AI Capex Returns: The Trillion-Dollar FOMO Arms Race and Key ROI Indicators
A massive debate is shaping the equity markets in May 2026 regarding the return on investment (ROI) for the historic artificial intelligence capital expenditure boom. Hyperscalers and technology companies are deploying unprecedented capital, but early indicators signal that the "circular loop" of AI spending has not yet diffused into the broader enterprise economy.
The Scale of the AI Build-Out
According to the Goldman Sachs Global Institute's report, Tracking Trillions (published May 1, 2026), baseline estimates project ~$7.6 trillion in cumulative AI capital expenditure between 2026 and 2031 across compute, data centers, and power infrastructure.
- Annual Capex: Projected to reach $765 billion in 2026 and more than double to $1.6 trillion by 2031.
- Supply-Side Sensitivity: The cumulative investment scale is highly sensitive to core physical assumptions. The single most influential variable is the economic useful life of AI silicon (typically estimated at 4 to 6 years); small shifts in the replacement cadence can swing cumulative spend by hundreds of billions of dollars.
- Rising Costs: Next-generation AI-optimized data centers now cost $15 million to $20 million per megawatt, up from the $10 million standard of the 2010s, due to extreme cooling and power density requirements.
The ROI Challenge and Refuting Indicators (The Bear Case)
Jim Covello, Goldman Sachs' Head of Global Equity Research and a prominent AI skeptic, highlights that after two years of tracking, the actual economic returns of AI remain highly concentrated at the semiconductor level (specifically Nvidia), while the companies deploying the technology are seeing modest or negative financial impacts. Key early indicators of inadequate returns in mid-2026 include:
- Zero Return on Pilots: Covello cites an MIT Labs report showing that 95% of organizations are getting zero return on their AI pilots.
- Enterprise Losses: An EY survey found that 99% of sampled companies reported financial losses due to AI-related risks, averaging $4.4 million per company.
- Productivity Drains: Research from the Harvard Business Review reveals that AI-generated errors ("workslop") cost a 10,000-person organization over $9 million annually in lost productivity.
- Low Production Deployment: While over 70% of enterprises use basic tools like Microsoft Copilot, only 8.6% of enterprises report having AI agents actually deployed in production, with the vast majority stuck in "pilot purgatory."
- Tight Circular Loop: Nvidia's revenue is heavily dependent on a handful of buyers (four customers contribute 61% of its revenue). If this concentration exceeds 70%, or if hyperscaler capex decelerates by more than 20% year-over-year, it signals that the infrastructure surplus is a structural balance-sheet liability rather than a growth catalyst.
Confirming Indicators of Long-Term ROI (The Bull Case)
Conversely, several key metrics would confirm that the AI capex floor is solidifying and diffusing into the real economy:
- Non-Tech Sector Adoption: Non-tech sectors show early signs of high ROI when adoption succeeds (e.g., healthcare reports a 3.2x ROI and compressed purchasing cycles; financial services show 4.2x returns; and manufacturing shows a 61% cost decrease from AI supply chains). If annual AI spending in healthcare and manufacturing exceeds $20 billion annually, it indicates healthy diffusion.
- Deal Conversions & Budgets: Enterprise AI deal conversion rates sustaining above 45%, and Fortune 2000 AI budgets growing 30% or more with documented productivity gains.
The "FOMO" Arms Race & Investment Implications
A striking finding from Goldman Sachs is that the primary engine driving this capital deployment is not a rational ROI calculation, but corporate insecurity. Hyperscalers (Microsoft, Amazon, Google, Meta) have dramatically increased their AI capex in 2026 even as their stocks have lagged the S&P 500, prioritizing the fear of missing out ("FOMO") over near-term shareholder returns.
From an investment perspective, Goldman Sachs suggests a contrarian structure: go long hyperscalers and underweight semiconductors. If enterprise ROI eventually materializes, hyperscalers (currently priced with deep skepticism) have significant room to run. If ROI continues to disappoint, hyperscalers will eventually cut capex, triggering a cash-flow-relief rally. Semiconductors, conversely, are priced for an endless arms race and remain highly vulnerable to any spending cuts.