Who Will Earn the Returns on the AI Buildout?

The Artificial Intelligence (AI) Capacity Race Is Becoming an Economics Test.

The industry is finding ways to add chips, power and data center capacity. But can returns keep pace with the spending—and who will capture them?

In 2024, we examined whether power and physical infrastructure could constrain AI's expansion. Two years later, those constraints remain, but the buildout has accelerated well beyond graphics processing units (GPUs). Early evidence of monetization is emerging. The focus is shifting from what must be built to who can earn attractive returns on it.

Putnam portfolio managers Andy O'Brien, CFA and Bobby Gray examine the opportunity from different ends of the AI value chain. Andy focuses on the technology and infrastructure needed to support rising demand. Bobby follows the business models, competitive dynamics and division of profits. Their perspectives are complementary: One tests the durability of AI demand; the other tests the durability of AI returns.

One cycle. Two investment tests.

Andy O'Brien: The Buildout May Still Be Underestimated

I see powerful reasons for AI investment to continue. Companies are spending offensively to pursue new revenue and defensively to avoid falling behind without enough computing capacity or AI intellectual property. Overbuilding could waste capital; underbuilding could threaten a company's position in the next major computing platform.

AI adoption remains in its early stages, leaving potential room for further demand growth. Many businesses use AI for isolated tasks, but few have rebuilt core workflows around it at scale yet. As models take on more complex work, even modest adoption growth could create a disproportionate increase in infrastructure demand.

The opportunity is also broadening. GPUs remain essential, but the buildout increasingly depends on memory, networking, optical components, custom silicon and semiconductor equipment. Putnam's analysis suggests earnings expectations have continued to improve across the infrastructure stack, with gains extending beyond the leading GPU supplier.

Wider demand does not guarantee broad-based benefits. Bottlenecks can support pricing, but shortages attract capacity and substitutes. What matters more is which suppliers remain valuable after scarcity eases.

One measure I follow is monetization per gigawatt, or how much revenue a given amount of data center power capacity can support. The cost of building that capacity has risen, but so has its revenue potential. That could increase the value of existing computing fleets and give cloud providers greater confidence to invest. The market may still be underestimating both how much infrastructure AI will require and how attractive the returns on investment could be.  

Bobby Gray: Revenue Is Arriving. The Return Debate Is Not Over.

I approach the cycle from the other side of the equation. For much of the buildout, spending expectations rose faster than forecasts for revenue or profit. For Amazon, entering this year, consensus capital spending expectations for 2027 had increased roughly 120% over the prior two years, while revenue estimates for that year were essentially unchanged, according to the managers’ analysis. 

More recently, however, the picture has begun to shift. Revenue growth and backlogs have accelerated, suggesting monetization may be starting to catch up with spending. The key test is whether those gains justify the much higher level of spending. Cloud providers are facing more competition, even as large AI developers become increasingly important customers for computing capacity. That could change long-term pricing dynamics, margins, and risk. Enterprises can also route tasks among models based on cost, speed, and performance, while open-weight models may intensify competition.

Those dynamics make value capture difficult to predict. The company producing the most advanced model will not necessarily retain the greatest share of the economics. Established software providers may add AI to products customers already use, while AI-native applications create new workflows. Hyperscalers may benefit from owning infrastructure and customer relationships even when the preferred model changes.

Model performance matters, but I believe distribution, proprietary data, workflow integration, switching costs, and control of the customer may matter even more. AI can create enormous value for users without guaranteeing attractive profits for every provider.

Three questions for the payoff phase

Together, the perspectives of Andy and Bobby point to three questions that move investors beyond simple AI exposure.

1. Is demand durable and broadening? We believe investors should look beyond headline spending to adoption, workload intensity, and what they imply for infrastructure demand.

2. Do the economics improve with scale? Revenue growth must translate into cash flow, attractive returns on capital or stronger customer economics.

3. Who controls the customer and retains the value? It’s important to assess pricing power, distribution, switching costs—and ultimately, how interchangeable AI models and services may become.

Early Is Not an Investment Thesis

AI can remain a powerful investment theme without rewarding every participant equally. The buildout can continue to expand while the payoff becomes more selective.