Goldman's Bullish AI Cash Flow Thesis Has a Timing Problem Its Own Report Doesn't Notice

Goldman Sachs says AI agents will eventually make tech companies richer. Its own numbers suggest the money is flowing somewhere else.

The bank published research last week projecting that agentic AI will drive a 24-fold surge in token consumption by 2030, with inference costs falling 60–70 percent annually — a combination the authors say will produce a "gross margin inflection" at hyperscalers within three to twelve months. The framing was straightforward: AI infrastructure spending is large now, but the returns are coming.

Buried in the same institution's concurrent research, a different picture emerges. Goldman Sachs analysts have also reported that Nvidia captures roughly 75 percent of total AI compute spend at approximately 75 percent gross margins, that 95 percent of enterprise AI pilot programs return zero measurable return, and that data center debt issuance among hyperscalers doubled to $182 billion in 2025 alone as free cash flow compressed under the weight of capital expenditure. The same bank that says AI will bail out tech cash flows is also documenting, in its own words, that the economics of the AI buildout flow disproportionately to one layer of the stack.

"The concern in the generalist investor community is the sustainability of capex because the free cash flows of hyperscalers have been compressed," wrote analyst Jim Schneider in the May 20 research note. "What fixes that? The answer lies in the underlying economics of the problem. If you raise gross margins, you raise operating cash flow, and that gives you more headroom to spend."

That logic is sound in theory. In practice, the margin inflection Goldman is projecting depends on demand for agentic AI workloads materializing at scale before the chip shortage eases — a shortage the same report warns will persist for twelve to eighteen months. Those timelines do not fully align.

The $700 billion question

The four largest hyperscalers — Google, Microsoft, Amazon, and Meta — are projected to spend roughly $700 billion on AI capital expenditure in 2026, a 60-plus percent increase from 2025, according to Goldman Sachs estimates. That spending is being financed in part through debt. Data center debt issuance doubled year-over-year in 2025, a figure the bank's own research describes as a response to free cash flow compression.

The margin inflection thesis depends on two things happening simultaneously: inference costs continuing to fall at 60–70 percent per year, and agentic AI adoption accelerating fast enough to absorb the capex before the debt burden forces harder choices. Goldman's model projects that daily LLM queries will grow at a 40 percent compound annual growth rate to 11 billion by 2030, with agentic use representing 30 percent of those queries. But it also acknowledges that enterprise adoption remains nascent: less than a quarter of enterprises are currently using agentic AI in any form, and those that are have not yet reached full autonomous deployment.

"The important point is that the adoption rates are still relatively low today, especially in small to medium-sized businesses," Schneider noted. Goldman's own forecast is that only 12 percent of knowledge workers will be using agentic AI by 2030, rising to 37 percent by 2040 — a long-tail adoption curve that does not obviously reconcile with a three-to-twelve month margin inflection.

Where the money actually goes

The concentration of AI economics at the semiconductor layer is the part of the story that most coverage of the Goldman report has glossed over. Nvidia, which manufactures the GPUs powering nearly all current AI workloads, accounts for approximately 75 percent of total AI compute spend — a figure cited in Goldman's own concurrent research on AI investment risk. At those volumes, Nvidia's gross margins sit at roughly 75 percent. The hyperscalers doing the spending are not operating at that margin profile. They are buying the tools.

Enterprise adoption data compounds the picture. Research cited in the bank's work suggests that 95 percent of organizations currently using AI pilot programs are deriving zero return from them — meaning the demand signal Goldman is modeling as the driver of future cash flow is, in the present tense, not there in measurable form.

"95 percent of organizations were getting zero return on their AI pilots," according to research referenced by Goldman analysts. "In other words, a big chunk of the enterprise market moved out of the window-shopping phase," but pilot-to-production conversion remains the primary bottleneck.

The timing question

Goldman Sachs Research expects the chip shortage to last twelve to eighteen months as semiconductor makers build out new fabrication capacity. The report's margin inflection timeline is three to twelve months. Those windows overlap but are not identical — and the margin inflection depends on the demand side, not just supply. If enterprise adoption follows the twelve-to-fourteen-year historical pattern the bank itself projects for knowledge worker uptake, the cash flow mathematics change considerably.

For now, the AI buildout continues at pace. Capex is being committed, debt is being issued, and the token consumption models are being published. The Goldman report's headline — AI agents will increase tech cash flow — may ultimately prove correct. Whether that cash flow lands at the companies writing the checks, or at the vendor they are writing them to, is a question the bank's own data does not fully answer.