Chanos: AI Capex Returns Have Been Cut in Half, Could Fall Further to 10%
Taylor Wilson
Short-seller Jim Chanos warns that hyperscalers' incremental return on invested capital has fallen from 40% to roughly 20%, and could slide to 10% if spending growth continues — at which point management faces a blunt question: keep building data centers, or just buy Treasuries.
Returns halved — how bad is the burn?
Chanos cites data on Google, Meta, Amazon, Microsoft, and Oracle: their incremental ROIC — how much profit each extra dollar of investment generates — sat at roughly 40% about eighteen months ago. It has since dropped to about 20%.
This means → the same capital deployed today earns half what it did a year and a half ago. If spending growth holds, that figure could fall to 10%.
He expects the reckoning to hit between late 2026 and 2027. In plain terms = at that point, management must answer shareholders: why keep pouring hundreds of billions into data centers when the return is worse than buying Treasuries?
Oracle is singled out — its AI-infrastructure incremental ROIC is the lowest among peers, and its stock has fallen 65–70% from its peak.
Why is this worse than the dot-com bubble?
Chanos draws a direct comparison to the late-1990s telecom infrastructure bubble. His conclusion: this time is worse.
From 1998 to 2002, competitive local exchange carriers and fiber-optic builders spent a combined ~$100 billion over five years, roughly $20 billion a year. In the current AI cycle, a single company's annual data-center and AI-infrastructure spend already exceeds that five-year total.
The funding structure is also different. During the dot-com bubble, the enterprise customers writing the checks remained profitable — they simply cut orders. This time, the hyperscalers are the only profitable spenders; nearly every other participant in the ecosystem runs on venture capital or high-leverage financing.
This reflects a deeper fragility — once these few profitable giants hit the brakes, there is no second payer anywhere in the chain.
What does the depreciation accounting hide?
Chanos flags a widely overlooked accounting issue: large volumes of purchased but not-yet-deployed GPUs and data-center equipment sit on balance sheets under "construction in progress," and depreciation has not yet started.
In plain terms = these assets are losing economic and technological value from the moment they are bought — newer chips arrive, performance benchmarks shift — but because they are not yet "live," none of that loss shows up on the income statement.
The gap between purchase and deployment runs roughly 18 months. Even with a nominal useful life of five to six years, the effective amortization period stretches to 6.5–7.5 years. Chanos's own model uses 10-year depreciation, and "even then, we still can't make the economics of most data centers work."
This also distorts the macro picture: S&P 500 earnings-growth expectations sit at 28% this year, far above the long-run historical average of roughly 6% per year. One reason is that one party's massive capex shows up as revenue on another party's income statement, while the spender capitalizes the cost and keeps it off current expenses.
What does Neocloud's "asset-light pivot" really tell us?
Chanos calls out Nebius by name. The company recently announced a shift to an "asset-light" model — no longer owning data centers and GPUs, instead becoming a "compute management services" provider (like a hotel franchise operator), with capex borne by third parties and Nebius collecting a management fee.
He calls this a "pretty significant about-face." For two years, Neocloud companies claimed heavy assets were their core competitive advantage. Now one of the largest players says "we don't need to own these assets anymore."
This reflects deeper pressure: both legacy and new data-center companies realize that labor and equipment shortages are driving construction costs sharply higher, and future maintenance capex will far exceed the figures previously presented to investors — "so they're rushing to offload assets."
He also notes that several deals around compute leasing at SpaceX's xAI data center — including agreements of roughly $1 billion each with Anthropic and Google — carry extremely short exit clauses, as brief as three months. Chanos views these deals as "having a strong promotional element."
Why is the interest rate the biggest hidden bomb?
In Chanos's framework, interest-rate risk is the most underestimated systemic threat in the entire AI-infrastructure bubble.
Data-center assets are currently trading at capitalization rates of 5–7%, while the U.S. 10-year Treasury yield sits at roughly 4.5% — a razor-thin spread. Sponsors layer leverage and aggressive mezzanine financing — high-risk debt that sits between equity and senior loans — to promise equity investors returns of 15%.
This means → the entire return structure rests on the assumption that rates do not rise. "If rates hit 6% or 7%, this all collapses. It will detonate one asset class after another."
He observes that CCC-rated junk-bond spreads — the lowest-quality tier — have already begun widening, but the BBB-to-BB range has not followed. In plain terms = credit markets are not panicking yet, but once rates visibly approach 5%, spread widening will be the key early-warning signal.
How does Chanos sum up the market right now?
He captures the pricing logic in one sentence: "In a bull market, people pay a premium for 'promises.' In a bear market, they only pay a discount for 'reality.' We are clearly in the former."
This means → current valuations are anchored to "the AI future that will be delivered," not "the AI present that has been delivered."
The core watchpoint, he argues, is hyperscaler capex decisions — whether this thesis is validated by late 2026 depends on whether these giants begin to pull back.
Content is for reference only, not financial advice.