The artificial intelligence infrastructure boom that has driven corporate earnings to record highs faces an uncomfortable math problem: the cost of building data centers is rising so steeply that new facilities may never generate enough profit to justify their construction. Industry analysts and regulators are now grappling with whether the economics of AI infrastructure can sustain the investment frenzy that has defined 2025 and 2026.

What Are the Real Costs Behind AI Data Center Construction?

The numbers tell a sobering story. Nvidia CEO Jensen Huang recently estimated that building a single one-gigawatt data center could cost between $80 billion and $100 billion. That's an enormous upfront investment for what sounds like a massive facility, but the revenue picture is far less impressive. Companies that analyze energy infrastructure, including Cleanview and Lancium, estimate that a one-gigawatt data center operating AI models would generate roughly $10 billion to $12 billion in annual revenue.

Do the math, and the problem becomes clear: at those revenue levels, it would take 8 to 10 years for a new data center simply to recoup its initial construction costs, before accounting for ongoing operating expenses or generating any actual profit. That timeline creates a critical vulnerability because the graphics processing units (GPUs) and specialized AI chips that power these facilities typically become obsolete or significantly less valuable within 5 to 7 years as newer, more efficient hardware emerges.

How Are Regulators Responding to Data Center Power Demands?

The Federal Energy Regulatory Commission (FERC) took direct action on this crisis in mid-June 2026, recognizing that the grid itself cannot keep pace with data center growth. On June 18, FERC voted unanimously to order six of the country's largest grid operators to explain, within 30 days, how they would connect new AI data centers to the power grid faster. If regulators did not approve their answers, the operators had 60 days to rewrite their own connection rules entirely.

The order targets six regional transmission organizations and independent system operators that run the wholesale grid across most of the country outside Texas and the Southeast. These operators now face specific requirements:

• Capacity Assessment: Determine how much spare generating capacity exists on their systems to serve new large loads
• Queue Management: Develop rules ensuring large data center loads do not get stuck in interconnection queues for years or jump ahead of smaller customers
• Cost Allocation: Prevent new substation and transmission costs from being spread across residential electricity bills
• Full Cost Recovery: Require data centers to cover the complete cost of grid work tied to their own connections

The most contentious issue is cost allocation. For years, consumer advocates and utilities have fought over whether ordinary ratepayers should subsidize infrastructure built primarily for trillion-dollar cloud companies. FERC just made full cost recovery a federal requirement rather than leaving it to state-by-state negotiation.

The six grid operators affected are PJM Interconnection (covering the Mid-Atlantic and parts of 13 states), MISO (15 central U.S. states), Southwest Power Pool (14 central and southern states), California ISO, ISO New England, and NYISO. PJM is considered the most critical to watch, as it serves the Virginia data center corridor, one of the world's densest concentrations of cloud infrastructure, and already has the largest backlog of data center interconnection requests.

Why Is Data Center Power Demand Growing So Rapidly?

The numbers behind the growth are staggering. U.S. data center power demand stood at 61.8 gigawatts in 2025 and is projected to reach 75.8 gigawatts in 2026, then climb to 108 gigawatts by 2028 and 134.4 gigawatts by 2030. That represents more than a doubling in five years, concentrated in regions that overlap almost exactly with the six grid operators now under FERC's order.

The Electric Power Research Institute (EPRI) has modeled data center growth as a share of total U.S. electricity consumption. In 2023, data centers accounted for roughly 4 percent of total U.S. electricity load. By 2030, that share is expected to climb to somewhere between 9 and 17 percent, depending on how aggressively artificial intelligence buildout continues. This growth is happening in a country where building a new high-voltage transmission line can take longer than constructing the data center that needs it, creating a fundamental infrastructure mismatch.

What Does This Mean for AI Investment and Corporate Earnings?

The profitability crisis extends far beyond individual data center projects. American corporate earnings are currently nearly 60 percent above their historical trend, driven largely by the AI boom and the hyperscalers,Alphabet, Microsoft, Meta, and Amazon,that dominate cloud infrastructure. Analysts forecast S&P 500 earnings to grow more than 27 percent over the next 12 months, with earnings per share expected to exceed trend by more than 85 percent next year and by 100 percent by mid-2028.

However, these forecasts assume that hyperscalers can maintain their traditional business model of low capital expenditure and sustained competitive advantages. That assumption is breaking down. The hyperscalers have become significantly more capital-intensive because of massive spending on AI infrastructure, and some of their competitive advantages have weakened considerably due to intense competition throughout the AI value chain.

One critical vulnerability is customer switching. Large language model (LLM) users, including enterprises and developers, tend to switch between platforms depending on quality and cost. When Anthropic's Claude outperformed OpenAI's ChatGPT in March 2026, users were able to migrate to another platform with almost no friction, according to the Ramp AI Index. This means hyperscalers will either depend on users' preferences for a particular model or have to compete with other data centers for the new leader's business. In the first case, revenue growth will slow; in the second, margins will decline as competition intensifies.

The financial interdependencies in the AI ecosystem create additional risk. According to the Bank for International Settlements (BIS) Annual Economic Report 2026, researchers found that in 2025, more than half of hyperscalers' revenues and nearly all chip manufacturers' revenues came from what they termed "circular financing" rather than real contracts. If this circular mechanism stops because of lower capital expenditure or declining margins, overall S&P 500 profitability could revert to its historical trend, erasing the extraordinary gains of the past two years.

How Are Microreactor Companies Positioning Themselves?

Recognizing that the grid cannot supply enough firm, carbon-light power fast enough, hyperscalers have begun signing nuclear power agreements. This has created a rush among small modular reactor (SMR) developers to position themselves as solutions to the data center power crisis. Two companies dominate this space: Oklo and NuScale, both trading on U.S. exchanges as pure-play SMR stocks.

Oklo is developing a fast fission microreactor called the Aurora, designed to use recycled nuclear fuel. The company targets data centers, remote industrial sites, and the U.S. military as customers. Oklo went public via a special purpose acquisition company (SPAC) and was chaired by OpenAI CEO Sam Altman until April 2025, an association that continues to drive retail investor interest. However, Oklo has no commercial reactor operating and generating revenue. The company has signed letters of intent with several parties, but these are preliminary agreements, not financed power purchase agreements. A January 2026 agreement with Meta supporting a planned 1.2 gigawatt campus in Ohio with a prepayment mechanism represents progress, but first power is targeted for as early as 2030.

NuScale is further along in regulatory approval. It is the only SMR developer with a U.S. Nuclear Regulatory Commission (NRC) standard design approval for its light-water reactor design, and it added a standard design approval for its uprated 77 megawatt-electric module in 2025. However, NuScale lost its anchor project, the Carbon Free Power Project in Idaho, after cost estimates rose sharply and utility subscribers withdrew. The company has since refocused on international projects and a utility-led U.S. program with the Tennessee Valley Authority (TVA) and ENTRA targeting up to 6 gigawatts of NuScale modules. Like Oklo, NuScale has no operating reactor generating commercial revenue.

Both companies face a critical gap between their technology narratives and actual cash flow. Oklo carries higher technology risk but also more optionality if its fuel-recycling approach works at scale. NuScale has demonstrated more regulatory progress but faces the harder question of whether it can find customers willing to accept its cost structure. Every year of licensing delay is another year of cash burn without revenue, and both companies fund operations through equity raises and government grants, which dilute existing shareholders.

Steps to Understanding the Data Center Economics Challenge

• Track Construction Costs: Monitor announcements from hyperscalers about data center construction budgets and timelines, as rising costs directly impact the payback period for new facilities
• Follow FERC Filings: Watch for the 60-day tariff filings from the six grid operators named in the June 2026 order to understand how interconnection rules will change and what costs data centers will bear
• Assess SMR Progress: Review quarterly updates from Oklo and NuScale on licensing milestones and customer agreements, as these will determine whether nuclear power can actually solve the data center power crisis
• Monitor Hyperscaler Earnings: Pay attention to capital expenditure guidance from Alphabet, Microsoft, Meta, and Amazon in earnings calls, as declining investment would signal a shift in confidence about data center profitability

The artificial intelligence boom has created extraordinary wealth for investors and record earnings for technology companies. But that boom rests on an assumption that data centers will eventually generate enough revenue to justify their construction costs. If the economics continue to deteriorate, if grid connection timelines extend further, or if nuclear power fails to materialize as a solution, the entire investment thesis could unwind. The next 12 to 24 months will be critical in determining whether AI infrastructure becomes a sustainable business or a cautionary tale about the limits of exponential growth.