Cathie Wood said 90% of data center financing sits in the United States. That is the whole claim. No denominator. No timestamp. No distinction between equity, private credit, and asset-backed issuance. No split between training clusters and inference capacity. The number arrived without a methodology and circulated without a rebuttal. By the time it reached crypto Twitter, it had hardened into fact.
I have spent nineteen years watching numbers like this. In late 2017, I bypassed the ICO frenzy and spent six weeks inside the 0x v0.9.9 exchange protocol, tracing three unsigned integer overflows in fillOrder. The marketing deck said one thing. The compiled contract said another. That gap โ between what a system claims and what it executes โ is where I work. The 90% figure has the same gap, except there is no contract to decompile. Reversing the stack to find the original intent, I find no stack at all. I find a sentence.
So the question is not whether 90% is high or low. The question is what the number measures, who benefits from its ambiguity, and which failure modes it conceals. Those are answerable. The headline is not.
What "Financing" Actually Compiles To
Data center capital formation is not one market. It is at least four, and they do not share a risk profile.
The first is hyperscaler cash flow. Microsoft, Google, Amazon, and Meta fund capacity from operating earnings. This is the cleanest capital โ no leverage, no refinancing risk, no covenant. The second is private credit and structured debt: securitizations, joint ventures, and project finance that move the liability off the sponsor's balance sheet and into a capital market. The third is vendor and sovereign co-investment, where chip suppliers and state funds take equity to secure allocation. The fourth is the crypto-native layer โ tokenized revenue, yield wrappers, and DePIN compute networks โ the newest, and the least audited.
When a single number collapses these four into "90% in the US," it destroys the only information that matters: the leverage ratio. Cash flow is not debt. A project funded from retained earnings fails gracefully. A project funded at 5x leverage against optimistic depreciation schedules fails by margin call. The 90% claim cannot distinguish the two.
This matters more now than a year ago. The financing mix is rotating. Earlier capacity was funded from the balance sheet. Newer capacity is funded from capital markets โ private credit, ABS, and JV structures that did not exist at this scale three years ago. The risk has not disappeared. It has been transferred from the companies that build the data centers to the investors who lend against them. In crypto we have a name for that transfer. We call it a yield product.
What would verification require? A ledger. A public, timestamped, categorized record of every instrument โ equity, debt, ABS, JV, tokenized wrapper โ with its jurisdiction and its base. No such ledger exists. The claim rests on private data assembled by a fund with a stated market interest. If a number cannot be traced to a source you can independently re-derive, it is not data. It is a press release. This is the standard I apply to smart contracts, and it should apply to market statistics. A claim that cannot be reproduced is not evidence. It is narrative with a decimal point.
The Caliber Problem
Here is the part the source never resolves. "90% of financing in the US" is a statistic without a base. Change the base and the number moves.
If the base is market-based financing โ venture, private credit, ABS, public equity โ US dominance is almost tautological. America has the deepest capital markets on earth. China funds compute through state-directed vehicles, telecom operators, and the "East Data West Computing" program. Those flows never touch a private credit fund. They are invisible to the denominator. So the 90% measures the institutional plumbing, not the total capital. The number is a statement about how America funds things, not about how much compute the world is building.
If the base is total compute investment โ measured in dollars or FLOPs โ the picture changes. China's state model deploys capital through different channels at different speeds. The US leads on capital mobilization. It does not obviously lead on build speed, land approval, or grid interconnection. Those variables determine when a GPU actually runs.
I traced a version of this trap in 2021, when I analyzed the ERC-721 metadata standard and found 40% of popular collections resolving to centralized IPFS nodes. "On-chain ownership" was an illusion. The token was on-chain. The image was not. The metadata was a lie until proven immutable. The 90% figure has the same structure: a claim that sounds like an on-chain fact but resolves to an off-chain assertion. Nobody has published the ledger. And in a market where capital is being tokenized, that absence is not neutral. It is the whole risk.
Circular Financing and the Depreciation Assumption
Now the failure mode the headline conceals most effectively.
AI infrastructure financing is rotating from cash flow toward debt, and it is doing so inside a loop. Chip vendors invest in cloud providers. Cloud providers buy chips. Cloud providers raise debt against future capacity. That capacity is pre-sold to AI labs, which are financed by the same investors. This is circular financing, and it inflates the demand signal.
I reverse-engineered this pattern once already. In 2022 I spent four weeks inside the LUNA/UST loop, mapping the exact block where the peg-breaking feedback became mathematically irreversible. The mechanism was not fraud. It was a self-referential incentive structure that looked stable until it could not be. The seigniorage model had a fixed point, and the fixed point was zero.
The AI data center loop has the same topology. Different assets, different counterparties, same self-reference. When the buyer, the seller, and the financier are the same three entities, the price is not a discovery. It is an echo.
Underneath the loop sits a single accounting assumption: GPU economic life. Most models assume three to six years. That number sets depreciation, which sets reported profit, which sets the equity value that supports the next debt issuance. If technical iteration outpaces the schedule โ and every accelerator generation argues that it does โ then depreciation is understated and profit is overstated. The correction is not a write-down. It is a repricing of the entire capital stack.
This is why the tax incentive matters more than the headline. American data centers enjoy sales tax exemptions, property tax abatements, and accelerated bonus depreciation. These levers set the after-tax IRR that makes a marginal project financeable. Remove them and the marginal project dies. The 90% concentration is not a sign of strength. It is a sign that the returns are policy-constructed.
Financing is also the wrong bottleneck to fixate on. Capital is fungible and fast; electrons and transformers are not. Grid interconnection queues in the major US data center corridors now run into multi-year waits, and high-voltage transformer lead times stretch to two years or more. A project can close its financing in a quarter and still wait three years for a power connection. This is the constraint that makes the 90% number almost irrelevant to actual compute delivery. Capital arrives on a spreadsheet's schedule. Power arrives on physics' schedule. The two do not synchronize.
The Tokenized Wrapper
The newest layer deserves its own line item. As data center revenue becomes collateral, it becomes securitizable, and as it becomes securitizable, it becomes wrappable. Tokenized yield products backed by AI infrastructure cash flows are already being structured, and their retail distribution channel is crypto. The structure repeats a pattern I have documented since 2021: an off-chain cash flow, an on-chain claim on that cash flow, and a gap between the two that only closes when the cash flow stops. The wrapper does not create risk. It repackages risk and sells it to the last buyer in the queue.
The DePIN Counter-Claim
There is a second-order question the source ignores entirely. If financing is this concentrated and this policy-dependent, what prices the alternative?
Decentralized physical infrastructure networks โ DePIN compute markets โ sell the opposite proposition: distributed, permissionless, and priced in tokens rather than in tax-advantaged project finance. On paper they are the natural hedge against concentration risk. In practice they are a rounding error against hyperscaler capex, and their token prices trade as a leveraged bet on the same AI narrative. When centralized capacity corrects, DePIN does not decouple. It amplifies. I audited a verifiable-compute protocol in 2026 and found a gas optimization that cut verification cost by 40%. The cryptography worked. The economics still inherited the same demand assumptions as the centralized stack. Decentralization changes the ledger, not the beta.
The Blind Spot
The consensus reading of the 90% figure is bullish: America owns AI infrastructure. The contrarian reading is that the number describes risk concentration, not advantage.
Every dollar of that 90% is exposed to the same four variables โ interest rates, tax policy, grid capacity, and the depreciation schedule. A diversified global build would spread those exposures. A concentrated one amplifies them. When the correction comes, it does not diffuse slowly. It arrives as a correlated unwind across a single capital market, in a single jurisdiction, holding a single asset class.
The crypto-native layer makes this worse, not better. Data center revenue is being wrapped into yield products and tokenized structures, the same way stablecoin yield products wrapped maturity mismatches in 2021. Those products worked in the up-cycle. They were engineered to fail first in the down-cycle. The wrapper is always the last to be told and the first to break.
The other blind spot is that nobody has verified the denominator. The 90% figure comes from a single source with a stated directional interest. Cathie Wood is a structural AI bull. Her funds are bound to the AI narrative. That does not make her wrong. It makes her unverified. Truth is not consensus; truth is verifiable code โ and there is no code here, only a claim.
Takeaway
Watch the tax legislation, not the headline. Watch the depreciation schedules buried in the debt issuance, not the capex guidance. Watch grid interconnection queues, because financing that cannot buy a power connection buys nothing.
The 90% number will keep circulating. The narrower question is when the loop unwinds, which layer breaks first โ the equity, the debt, or the tokenized wrapper sold to retail? My money is on the wrapper. It always is.