The Trillion-Dollar Bet
Why companies keep investing in AI even as the returns are still lagging
Over the past five
years, global corporate capital has carried out one of the largest resource
reallocations in recent economic history. No previous technology—not industrial
electrification, not the interstate highway system, not the broadband buildout during
the dot-com bubble—concentrated so much capital spending in so little time as
artificial intelligence has between 2021 and 2026. The question now dominating
boardrooms is no longer whether it is worth investing in AI, but whether the
current pace of investment is sustainable, and whether the market—understood as
the real revenue being generated, investors' willingness to keep financing the
buildout, and the system's productive capacity to absorb that much capital—can
keep pace over the next five years.
This article examines three linked questions now shaping the boardroom conversation: how much has actually been invested, whether that investment is translating into measurable economic return, and how feasible it is to sustain—or even accelerate—that pace through 2030.
FIVE YEARS OF UNPRECEDENTED SPENDING
The starting point is easy to state and hard to absorb: capital expenditure by the five largest U.S. hyperscalers—Microsoft, Amazon, Alphabet, Meta, and Oracle—rose from roughly $256 billion in 2024 to somewhere between $660 billion and $725 billion committed for 2026, a jump of 67% to 77% in a single fiscal year. Goldman Sachs projects that cumulative hyperscaler spending between 2025 and 2027 will reach $1.15 trillion, more than double the $477 billion spent over the prior three years. Research firms such as TrendForce push the figure even higher when the nine largest cloud providers are included: up to $830 billion in 2026 alone.
Placed in historical perspective, a recent analysis cited by the Wall Street Journal puts AI-related capital spending, as a share of U.S. GDP, above the peak investment levels of the Manhattan Project, the Apollo program, the buildout of the electrical grid, and the construction of the interstate highway system—surpassed only by the Louisiana Purchase and, possibly, the peak of 19th-century railroad construction.
On top of this infrastructure figure comes corporate spending on adoption, licensing, and internal AI capability-building, which the 2026 AI Index from Stanford HAI—the field's most widely cited academic benchmark—estimates at $581.7 billion globally for 2025, more than double (+129.9%) the prior year. Of that total, pure private investment—venture capital, funding rounds, and acquisitions—totaled $344.7 billion, up 127.5% year over year. Organizational adoption, per the same report, reached 88% of surveyed companies.
Financing this expansion no longer relies solely on operating cash flow. Big Tech issued a record $428 billion in corporate bonds during 2025 to fund data-center construction, and some industry projections anticipate up to $1.5 trillion in additional debt issuance in the years ahead. Hyperscaler capital intensity—capex as a share of revenue—now sits between 45% and 57%, levels that historically belonged to utilities or heavy industry, not high-margin software companies.
IS THIS INVESTMENT PAYING OFF?
This is where the narrative splits sharply, and where serious analysis—as opposed to the industry's promotional talking points—has to distinguish between the macroeconomic level and the level of the individual firm.
At the macro level, the story is one of genuine growth. Frontier-lab revenue has surged at a pace with no precedent in corporate history. Anthropic went from an annualized revenue run-rate of roughly $1 billion in January 2025 to about $30 billion by April 2026—a 30x increase in fifteen months—and projects its first operationally profitable quarter. OpenAI reached a $25 billion annualized run-rate by late February 2026, though it continues to project losses of $14 billion to $17 billion for this fiscal year alone, with a break-even point the company itself has pushed out to 2029–2030. Harvard economist Jason Furman has estimated that AI-related infrastructure investment accounted for roughly 92% of U.S. GDP growth in the first half of 2025—a figure that, read carefully, says as much about AI's strength as it does about the fragility of the rest of the economy.
At the level of the company adopting AI—not the one selling it, but the one buying and implementing it internally—the picture is markedly more sober. The most consequential study of the past twelve months comes from the MIT NANDA Initiative, "The GenAI Divide: State of AI in Business 2025," which analyzed 300 public generative-AI deployments, 150 executive interviews, and surveys of 350 employees. Its central finding: 95% of generative-AI pilot projects produced no measurable impact on the P&L; only 5% of implementations succeeded in integrating into workflows in a way that generated significant financial value. The study identifies what it calls a "learning gap": the problem does not lie in model quality but in organizations' inability to integrate generic tools into specific business processes. Tellingly, the study found that companies concentrate most of their generative-AI budgets in sales and marketing functions—the ones with the lowest observed returns—while back-office functions, such as administrative process automation and customer service, are the ones producing the most consistent savings.
This pattern is not an isolated case. U.S. Census data on companies with more than 250 employees suggest that generative-AI adoption plateaued during 2026 and, in some segments, began to decline. A widely cited Atlassian study found that 96% of companies failed to achieve significant productivity gains from their most recent AI tools.
The most rigorous academic counterpoint to this debate comes from MIT itself: economist Daron Acemoglu, the 2024 Nobel laureate, estimates in his paper "The Simple Macroeconomics of AI" that only about 5% of tasks in the U.S. economy will be able to be profitably performed by AI over a ten-year horizon, which would translate into a GDP increase of just 1.1% to 1.6% over ten years—an effect Acemoglu describes as "non-trivial, but modest," far below Goldman Sachs' projection of $7 trillion or the $17 trillion to $25 trillion in annual value estimated by the McKinsey Global Institute. This gap between macroeconomic projections—ranging from modest to transformative—is itself the best evidence that the sector still lacks consensus on the aggregate return on this investment.
Some nuance to the pessimism is warranted, however: Stanford's AI Index documents that U.S. consumer surplus from generative-AI tools reached about $172 billion annually in early 2026, up 54% in a single year, with the median value per user tripling between 2025 and 2026. The problem, the report's own authors note, is that this value is being captured mostly by end users and consumers—via free or low-cost tools—rather than by the companies that financed the underlying infrastructure. It is, in other words, a surplus leaking down the value chain before it adequately compensates those who built the foundation.
HOW MUCH MORE WILL NEED TO BE INVESTED THROUGH 2030
Available projections agree that the investment cycle will not only continue but will accelerate before it stabilizes. McKinsey estimates that global demand for data-center capacity could nearly triple by 2030, with roughly 70% of that demand driven directly by AI workloads. A recent analysis built on that projection puts total required capital spending at $6.7 trillion by the end of the decade, of which $5.2 trillion would go specifically to AI processing infrastructure and $1.5 trillion to traditional IT workloads.
Layered on top of this are commitments already announced that go beyond the traditional hyperscalers: the Stargate project, with a stated ambition of $500 billion over five years among OpenAI, SoftBank, Oracle, and other partners; growing sovereign investment from Saudi Arabia, the United Arab Emirates, and Japan; and individual corporate announcements such as South Korea's SK Group commitment of more than $500 billion under letters of intent with Nvidia. The operational takeaway for any board is that aggregate sector spending will not merely fail to slow in the near term—it will likely double from the already-elevated 2026 levels before consolidated evidence of return emerges at the level of the average adopting company.
CAN THE MARKET SUSTAIN THIS PACE?
This is the question separating structural optimists from cyclical skeptics, and both camps now have real—not merely speculative—evidence behind them.
The case for sustainability rests on three pillars. First, revenue concentration: Anthropic reports that more than 500 companies now spend over $1 million annually on its platform, with eight of the Fortune 10 among its customers, and that enterprise customers generate three to five times more revenue per token than consumer users, with more predictable and cheaper-to-serve usage patterns. Second, the efficiency trajectory: Anthropic reached a revenue run-rate comparable to OpenAI's while spending a fraction on model training, suggesting the cost curve per unit of intelligence continues to improve faster than aggregate infrastructure spending. Third, the scale of adoption: with 53% global population adoption in just three years—faster than the personal computer or the internet, according to Stanford HAI—the installed base of users and use cases keeps expanding, supporting the thesis that monetization, though lagging, will eventually catch up with infrastructure.
The case against is equally solid. Torsten Sløk, chief economist at Apollo Global Management, has warned that valuations of AI-linked stocks already exceed, by some measures, those of dot-com companies in 1999. A systemic-risk analysis from Oliver Wyman warns that a loss of investor confidence would trigger capital-spending cutbacks that would compound the GDP slowdown, given that much of recent economic growth depends on this very investment. And the financing pattern itself—increasingly reliant on corporate debt issuance rather than operating cash flow—introduces a fragility that did not exist in previous technology investment cycles financed primarily with equity capital.
What the research from Stanford, MIT, and Harvard reveals, taken together, is that the sustainability of the cycle does not hinge on a binary answer but on a widening split between two populations of companies. A small group of infrastructure and frontier-model providers—along with a still smaller number of adopting companies that have managed to integrate AI into specific, measurable processes—are capturing real and growing returns. The vast majority of organizations that adopted generative AI during the 2023–2025 enthusiasm peak have not, to date, managed to translate that adoption into verifiable financial impact, and they run the risk of having their AI budgets cut in the next round of capital discipline.
IMPLICATIONS FOR DECISION-MAKING
For executive teams, the correct reading of this evidence is not "invest less," but invest differently. The MIT study itself offers the variable that best predicts success: not the size of the budget nor the sophistication of the model used, but the degree of operational integration between the tool and a specific business process, with clear metrics and clear ownership. The companies that made it into NANDA's successful 5% shared a common pattern: they started from a concrete operational problem, worked with specialized external vendors rather than building everything in-house, and measured impact in business terms—not in adoption figures or "active user" counts.
The strategic lesson, ultimately, is the same one that has accompanied every general-purpose infrastructure cycle since electrification: capacity-building precedes—almost always by years—the widespread capture of value. The question every board needs to answer is not whether AI will generate returns—the evidence from Stanford and from Anthropic suggests that, for those already achieving it, the return is substantial and growing—but whether the organization itself has the implementation discipline needed to land inside that 5%, rather than funding, once again, a pilot that will never reach production.
Sources: Stanford HAI, AI Index Report 2026; MIT NANDA Initiative, "The GenAI Divide: State of AI in Business 2025"; Daron Acemoglu (MIT), "The Simple Macroeconomics of AI"; Goldman Sachs Research; McKinsey Global Institute; Harvard University (Jason Furman); Oliver Wyman; Apollo Global Management.
