The Number That Did All the Work
There is a figure making the rounds that has already accomplished more than any model release, any benchmark, any keynote in the last eighteen months. Ten point three trillion dollars. That is the sum the Brookings Institution is reported to have projected for cumulative global AI infrastructure investment through 2032, and it arrived in my feed — as it arrived in yours — by way of Crypto Briefing, a channel built for people who trade risk assets, not for people who read think-tank methodology appendices.
That detail matters more than the number itself. When a macro forecast crosses from the domain of policy research into the domain of high-beta retail capital, it has stopped being a forecast. It has become an input. And when an input is a nine-figure-plus abstraction with no visible denominator, the first thing a forensic analyst should do is not celebrate it. The first thing is to ask what it is measured in, who is measuring it, and why this particular audience is hearing it now.
The trap isn't that the number is wrong. The trap is that the number is unfalsifiable in its current form — and unfalsifiable numbers are the most dangerous kind in a capital cycle, because they cannot be disproven until the money has already been spent.
I have watched this movie before. In 2017 I sat in a co-working space in Buenos Aires and audited the tokenomics of more than fifty ICO whitepapers, and the single most reliable signal of an impending collapse was not the quality of the code or the pedigree of the team. It was the presence of a large, round, unsourced number in the first three paragraphs. Fifty billion in addressable market. Ten million users by 2020. Numbers that big don't inform. They recruit.
So before we accept $10.3 trillion as a description of the future, let's treat it as what it actually is: a rhetorical instrument, a supply-side capital narrative, and — if the reflexivity holds — a self-fulfilling script that may end up building the very overcapacity it claims to predict.
Context: Where the Number Sits in the Global Liquidity Map
Let me place this figure on the map before I start cutting into it.
The most important thing to understand about the $10.3 trillion projection is that it is not an outlier. It sits at the upper edge of a cluster of institutional estimates that have been inflating in lockstep over the past twenty-four months. McKinsey, in 2024, put global data center capital investment needs at roughly $6.7 trillion by 2030, with AI-related spend accounting for around $5.2 trillion of that. Morgan Stanley's numbers have clustered in the $3 trillion range. Brookings, at $10.3 trillion through 2032, is not a maverick — it is the top of a bell curve that is itself migrating to the right.
Read that migration carefully. It is not that the underlying physics changed. It is that the consensus has been repeatedly revised upward by the same mechanism that revises every consensus upward in a capital boom: each new estimate is anchored to the last one, plus a confidence premium. The estimate is not a measurement of demand. It is a measurement of conviction.
Now overlay the liquidity map. This projection was published into a specific monetary environment. The Fed has been navigating the late stage of a tightening cycle, with the question of rate normalization still unresolved as of mid-2025. Global M2 has been oscillating rather than expanding. The cost of capital, in real terms, is meaningfully higher than it was during the 2020-2021 buildout of everything. And yet the single largest capital allocation program in the history of industrial civilization is supposedly about to be financed inside that environment.
That is not impossible. It is, however, an extraordinary claim, and extraordinary claims financed by other people's money deserve extraordinary scrutiny.
Here is the piece of plumbing that almost no one reports. AI infrastructure is being financed increasingly through structures that hide their true leverage. Special purpose vehicles. Joint ventures. Data center asset-backed securities. Private credit funds stepping in where banks have stepped back. GPU-collateralized lending. And the most elegant circularity of all: chip vendors taking equity stakes in their own customers, who then use the proceeds to buy more chips. Each of these structures is legitimate in isolation. Stacked together, they convert what looks like a technology capex cycle into something that behaves, on the balance sheets of the non-bank financial system, much more like a real estate cycle.
This is why the phrase "financial complexity" appears in the reporting, and why it is the only sentence in the entire coverage that carries any warning weight. It is the think-tank equivalent of a doctor saying, quietly, "we should keep an eye on that."
And that, finally, is why this story reached a crypto audience. Mining operators pivoting to HPC hosting. Tokenized compute markets. Decentralized GPU networks pitching themselves as the cost-efficient alternative to hyperscaler rent. The narrative being sold to that audience is a specific one: that the AI buildout will overflow the walls of the hyperscalers and spill into decentralized infrastructure, and that early positioning captures the spillover.
That narrative may be true. But note what it requires — it requires the centralized buildout to succeed enough to create the demand, and to fail enough to leave room. I have spent a decade watching crypto investors get seduced by exactly this kind of "both and" thesis. They rarely survive contact with a capital cycle.
Core: The Arithmetic of a Promise
The single most useful thing I can do with $10.3 trillion is divide it. Numbers this large hide their teeth until you break them apart.
Let me assume, for the sake of argument, that the figure represents cumulative capital expenditure spanning roughly 2025 through 2032 — seven years. That is the most benign reading, and I want to give the projection every benefit of the doubt before I strip it down.
At seven years and $10.3 trillion, you are looking at an average of roughly $1.47 trillion per year in capital deployment. Now apply straight-line depreciation, which is the crude but honest way infrastructure accountants think. That deprecation charge alone lands at about $1.47 trillion annually. To cover the cost of capital on that asset base — and let's be generous and assume a weighted average cost of capital of eight to ten percent, which is what a mature utility-grade asset should demand — you need annual revenue in the range of $2 trillion to $2.5 trillion just to break even on the capital, before we even talk about profit.
Now put that next to reality. The entire global cloud computing market — AWS, Azure, Google Cloud, Oracle, everyone — generated somewhere in the neighborhood of $600 to $700 billion in 2024. The largest SaaS companies on earth report annual revenues in the tens of billions, not hundreds. The total enterprise software market is roughly $600 billion to $700 billion a year.
So the required revenue pool to justify $10.3 trillion is roughly three to four times the size of the entire existing cloud and enterprise software market combined. And this is the most generous reading of the projection. If $10.3 trillion is instead a 2032 single-year figure — which the coverage does not clarify — then the required compound annual growth rate from a 2024 base of roughly $230 billion in hyperscaler capex would need to run near fifty percent annually for eight consecutive years. There is no precedent for that in the history of commercial capital formation. Not rail. Not telecom. Not semiconductors.
This is the gap nobody in the headlines wants to name. The investment-commitment side of the ledger is quantified to the trillion, and the revenue-verification side of the ledger is quantified to nothing at all. That asymmetry is not an accident. It is the defining feature of a supply-side narrative.
The CAGR test tells the same story. If you take the combined capex of Microsoft, Google, Amazon, and Meta — roughly $230 billion in 2024, with 2025 guidance clustering around $320 billion — and you ask what it takes to get from there to $10.3 trillion cumulative by 2032, you get a path that is optimistic but not insane: something like a sustained twenty-five percent annual growth rate. That is a real, walkable road. It is also the kind of road that has a sharp turn at the end of it, because capital intensity like this compresses margins, and compressing margins is the thing that ends capex cycles.
I learned to run this exact test in the spring of 2020, when I sat down to model the incentive structures of Compound and Aave during what the market was still calling DeFi Summer. The math that mattered was never the headline yield. It was the question of where the yield came from. When I decomposed those farms, the answer was consistent: a large fraction of the advertised return was a claim on future token issuance, discounted into the present. It was borrowed from tomorrow's balance sheet, dressed as today's income.
I am not saying AI infrastructure is a Ponzi. I am saying the accounting instinct generalizes. Whenever you see a large investment figure with no matching revenue figure, the correct analytical move is to ask which future period the value has been borrowed from. In DeFi in 2020, the borrowing was from token emissions. In AI infrastructure in 2025, the borrowing is from an application-layer revenue curve that does not yet exist at the required scale.
Let me be precise about the shape of that curve, because it determines everything. For the $10.3 trillion to be rational, you need an application layer that generates somewhere between two and two and a half trillion dollars in annual revenue by the early 2030s. As of 2025, the entire global market for AI-native applications is well under one hundred billion dollars. That is not a rounding error in the gap. That is a gap of more than an order of magnitude, with a seven-year clock running on it.
The historical precedent here is uncomfortable for optimists. In the late 1990s, telecom operators laid fiber at a rate that assumed internet traffic would compound at roughly the pace of the dot-com equity indices. The fiber was real. The demand curves were not. The result was one of the largest asset impairments in corporate history — and, crucially, that impairment was not the end of the internet. It was the beginning of cheap internet. The builders got wiped out. The users got a windfall. I will come back to this, because it is the most under-priced scenario in the entire debate.
The Power Constraint: Where the Arithmetic Meets Physics
Here is where I break ranks with the capital-cycle crowd and join the electrical engineers.
Every dollar in that $10.3 trillion ultimately resolves into one of five physical inputs: compute silicon, servers, buildings and mechanical systems, power delivery, and network interconnect. The reasonable split — and this is my own decomposition, since the projection provides none — runs roughly forty to fifty percent into compute hardware, twenty to twenty-five percent into data center construction and mechanical and electrical systems, fifteen to twenty percent into power generation and grid connection, and the remainder into networking.
Take the compute line and push it. Forty-five percent of $10.3 trillion is about $4.6 trillion. At current AI server rack economics — call it three to four million dollars per rack including networking, a number that will drift with scale — you are looking at a deployment on the order of fifteen to twenty million racks. For context, the aggregate installed base of AI accelerators as of 2024 is measured in the low millions of units. We are being asked to believe in an order-of-magnitude deployment expansion inside eight years.
Silicon can scale. Taiwan Semiconductor can build fabs. HBM suppliers can add lines. Advanced packaging can expand, expensively. The supply chain will creak, but it has demonstrated an ability to respond to price signals that would make a commodities trader weep with joy.
Power cannot scale the same way. And this is the constraint that will kill more of the $10.3 trillion than any demand shortfall.
Global data center electricity consumption is estimated to have run around 460 terawatt-hours in 2024. The trajectory most analysts are modeling puts that somewhere between 1,000 and 1,300 terawatt-hours by 2030. To be clear about what that means: you are adding the equivalent of a mid-sized European country's electricity demand to a grid that, in the United States, already has interconnection queues of three to seven years in most major markets. In some regions it is longer. In some it is effectively closed.
Here is the sentence that should be printed on every AI infrastructure pitch deck: the interconnect queue is longer than the chip delivery lead time. You can buy the GPUs. You can build the shell. You cannot conjure the electrons on the schedule the financial model assumes, because the binding constraint is not money — it is transformers, transmission lines, turbine manufacturing slots, and the permitting process, and none of those respond to a term sheet.
I have watched this exact pattern before, in a different asset class. When I mapped the Terra/Luna collapse in 2022, the failure looked like a stablecoin failure. It was not. It was a liquidity failure dressed as a stablecoin failure. The algorithmic mechanism was the visible symptom; the actual cause was that the system had an implicit dependency — on sustained capital inflow and on shallow secondary liquidity — that could not survive a regime change in the macro environment. When the Fed tightened, the whole interconnected layer of leverage that had grown around Terra went from asset to liability in roughly seventy-two hours, and the margin calls cascaded across centralized venues that had no idea they were correlated.
The AI infrastructure complex has the same structural vulnerability, in physical rather than financial form. The buildout has an implicit dependency — on power being available on demand — that the financial models treat as infinite. The trap isn't the capital. The trap is the assumption that energy is a variable cost rather than a hard ceiling.
And notice the geographic consequence, because it is where the macro story quietly becomes a geopolitics story. Data centers used to be sited near users for latency. They are increasingly being sited near power. The Texas grid. The Nordics. The Gulf. Southeast Asia. This is not a technology decision. It is a resources decision, and it turns compute into the newest form of a commodity that must be located where the fuel is. Which is a very old story. It is the story of aluminum smelting, of oil refining, of steel. Compute is joining the heavy industries.
The Chip Layer and the Cost Curve
There is a counterargument, and it is a serious one. Unit compute costs fall. They have fallen for fifty years. If the cost per unit of inference declines fast enough, the total investment required to reach a given capability level could fall even as the capability ceiling rises. In that world, the $10.3 trillion forecast could prove too high and too pessimistic at the same time.
I take this seriously, but I do not think it saves the projection, for a reason that has nothing to do with physics and everything to do with psychology. The forecast is not a demand model. It is a competitive expenditure model. In an arms-race dynamic, the participants do not stop spending when the unit cost falls. They spend the savings on more units. Jevons' paradox applies to compute every bit as much as it applied to coal in the nineteenth century. Cheaper chips do not reduce the bill. They raise the ambition.
That is why the supply-chain forecast implied by the projection is so aggressive. If silicon is roughly half the spend, and if Nvidia retains a fifty to seventy percent share of accelerators through the period, the company's data center revenue would need to move from roughly $115 billion in 2024 to a multi-hundred-billion annual run rate by the early 2030s. The bottlenecks concentrate upstream: CoWoS advanced packaging capacity, HBM supply from a three-player oligopoly, and leading-edge foundry allocation at 3nm and below. Each of those is capital-intensive to expand, and each expansion takes years, not quarters.
This is also where I have to fold in a lesson I learned the hard way in 2024. When the spot Bitcoin ETFs launched, I built a model comparing BlackRock's IBIT net inflows against Fidelity's FBTC on a weekly cadence, cross-referenced against on-chain reserve changes. What the model told me — against a market consensus of an immediate parabolic rally — was that the supply shock would be slow and structural, unfolding over roughly eighteen months, and that the price path would look more like consolidation than ignition. That call was right, and it was right for a reason that generalizes: institutional capital does not arrive as a wave. It arrives as a drip, and it rebalances on a schedule that has nothing to do with retail excitement.
Apply the same lens here. The $10.3 trillion, if it materializes at all, will not arrive linearly. It will front-load into the build phase and back-load into the impairment phase, and the impairment phase is where the interesting money gets made — or lost, depending on which side of the ledger you sit.
Financing Structure: The Hidden Denominator
The projection does not tell us who pays. That omission is more consequential than any number in the headline.
If the buildout is financed primarily from hyperscaler operating cash flow — the Microsofts, the Googles, the Amazons, the Metas — then the cycle is largely self-funded and the systemic risk is contained. These are among the most cash-generative enterprises in commercial history, and they can absorb a great deal of capital misallocation before anyone's solvency is at stake.
But the coverage hints at something else. The phrase "financial complexity" points toward structured finance, and structured finance points toward the non-bank financial system. Over the last two years, data center financing has migrated toward SPVs, joint ventures with infrastructure funds, asset-backed securities collateralized by future compute contracts, and private credit lenders who have moved aggressively into the space. GPU-collateralized debt has appeared, with depreciation assumptions — typically three to six years for silicon against twenty-plus years for the shell — doing an enormous amount of unexamined work.
The depreciation mismatch is the fuse. If GPUs have an economic life closer to three years than six, then the collateral underlying a meaningful slice of this financing is worth a fraction of its stated value, and the impairment lands on lenders who assumed otherwise. This is not speculation about intent. It is arithmetic about duration, and it is exactly the kind of mismatch that only becomes visible when the cycle turns.
I remember running this decomposition during the DeFi liquidity trap analysis in 2020. The advertised yield on those farms assumed that the incentive tokens retained value over the farming period. When they did not, the yield collapsed, and the collapse did not announce itself. It simply appeared, one block at a time, until the position that looked profitable was underwater. Structured finance works the same way. The assumptions sit quietly in the model until the regime changes, and then they are all repriced at once.
The Contrarian Angle: Reflexivity and the Builders-Get-Wiped Scenario
Now I want to argue the part that makes me unpopular.
The single greatest risk from the $10.3 trillion forecast is not that it is wrong. It is that it is right enough to be believed, and believed enough to be loaded onto capital structures that cannot survive the transition from build to operate.
This is the reflexivity problem, and it is the most important thing in this entire essay. When a respected non-profit research institution publishes a projection of this magnitude, and when that projection is then amplified through channels with the highest risk appetite in the market, the projection does two things simultaneously. It describes a possible future, and it changes the probability of that future via the capital it mobilizes. Third-party forecasts are not neutral observations in a capital cycle. They are catalysts inside one.
The 2000 telecom cycle demonstrates this with brutal clarity. Forecasts of internet traffic growth — themselves reasonable — were used to justify fiber deployment at a rate that assumed the equity market's rate of return as a physical law. The fiber got built. The returns did not arrive on schedule. The builders were vaporized. And then, a decade later, the users of that fiber got the cheapest bandwidth in history, which is how the modern internet scaled.
I expect the same shape here. The builders of AI infrastructure at peak valuation will absorb the impairment. The application layer that follows — built on cheap, abundant, already-depreciated compute — will capture the surplus. The path of maximum pain is not "AI fails." It is "AI succeeds, but on the wrong balance sheets."
This is the decoupling thesis I want to put on the table, and it is the thing the coverage never mentions. The narrative being sold to the crypto audience assumes a smooth continuum: buildout creates demand, demand overflows central infrastructure, decentralized compute picks up the slack at better cost. But that continuum is a fantasy of the transition. What actually tends to happen in a capital cycle is a violent discontinuity between the build phase and the operate phase. The operators who own capacity at the peak of the build are the ones most exposed to the repricing. The operators who buy distressed capacity after the repricing are the ones who win.
If you want to find where this thesis is already visible, look at the decentralized compute networks — Render, Fetch.ai, and their peers — that I began mapping in earnest through 2026. My working hypothesis then was that a web3 compute market could undercut centralized cloud on cost, because it taps idle capacity rather than building dedicated capacity. I still think that is directionally right, but the nuance I missed initially is timing. Decentralized compute is not a competitor to the hyperscaler buildout. It is a scavenger of the hyperscaler buildout. It thrives on the surplus, and it cannot exist at scale until the surplus arrives. That means the decentralized compute thesis is actually a bet on the impairment of the centralized buildout, not on its success. Anyone positioning as if the two are allies is misreading the mechanism.
The same inversion applies to the DePIN and mining-adjacent pitch. The economically interesting event is not the buildout. It is the moment when GPU rental prices start to soften, when utilization rates fall below the threshold that services debt, and when distressed operators discover that their depreciation schedule was optimistic. That moment is the entry point, not the peak.
There is one more inversion worth naming. In an arms race, the competitive moat shifts from the part of the stack that can be copied to the part that cannot. Through 2023 and 2024, the moat was model capability — and model capability diffuses. It gets replicated, open-sourced, distilled, and commoditized within quarters. Through 2025 and beyond, the moat is shifting to capital, energy, and supply-chain control. Those are much harder to copy. But here is the twist: the difficulty of copying them is precisely what invites overinvestment. When the barrier to entry is a trillion dollars, everybody with a trillion dollars shows up, and the barrier protects no one.
I will note, as an aside that ties directly to my own view on Layer 2 economics, that this is the same failure mode that has dogged ZK Rollup scaling. The proving cost curve looked like a problem of scale, and it was assumed that volume would drive marginal cost toward zero. It hasn't. Operators are bleeding on proving costs that only make sense if gas returns to bull-market levels. The assumption of future abundance as the justification for present expenditure is a pattern, not an exception, and it shows up in every infrastructure buildout that mistakes a technical capability for a commercial model.
Which brings me to the piece nobody wants to discuss: the possibility that the forecast is not just aggressive but physically impossible to execute. If the power interconnection queue is the binding constraint, then the $10.3 trillion cannot be deployed on the timeline given, regardless of how much capital wants to move. The forecast could fail not because demand disappointed but because the electrons did not show up. In that scenario, the ultimate irony is that the deployment shortfall converges with the demand shortfall, and the industry arrives at the impairment simultaneously from both sides.
Chaos is just data that hasn't been priced yet. And in a capital cycle, the most valuable data is always the constraint that everyone assumed was a footnote.
Takeaway: Positioning for the Build, the Bill, and the Bargain
I am not a bear on AI. I am a bear on the assumption that the investment curve and the revenue curve are the same curve. Those are two different lines, running on two different clocks, and the distance between them at any given moment is the definition of the risk.
So let me leave you with the positioning framework I am actually operating on.
First, the physical layer is where the certainty lives. Whoever controls power, cooling, transformers, turbines, and interconnect capacity controls the binding constraint of the entire buildout. The interesting observation is that the biggest beneficiaries of the AI boom may not be AI companies at all. They may be traditional industrial and energy firms that happen to own the scarce physical assets. That inversion is the most under-priced trade in the room.
Second, watch the revenue line, not the capex line. The capex line is loud. It gets reported quarterly, it moves stocks, it generates headlines. The revenue line is quiet, it is only now beginning to be disclosed separately, and it is the only thing that matters for whether the $10.3 trillion is real. When AI-related revenue starts being broken out on earnings calls, that disclosure will be the single most important data point in the sector. If it appears and it is small, the cycle turns. If it appears and it is large, the cycle extends.
Third, prepare for the impairment window. Every infrastructure cycle ends in a period of asset repricing, and the operators who buy capacity in that window — cheap, depreciated, already-built — capture the surplus that the builders gave up. The bargain is not in the build. It is in the bill. Position for the bill.
Fourth, monitor the financing structure, not just the financing volume. Data center ABS issuance, private credit exposure, GPU-collateralized lending spreads — these are the instruments that will telegraph stress before it reaches the equity market. When those spreads widen, you are watching the fuse get lit, and you will have weeks, not months, to react.
Fifth, and most importantly, hold the question that this entire essay has been circling. Is $10.3 trillion a description of what will happen, or a script that the market is now compelled to follow until it cannot? The number itself is not the story. The story is that a number this large, this unverified, and this amplified has already begun to allocate capital on the basis of an assumption about the future that no one can currently test.
The builders will build. The electrons will decide. And the users — the ones who show up after the impairment, on cheap capacity built by someone else's money — will inherit the thing everyone was arguing about.
That is not a forecast of doom. It is a forecast of sorting. The trap isn't the $10.3 trillion. It's the illusion of infinite growth priced as if the constraint were money, when the constraint was always the current.
