Goldman Sachs estimates tech companies will spend $7.6 trillion through 2031 building thousands of data centers to power AI. A May 2026 study from Gartner found that businesses replacing workers with AI agents often fail to generate a return on investment. This week, the Nasdaq Composite dropped nearly 3% as investors began pricing in the possibility that the most expensive infrastructure buildout in corporate history may not have enough customers to justify it, according to CBS News.
The tension is straightforward: the supply side of AI is scaling at a rate the demand side cannot match. And for agent builders and operators, the Gartner finding is the sharpest data point in a growing body of evidence that agent automation does not automatically translate to financial returns.
The Supply Side: $7.6 Trillion and Counting
Goldman Sachs published its “Tracking Trillions” analysis in May 2026, identifying four variables that determine the total scale of AI infrastructure spending: the economic useful life of AI silicon, the cost and complexity of next-generation data centers, the chip and architecture mix, and elongation from power, labor, and equipment bottlenecks. Small shifts in any of these assumptions move cumulative spend by hundreds of billions.
The hyperscalers driving this buildout are Alphabet, Amazon, Meta, Microsoft, and Oracle. All five are spending heavily on data centers and chips in anticipation of strong demand for AI services, while large language model developers like OpenAI and Anthropic pay to use their infrastructure.
Kate Brennan, associate director of the AI Now Institute, told CBS News that the returns are not materializing: “The returns are not coming in, and the claims that are being made, in terms of efficiency or productivity numbers, are not netting out.”
Brennan described the current adoption wave as supply-driven rather than demand-driven: “The current push for AI adoption that we’re seeing is directly coming from the financial incentives of AI firms.” The hyperscalers and AI companies are making a “deliberate push for AI everywhere, no matter whether the demand is there or if customers want it or not.”
The Demand Side: Consumers Won’t Pay, Workers Don’t Trust
The consumer side of the equation is documented in two studies published in June 2026.
Pew Research Center found that 40% of American adults believe AI will be a negative societal force over the next two decades, compared to 16% who believe it will be positive. Americans are increasingly using AI, but their engagement is often involuntary: Google surfaces AI responses at the top of search results, and customer service lines route callers to AI agents before offering human alternatives.
A Bank of America Institute study on consumer AI usage found that few consumers appear willing to pay for AI services. Usage is growing, but willingness to pay is not keeping pace with the infrastructure investment required to deliver those services.
These are not niche findings. They describe a structural problem: the companies spending trillions on AI infrastructure are simultaneously discovering that the end users they need to justify those investments either don’t want AI, don’t trust it, or won’t pay for it.
The Agent-Specific Problem: Replacing Workers Without Replacing Value
The Gartner study from May 2026 zeroes in on the specific failure mode most relevant to NCT’s audience. Businesses that replace workers with AI agents often fail to generate a return on investment. The finding does not claim agents are incapable of creating value. It claims that the most common deployment pattern, direct worker replacement, is not delivering financial returns.
This matters because direct replacement is exactly the pitch that enterprise AI vendors have been making for the past 18 months. Deploy agents, reduce headcount, cut costs, increase margins. The Gartner data suggests this sequence breaks down somewhere between “deploy agents” and “increase margins.”
The reasons are not mysterious. Agent deployments carry costs that worker-replacement math often ignores: integration time, supervision overhead, error remediation, retraining adjacent workers to collaborate with agents, and the ongoing token and compute costs of running agents at scale. When an enterprise replaces a $70,000/year employee with an agent that requires $40,000/year in compute, $15,000 in integration work, and $20,000 in ongoing supervision and error correction, the math doesn’t favor the agent.
NCT has reported extensively on the scaling side of this equation. AWS attributes $15 billion in annual recurring revenue to AI agent adoption. Vistra logged 80,000+ AI actions and 13,000+ productivity hours. Salesforce published governance frameworks drawn from 20,000+ production agents. But deployment volume does not equal ROI. Running 20,000 agents proves operational capability. Whether those agents generate net positive returns after all costs is a separate question that Gartner’s data suggests many enterprises cannot yet answer affirmatively.
The Capex Payback Problem
Economist Ed Yardeni of Yardeni Research examined annualized revenue estimates for OpenAI and Anthropic to assess whether they are adding users fast enough to cover their spending commitments with the hyperscalers. He calls this a “capex payback test.”
His conclusion, reported by CBS News: “We find that the AI ecosystem is not fully end-user revenue-backed yet, but it is not entirely speculative either. Expected 2030 revenues make the math look much better. But those forecasts depend on a big assumption: AI revenues must continue to scale, and compute efficiency must improve, or both.”
The conditional is critical. The current investment thesis requires sustained revenue growth through 2030, combined with compute efficiency improvements that reduce the cost of delivering AI services. If either assumption fails, the payback math deteriorates.
Yardeni’s framework applies directly to agent infrastructure. Agent compute costs are not falling as fast as base model inference costs, because agents require multi-turn reasoning, tool calls, and persistent memory. An agent that makes 15 tool calls to complete a task consumes 10-50x the tokens of a single model inference. Compute efficiency gains at the model layer may not translate to equivalent cost reductions at the agent layer.
The Dotcom Parallel and Its Limits
Wall Street analysts are explicitly drawing comparisons to the dotcom bubble of the late 1990s. Qian Wang, global head of capital market research at Vanguard, and senior global economist Kevin Khang wrote this week that the AI landscape is likely to yield uneven outcomes, according to CBS News: “Some firms may emerge as more profitable and with significant competitive advantages, while others could find their core businesses obsolete in a new AI economy.”
Jonas Goltermann, chief markets economist at Capital Economics, told CBS News the rally in AI-related equities is winding down, and expects tech stocks to drop, perhaps sharply, in 2027.
The dotcom comparison is instructive but imprecise. The late-1990s internet bubble involved companies with zero revenue building infrastructure for theoretical demand. The 2026 AI buildout involves companies with massive existing revenue bases (Alphabet, Amazon, Microsoft) diverting capital toward AI infrastructure. These companies have massive existing revenue bases and won’t go bankrupt. Their real exposure: AI spending becomes a multi-year drag on returns while demand catches up, leaving enterprise customers who deployed agents on vendor promises locked into infrastructure commitments with negative ROI.
The Token Spending Backlash
The demand-side skepticism is already manifesting inside enterprises. NCT reported this week that corporations are implementing token budgets and agent loop oversight, reversing the move-fast-burn-tokens culture of early 2025-2026. Companies that deployed agents aggressively are now discovering that uncapped token spend creates unpredictable cost structures, and are pulling back.
This is the micro version of the macro problem Goldman Sachs identified. At the infrastructure level, hyperscalers are spending trillions on capacity. At the enterprise level, customers are capping how much of that capacity they actually consume. If every enterprise implements token budgets that limit agent activity, the aggregate demand for AI compute may fall well below what the $7.6 trillion buildout was designed to serve.
The RAISE US Counter-Signal
Against this backdrop, the RAISE US initiative, backed by OpenAI, Anthropic, Amazon, Microsoft, Bank of America, and JPMorgan Chase, launched a $500 million workforce retraining program targeting AI job displacement. The initiative, led by former Commerce Secretary Gina Raimondo and former Indiana Governor Eric Holcomb, plans to deploy capital across partner states over 3-4 years.
The timing is revealing. The same companies spending trillions on AI infrastructure are simultaneously funding programs to retrain workers displaced by that infrastructure. The framing is humanitarian; the logic is risk management. AI companies recognize that uncontrolled job displacement creates the political backlash and regulatory pressure that could constrain their addressable market. RAISE US is a social license investment: a $500 million cost of doing business to protect a multi-trillion dollar bet.
But it also validates the Gartner finding. If AI agent deployment reliably generated positive ROI, there would be less workforce displacement to manage. The need for a $500 million retraining fund suggests the displacement is real while the returns are not.
Where the Math Breaks
The AI infrastructure buildout rests on a chain of assumptions:
Assumption 1: AI compute demand will grow exponentially through 2031. Goldman Sachs models this across multiple scenarios, noting that the $7.6 trillion figure is sensitive to silicon replacement cadence, data center costs, and power constraints.
Assumption 2: Enterprise customers will convert pilots to production deployments at scale. The evidence here is mixed. Salesforce has 20,000+ production agents, but Gartner says many of those deployments don’t generate positive returns.
Assumption 3: Consumers will pay for AI services. Pew and Bank of America data suggest otherwise, at least for now.
Assumption 4: Compute efficiency improvements will reduce per-unit costs fast enough to make AI services profitable at current pricing. For base model inference, this is plausible. For agent workloads with multi-turn reasoning and tool calls, the efficiency curve is shallower.
Assumption 5: The political and regulatory environment will remain permissive. RAISE US exists because the companies driving the buildout are already hedging against this assumption failing.
If any two of these assumptions fail simultaneously, the payback timeline extends beyond what investors will tolerate. This is what the Nasdaq’s 3% drop this week is pricing in: not catastrophic failure, but the growing probability that the payback curve is longer and less certain than the infrastructure investment assumes.
The Production Test for Agent Builders
For teams building and deploying agents, the Gartner finding is a direct operational warning. The question every agent deployment should answer before scaling: after all costs, including compute, integration, supervision, error remediation, and displaced worker retraining, does this deployment generate positive returns within 12 months?
If the answer is “we’ll know in 2030,” the deployment is speculative infrastructure, not a proven business case. And as the Nasdaq selloff and enterprise token backlash demonstrate, the market’s patience for speculative AI infrastructure is not infinite.