Somewhere in northern Virginia, a substation hums at a frequency no one designed for. The grid there was built for homes and offices, not for the relentless arithmetic of transformer training. And yet this unremarkable piece of electrical infrastructure has become one of the most consequential financial instruments of our decade โ not because it yields, but because it absorbs.
I have spent the last four months watching a single number. Over the trailing year, the hyperscale cloud providers committed roughly two hundred billion dollars to AI-related capital expenditure. By some forecasts, that figure climbs past three hundred billion in the year ahead. Against this, the direct revenue generated by generative AI โ API calls, enterprise seats, cloud inference โ sits somewhere in the tens of billions. The scissors are open, and they are widening.
This is not, on its face, a crypto story. It is a story about capital โ where it goes, what it displaces, and who is left holding a ledger of obligations when the music slows. But it is a story that crypto readers should read more carefully than anyone else, because the liquidity they keep waiting for has quietly left the room.
To understand the squeeze, you have to understand what AI capital actually is. It is not venture money chasing a narrative, and it is not retail speculation chasing a chart. It is heavy capital โ the kind that pours into concrete, copper, and silicon, and does not come back quickly. When a hyperscaler commits a billion dollars, that money does not evaporate into salaries and software licenses. It becomes data centers, GPU clusters, substations, cooling towers โ assets with a three-to-five-year depreciation clock, negligible liquidity, and almost no secondary market at scale.
In the language I use when I audit governance contracts, this is a commitment device. Once written, it is extraordinarily difficult to unwind.
That physical rigidity is the technical premise most macro commentary skips. And it is precisely why the effect on sovereign finance is not a metaphor. Expected returns on American AI assets lift the entire dollar complex. Capital that once financed infrastructure in Jakarta or Sรฃo Paulo now finds a higher risk-adjusted return in a Virginia data campus. The mechanism is not mysterious; it is the old "dollar smile" wearing a new face. High expected returns pull capital home, the dollar strengthens, and economies that borrow in dollars feel the cost of that borrowing rise without any decision of their own.
Meanwhile, the same demand for capital that draws money inward also pushes real interest rates higher. Here the arithmetic turns uncomfortable for the very government hosting the boom. American federal debt has crossed thirty-six trillion dollars. Annual interest expense now approaches โ and in some quarters exceeds โ one trillion. Every basis point of rate pressure is not an abstraction; it is a line item. When private capital competes with sovereign borrowing for the same pool of savings, the sovereign does not always win. That is the literal meaning of "feeling the squeeze," and it is a strange kind of irony: the nation exporting the AI revolution is also financing it at a rising cost.
The academic literature calls America's ability to borrow cheaply in its own currency the "exorbitant privilege." The squeeze narrative suggests that privilege is fraying โ that AI's demand for capital is raising the very cost that privilege once suppressed. Both things can be true at once. The dollar can be the world's reserve and still feel the strain of financing a boom it did not fully choose. That duality is the honest center of this story, and it is the part that resists simple moralizing.
The crowding-out is not hypothetical. When the private sector's appetite for capital rises, the sovereign's cost of capital rises with it. A government rolling over trillions in debt is a price-taker in a market where AI is a price-maker. The consequence is a slow, structural tightening of fiscal space โ invisible quarter to quarter, decisive over a decade.
The scissors gap is the fault line. Capital expenditure is racing ahead of revenue at a ratio no mature industry would tolerate. I watched a similar pattern in 2020, when I retreated to a cabin outside Seattle to model the composability risks inside Yearn Finance's vaults. What I found then was a system in which each layer of leverage looked safe in isolation and catastrophic in aggregate. The AI capital cycle rhymes with it. Chipmakers invest in capacity. Cloud providers buy that capacity. Model labs rent it, funded by the same investors who hold the chipmakers. The money circulates, and each lap inflates the apparent scale of "investment" without necessarily creating an external dollar of demand.
Strip out the circular and related-party transactions, and the true external inflow is smaller โ and harder to see โ than the headline numbers suggest. I am not claiming fraud. I am claiming reflexivity. When I audited MakerDAO's early governance contracts in 2017, I found a stability fee calculation that threatened user solvency; the flaw was invisible precisely because everyone assumed the system was self-correcting. Markets rarely are.
There is a counter-argument, and I hold it honestly. If AI productivity gains materialize on the timelines some economists project, then today's capital expenditure is not a bubble but a rational pre-positioning โ building the roads before the traffic arrives. The heaviness of the assets, the very thing that makes the downside severe, is also the thing that makes the bet credible. Nobody builds a substation on a whim. So the honest answer to "bubble or investment?" is that we cannot yet tell, and anyone who claims certainty is selling something. The difference between a visionary and a gambler is settled only by the ledger of time.
The energy layer deserves its own accounting. AI data centers are projected to consume a share of national electricity that no grid planner anticipated even three years ago. In parts of Virginia and Ireland, interconnection queues have become the real bottleneck โ not capital, not chips, but power. This is the least reversible commitment of all. You can cancel a software subscription in a click; you cannot un-build a transmission line. The capital that flowed into AI is locked not merely in hardware but in geography and time, and that lock is the physical substrate of the financial risk everyone now senses but few can size.
What I can tell you is that this cycle is financed by capital that is mobile and expectations that are fragile. If the revenue curve fails to bend upward, the adjustment will not be gentle. Heavy assets do not exit gracefully; they are written down, and the write-down travels. The same concentration that makes AI the most productive frontier in technology also makes it the most systemically fragile โ a handful of firms, a handful of assets, a handful of grids. The IMF and the BIS have begun whispering about this concentration risk. They are right to whisper.
And here the story turns toward the periphery. The popular framing casts developing economies as pure victims โ capital drained, currencies pressured, growth suppressed. The truth is more textured. Supply-chain economies benefit: the memory fabs of Korea, the assembly floors of Southeast Asia, the advanced packaging in Taiwan. Sovereign funds in the Gulf are not being robbed; they are actively buying into American AI, converting oil into compute. The damage falls unevenly, on economies that depend on foreign capital inflows and borrow in hard currency. That is not a story of the Global South against the North. It is a story of differentiation โ and differentiation is where the mispricing lives.
We should name the beneficiaries plainly, because a ledger that omits them is not a ledger but an argument. The chip designers, the cloud operators, the model labs, and the supply-chain economies they depend on are all, for now, on the winning side. The article that prompted this analysis focused on the losers, and that focus is not wrong โ it is simply partial. A complete account must hold both, because the reallocation is not zero-sum. It is a transfer, and transfers have two ends.
I keep returning to a lesson from 2021, when I helped three Indigenous artists launch a deliberately non-speculative collection on Tezos. We raised fifteen thousand dollars and built something that outlasted any yield farm I have ever seen. The lesson was not that small is virtuous. The lesson was that value which is real behaves differently from value which is merely priced. We minted souls, not just tokens. AI capital, for all its magnitude, has not yet proven which of the two it is holding.
There is a second layer here that financial commentary misses entirely, and it is the one I care about most. If AI systems are to be trusted with the machinery of finance, identity, and governance, then their accountability must be verifiable โ not asserted. In 2026 I worked with a small team of ethicists and engineers to design a decentralized identity framework for AI agents on Polkadot, using zero-knowledge proofs to attest ethical compliance without exposing sensitive data. The point was never to surveil the machine. The point was to make its commitments checkable by anyone. The same principle applies to capital. If AI investment is to be trusted with the world's savings, its claims must be auditable โ capex reconciled against revenue, circular flows disclosed, concentration measured. Truth emerges when the ledger is transparent. Right now, much of the AI capital ledger is opaque, and opacity is where fragility hides.
This is where my own discipline offers a warning rather than a prediction. Decentralized finance taught us that composability is a gift and a weapon. It lets value flow freely, and it lets failure propagate at the speed of a block. AI capital is composable too โ chip to cloud to lab to chip โ and it is uninsured in the way that matters. When I wrote "The Silence After the Crash" in the aftermath of LUNA, my argument was simple: decentralization without accountability is not freedom, it is anarchy. The same sentence applies to a capital cycle that no single actor fully sees and no single actor fully controls.
For the crypto market, the implication is quieter than a headline and more consequential than a drawdown. The marginal dollar that once sought yield in a lending pool now seeks yield in an AI equity. This is not theft; it is competition, and it is being won by the asset class with the clearer story of the future. The tokens that survive this reallocation will be the ones that offer something the AI complex cannot โ verifiable scarcity, permissionless settlement, and a community that owns its own ledger rather than renting exposure to someone else's.
So the contrarian point is this. The article that prompted these notes frames AI investment as a homogeneous macro tide, a single force pulling global cash toward America and squeezing everyone else. That framing is useful, and it is also misleading. AI capital is not one thing; it is at least two โ the durable bet on capability, and the reflexive spiral of related-party enthusiasm. Conflating them is how you miss both the opportunity and the risk. And when a crypto publication reports on AI draining global liquidity, remember whose anxieties are being articulated. The crypto reader is not a neutral observer of this story. The crypto reader is one of its characters.
I do not know whether the scissors will close through revenue or through a write-down. I know only that heavy assets, mobile capital, and concentrated expectations form an unstable triad, and that the instability will express itself first at the edges โ in the currencies of nations that never chose this cycle, in the grids that never agreed to carry it, in the liquidity pools that quietly emptied while everyone watched the frontier.
In the chaos of DeFi, I found my silence. I find it again here, watching a substation in Virginia hum for a future that has not yet been invoiced. The question is not whether AI will reshape the global ledger. The question is whether we will build the accountability into it before the ledger builds the lesson for us.