This week, the AI ROI question moved all the way up the stack. It's no longer just "is AI paying off for my team?" It's "does the entire chain — from data centers to model makers to the companies buying the tools — actually pay back?" And the honest answer is: not yet. The money at the top is waiting on the work at the bottom.
Two Stanford economists put a number on it: a nearly $1 trillion gap between what the largest hyperscalers have spent and the AI revenue they've taken in since 2024. Meanwhile, a new PYMNTS study found that enterprises with AI embedded in three or more business functions were far more likely to report a return (93%) than those with AI embedded in none (25%). Those two findings are the same story told from opposite ends.
Here's why that matters operationally. Every dollar of that infrastructure gets recovered one way: customers using AI deeply enough that it becomes worth paying for, renewing, and expanding. Not logging in. Not experimenting. Embedding it into the daily work. Most organizations are still in the "broad but shallow" stage — lots of licenses, lots of pilots, a handful of enthusiasts — and wondering why the returns feel thin. The evidence this week says the returns are thin because the use is thin. Here's the week's data, and what it means for leaders and professionals trying to make AI pay.
The spend-to-revenue gap at the top is roughly $1 trillion. Stanford economists Jared Bernstein and Ryan Cummings estimate a nearly $1 trillion gap between hyperscaler spending — Alphabet, Amazon, Meta, Microsoft, Oracle and SpaceX — and the AI revenue those companies have taken in since 2024. Their conclusion: AI revenue would need to triple or quadruple next year and every year after for a decade "for this to work out." Goldman Sachs agrees that hyperscaler AI revenue remains below what's needed to break even on capex, but is more optimistic that the investments ultimately pay off. This is analysis and projection, not audited accounting — but it frames the pressure that flows downstream to every buyer. (Axios, Sep 30)
The model makers are locking in fixed costs against future demand. Reuters, reviewing Anthropic's confidential IPO prospectus, reported roughly $518 billion in compute commitments over about a decade, with roughly 80% non-cancelable — alongside a 2025 net loss of about $42 billion, largely driven by a non-cash charge, and an expected second straight quarter of adjusted operating profit. Separately, Barclays estimates AI-related debt and equity raised by hyperscalers, data-center vehicles and neo-clouds reached about $346 billion year to date, versus $172 billion for all of 2025. Enterprise demand is real, but so are the obligations it now has to cover. (Reuters, Sep 28–29; Barclays estimate via Yahoo Finance)
Depth, not breadth, separates the companies reporting returns. PYMNTS surveyed 60 U.S. enterprises across 437 AI deployments. Among companies with no function where AI is embedded in regular work, 25% said AI already delivers a return. That rose to 55% with one or two embedded functions and 93% with three or more. Only 20% of deployments were embedded in a function's regular work; 37% were broad but not daily, and 43% were still limited or exploratory. Important qualifiers: the sample is small, returns are self-reported rather than measured, and the relationship is correlation, not proof of causation. (PYMNTS, "AI at Work," Sep 29)
Among the largest U.S. companies, measured value is still rare. a16z's State of Markets update found that nearly 30% of S&P 500 companies report some "quantifiable impact" from AI — but only about 2% report any tracked metric. The firm's summary: adoption "is broad, but for the most part, it's relatively shallow." (a16z, State of Markets II, Sep 30)
The pattern that matters
Put those four findings side by side and a single chain appears. Hyperscalers are spending ahead of revenue. Model makers are signing long, mostly non-cancelable commitments. Capital markets are financing the buildout at roughly double last year's pace. And all of it depends on the same final link: enterprises turning AI from something available into something embedded.
That last link is where value is being lost. Most organizations measure AI adoption by access — seats purchased, people trained, prompts run. But access is the shallow end. The PYMNTS data suggests the return doesn't climb steadily with more tools; it climbs when AI becomes part of how a function does its regular work. The a16z data shows what happens when that doesn't happen: plenty of companies can describe an impact, very few can point to a metric they track.
What leaders often misunderstand is that breadth feels like progress. Rolling AI out to every department looks like momentum on a slide. But ten departments at 10% depth may produce less measurable value than two departments at full depth — because only the deep deployments change cycle times, error rates, capacity and cost in ways you can actually see. Even vendors are shifting their framing; Google Cloud's new "ROI of AI 2026" report, built on 2,400+ executives, centers on a scorecard of ROI measurement methods rather than adoption counts.
The implication is practical: stop asking "how many teams have AI?" and start asking "in how many workflows is AI doing real work every day, with someone accountable for the result?"
The EMDUIT take — for the professional putting AI to work
Picture a property management company that gives every manager an AI assistant. Some use it to draft owner emails. Some summarize leases. A few try it for maintenance triage. Everyone agrees it's helpful. Nobody can say what it's worth. That's broad and shallow.
Now picture the same company choosing one workflow — maintenance requests — and embedding AI in it. Every incoming request is classified for urgency, matched to the right vendor, and drafted into a work order. A coordinator reviews and approves each one. The system captures time-to-dispatch, repeat visits and tenant follow-ups. Each month, someone reviews the numbers and adjusts the rules. Then they move to the second workflow, and the third.
That's the difference between having access, finishing an isolated task faster, and actually redesigning the work. It's also the difference between hoping AI pays off and being able to show it. This is why EMDUIT focuses on the full operating loop around AI: the task, the behavior, the guardrails, the handoff, the measurement, and the next improvement.
The industry has placed a trillion-dollar bet that AI will become part of how work gets done. Whether it pays off for you depends on whether it actually does. Broad AI adoption creates activity. Embedded AI creates returns.
Sources & methodology
Sources: Axios (Sep 30, 2026) reporting on Stanford economists Jared Bernstein and Ryan Cummings and Goldman Sachs analysis | Reuters reporting on Anthropic's confidential IPO prospectus (Sep 28–29, 2026) | Barclays estimates via Yahoo Finance | PYMNTS, "AI at Work: Why Deeper Enterprise Use Produces Stronger Returns" (Sep 29, 2026) | a16z, State of Markets II (Sep 30, 2026) | Google Cloud, "ROI of AI 2026" (vendor-published). Methodology note: The $1 trillion gap and revenue-growth requirement are economist analysis and projection, not audited figures. Anthropic figures come from Reuters' review of an unpublished prospectus. The $346 billion figure is a Barclays estimate that includes both debt and equity. PYMNTS returns are self-reported from a 60-company sample and show correlation, not causation. Google Cloud's report is vendor-published and cited for framing only.