity. It's not one broad bet on adoption. It's a collection of businesses with different economics, different risks, and very different paths to turning AI growth into lasting shareholder value.
This article is educational and analytical in nature. It does not constitute investment, legal, or tax advice, does not recommend any security or transaction, and is not an offer or solicitation.
