AI Infrastructure ROI Circularity Risk: The $2 Trillion Math Problem

On May 28, 2026, Anthropic closed a funding round near a $1 trillion valuation on a roughly $47 billion revenue run rate, with sovereign wealth and pension capital leading. Two days of headlines pushed the bull case for AI compute harder than any earnings call could. Underneath, Bain’s 2025 Global Technology Report still says the industry needs roughly $2 trillion in new annual revenue by 2030 to justify the compute that hyperscalers are already buying. The AI infrastructure ROI circularity risk lives in the gap between those two numbers.

This is not the 1999 telecom story rerun. The mechanism is different, and the binding constraint is not chip supply. It is the credit profile of one bondholder cluster.

Key facts at a glance

• Bain projects roughly $500 billion per year of AI compute capex by 2030, requiring about $2 trillion per year of new revenue to justify it, with an approximate $800 billion residual shortfall even after every reasonable offset (Bain press release, Sept 2025).
• Google’s planned capex was cited at roughly $90 billion in 2025 and $180 billion in 2026 by Steve Eisman on his own podcast (The Real Eisman Playbook Ep 62).
• Anthropic ran from roughly $1 billion ARR at the start of 2025 to a $30 billion run rate by mid-year and approximately $47 billion by May 2026 (VentureBeat, Sacra profile).
• OpenAI does not project profitability before 2030 and projects multi-year losses through 2028, per internal financial documents reviewed by Fortune (Fortune, Nov 2025).
• Gartner forecasts worldwide IT spending to grow 9.8% in 2025, with software and data center systems leading the mix (Gartner, Jan 2025).

The bull case

The bull case is that the denominator critics use is too small. Bain defines the AI prize against the labor base, not the software base, because the product being sold is task automation. Wedbush’s Dan Ives frames 2026 as the third inning of a multi-year monetization wave, not a peak (The Motley Fool, May 2026).

Anthropic’s curve is the cleanest live test. An 80x climb in run rate inside fifteen months is what real total-addressable-market expansion looks like when it actually happens, not a marketing chart (VentureBeat, TrendingTopics). Anthropic CFO Krishna Rao has said gross margins are positive on a unit basis once training cost is excluded. Microsoft’s FY2025 disclosure also shows it has booked roughly $5.9 billion of OpenAI-related gains over nine months (Microsoft 8-K exhibit), evidence that at least one hyperscaler is taking real money out of the loop, not only putting capital in.

If two private labs can compound this fast, the Bain target sits four or five doublings away on a curve that is already compounding, which is plausible rather than impossible.

The bear case

The bear case is mechanical, not rhetorical. Bain’s arithmetic does not require a recession. Even in the optimistic scenario, with every promised AI productivity saving reinvested and every on-premises IT dollar shifted into cloud, AI revenue still falls roughly $800 billion short of what is required to service the compute bill (Bloomberg, Tom’s Hardware). Futurum estimates 2026 hyperscaler AI capex alone at roughly $690 billion across the five largest spenders (Futurum Group).

Demand-side data is uglier than the run-rate charts suggest. Gary Marcus, drawing on the MIT Media Lab and PwC’s 2026 CEO survey, points to 95% of enterprise generative AI pilots showing no measurable ROI and 56% of CEOs reporting nothing from their AI investments (Marcus on AI, Philipp Dubach analysis). Without enterprise revenue at scale, the gap closes only via consumer subscriptions and venture capital, neither of which has historically funded utility-scale capex.

Independent operator analysis also flags that hyperscaler AI segment economics are not separately disclosed, which makes it hard to verify whether the buildout is value-accretive or earnings-dilutive at the segment level (Om Malik).

What the data shows

Two rows are growing. Five are static or structural. The bull case needs the growing rows to keep compounding. The bear case only needs the static rows to remain static.

The Oracle thread

Marcus points at OpenAI as the domino that breaks the loop. The accounting points at Oracle. Oracle has taken on a very large slice of OpenAI compute commitments and priced its bond covenants against that revenue. If OpenAI growth merely decelerates inside the curve Oracle assumed, Oracle credit gets repriced before any equity in the loop does. That credit-event path is faster than a slow demand shortfall, which is why we treat it as the acute risk rather than the consensus “AI bubble” narrative.

What would change our mind

• Hyperscalers begin disclosing AI segment gross margin and utilization, not only aggregate cloud revenue.
• One or more enterprise AI deployments show audited, year-over-year ROI above 25%, replicated across at least three industries.
• Oracle and similar contract-heavy compute lessors begin requiring credit protection or co-investment on customer compute contracts.
• A non-venture, non-hyperscaler revenue pool (governments, pharma, defense) commits at the tens-of-billions level on multi-year terms.

If any two of these arrive together, the circularity argument weakens materially. Until then, the math is the math.

What this does not tell you

• We do not know what percentage of hyperscaler future cloud revenue is concentrated in OpenAI and Anthropic. That share is not disclosed in any 10-K we have seen, and public estimates are informed guesses rather than facts.
• We do not know the gas turbine, transformer, and grid interconnection lead times that will gate 2027 and 2028 capex being deployed, only that several operators report multi-year backlogs.
• We do not know whether sovereign and pension capital led into Anthropic at a near-trillion-dollar mark is sticky or chasing the same fear-of-missing-out cycle that funded prior late-stage rounds.
• We do not know the next major model release date or its capability step, both of which could move enterprise adoption sharply in either direction.

FAQ

Does this mean AI capex is a bubble?

Not necessarily. Bain’s framing is a math problem, not a verdict. It says the revenue required to justify the buildout has to arrive. If it does, the capex was rational. If it does not, the capex was early. Both outcomes are still on the table.

What is the single most important number to watch?

The audited share of hyperscaler cloud revenue attributable to AI workloads, broken out by segment. Until that disclosure exists, the rest of the debate runs on inference.

Is Anthropic’s growth proof the bull case is right?

It is evidence, not proof. Two private labs scaling fast does not validate a $2 trillion industry-wide revenue requirement, but it does raise the ceiling on what is plausible.

Why does Oracle matter more than OpenAI in the credit chain?

Because Oracle is the public-debt counterparty. OpenAI’s stress flows through Oracle’s bond covenants before it surfaces in any equity print.