AI will change everything, but whether it pays investors is another story

Sponsored by

Who gets frontier AI’s cash burn surplus? AI will transform how we work; that’s the easy part. The tougher question: Who pockets the gains? Capital-intensive model builders may create enormous value and capture almost none of it. Inventing the future doesn’t mean you’ll get paid for it.

History is full of innovations that changed society while leaving shareholders empty-handed. One of the hardest things for investors to do is separate a technology’s usefulness from the durability of the profits earned by its producers.

Electricity is perhaps the clearest example. It transformed factories, homes, transportation, and communications. Thomas Edison helped create enormous value, but much of that value flowed outward to households, manufacturers, cities, workers, and consumers. Society captured the surplus.

The internet was different, at least for a narrow group of companies. Amazon, Google, Netflix, and others built ways to capture value. They owned the customer relationship, controlled distribution, and built data advantages, but their costs fell as usage increased. The more people used these platforms, the stronger their economics became.

That distinction matters, because investors do not own productivity. They own cash flows.

AI will almost certainly create value. It can help people write, code, research, design, analyze, and make decisions faster. It could also lower the cost of many cognitive tasks and enable new products. The question isn’t whether AI creates value. The question is: who keeps it?

For frontier AI models, the answer is not yet clear.

The cloud is grounded

The first challenge? Cost. Unlike the software platforms of the 2010s, frontier AI is not asset light. It may live in the cloud, but that cloud lives on the ground. AI requires datacenters, semiconductors, memory, networking equipment, electricity, cooling, land, permitting, skilled labor, enormous amounts of capital, and a lot of cement. Training AI models is expensive. Running them is expensive, too. Though every query is digital, they still consume physical resources.

This creates a different economic profile than many investors associate with technology. Internet 2.0 companies often benefited from powerful operating leverage: build the platform once, and each incremental user could be served at very low marginal cost. Some frontier AI models may face the opposite. More usage could mean more compute, more memory, more power, and more cost. Scale may grow revenue, but it may also grow the cost of goods sold.

When “good enough” is good enough

The second challenge is differentiation. Customers don’t always pay for the best technology. They pay for the technology that is “good enough” for the job at an acceptable price.

If a legal summary, software test, marketing draft, or customer-service response can be handled by a cheaper model, many customers won’t pay frontier-model prices for frontier-model performance. They will substitute. The token may look like a digital abstraction, but its price is whispering the market is tight. It tells users that compute, memory bandwidth, power, and inference capacity are scarce. When prices rise or budgets become visible, low-return uses migrate toward cheaper or more efficient models. Expensive intelligence gets saved for problems where the benefits justify the cost.

That does not mean frontier AI will fail, but it does mean the market may split. The most advanced models may be used for complex, high-value work, while everyday usage might shift toward smaller, cheaper, and more specialized systems. Adoption can rise while pricing power falls. Usage can explode while value capture disappoints. That’s an uncomfortable seat for investors.

There’s also a bargaining-power problem. If AI lowers costs for customers, competition may force those customers to pass savings on. If AI requires scarce chips, memory, and power, suppliers may capture the economics. If AI outputs are embedded inside someone else’s application, the owner of the workflow may capture the value. If models converge in quality, customers may pit providers against one another. AI creates value, but frontier model providers may not keep much of it.

This is why the distinction matters. A company can help change the world and still be a poor investment if it lacks pricing power, customer ownership, differentiated data, distribution control, or favorable unit economics.

Where’s the value capture?

Businesses that control scarce inputs, own trusted workflows, possess proprietary data, serve mission-critical use cases, or have the distribution to integrate AI within existing customer relationships are likely to pocket the gains.

The risk may be highest for businesses that spend heavily to stay competitive but can’t monetize that spending. These companies may adopt AI, fund AI, or sell AI-branded products yet still give away much of the benefit to others. That’s not a technology problem. It’s a business-model problem.

Where this leaves us

The market is still debating how large AI will become. That’s the easier question. The harder and more important question is how the value surplus will be divided.

Benchmarks don’t distinguish between companies that create value and companies that capture it. Indices own capital-intensive model builders, scarce suppliers, workflow owners, AI beneficiaries, and AI victims together. Conveniently for active managers, indices don’t sort the value creators from the value victims.

AI may become one of the most important technologies of our lifetimes, but value creation belongs to society. Value capture belongs to owners. The gap between the two? That’s the trade of the decade.