On November 20, 2025, a Crypto Briefing item crossed my terminal carrying a single load-bearing figure: $3.7 trillion. It was attributed to an analysis of AI data center revenue requirements and framed as evidence of "systemic risk." No methodology. No time frame. No author. No link to a primary document. Just the number, a currency symbol, and three commas.
I have read earnings disclosures for sixteen years. An unanchored number moves capital faster than a verified one ever does, because it cannot be falsified. So before accepting $3.7 trillion as either forecast or warning, I apply the first rule of any audit: establish the unit of account. Code is law only if the audit trail is unbroken.
The figure is not fabricated. It is a computed output — real arithmetic, undisclosed inputs. And its meaning swings by a factor of six depending on one definition the article withheld: whether $3.7 trillion denotes annual revenue, cumulative revenue, or capital-recovery revenue. That omission is not a footnote. It is the entire thesis.
To see why, you have to know what $3.7 trillion has to compete against. On the Gartner baseline, global IT spending in 2025 runs about $5.4–5.6 trillion. Global public cloud infrastructure — IaaS plus PaaS — is roughly $350 billion. Enterprise software sits near $650 billion. Digital advertising is about $700 billion.
Hold those four pools together and the first interpretation collapses. If $3.7 trillion meant annual revenue, it would equal roughly two-thirds of all global IT spending. No technology category in history has captured two-thirds of IT spend. I assign that reading a probability below 10%.
The second interpretation — cumulative revenue across six years — is aggressive but not absurd. It requires AI to construct a second cloud market on top of the existing one. Global cloud revenue over a comparable six-year window lands in the $4–5 trillion range. This reading means "replicate the cloud, then add half." I put it at 30–45%.
The third interpretation survives arithmetic. If $3.7 trillion is the capital-recovery requirement — the revenue needed to cover depreciation plus a return on a defined capex base — then the number is provenance-derived, not demand-derived. At a six-year depreciation schedule, 50% gross margin, and zero residual value, $3.7 trillion of revenue corresponds to roughly $1.85 trillion of capital expenditure. The math closes. The other two readings do not.
Here is why this matters to anyone holding crypto assets. The most aggressive buyers of the AI infrastructure narrative are precisely the audiences crypto media serves — holders of bitcoin miners pivoting to AI hosting, tokenized compute protocols, and stablecoin lenders financing data center debt. A missing unit of account is not an academic problem for them. It is a liability sitting on their balance sheet under a mislabeled line item.
The choice of outlet is itself a data point. "AI data centers carry systemic risk" resonates with a community already disposed to believe traditional finance is structurally fragile. That resonance explains the distribution. It does not validate the number. And the same community holds the very assets most exposed to the number's resolution — a mismatch worth noting before we reach the arithmetic.
Let me do the reverse engineering the article skipped. Work backward from $3.7 trillion.
At 50% gross margin, the revenue implies about $1.85 trillion of capex. At AI data center construction cost of $25–30 million per megawatt, that is 60–74 GW of new IT load. Convert to energy: 74 GW × PUE 1.2 × 8,760 hours ≈ 780 TWh per year. Reference point: all global data centers consumed roughly 415 TWh in 2024. This single build-out would nearly triple global data center power draw and require about 90–100 GW of new generation capacity — 80 to 100 large gas turbines, or 30 to 40 nuclear units.
Now the constraint capital cannot buy. The bottleneck is not chips. It is the interconnection queue. Large power transformers carry 3–5 year lead times. Grid interconnection queues in PJM and ERCOT run 4–7 years. Data center shells take 12–24 months to construct. The mismatch is structural: the shell is a 20–30 year asset; the GPU inside it is a 1–2 year asset. That duration mismatch is the core fragility of the entire financing structure, and the article compressed it into the phrase "infrastructure adaptation."
The second ignored signal is GPU rental deflation. H100 on-demand pricing fell from roughly $8 per hour in 2023 to $2–2.50 per hour by 2025 on third-party platforms. Any revenue model resting on static rental assumptions is already wrong. If $3.7 trillion assumes stable compute pricing, it overstates achievable revenue materially.
The third is depreciation. Hyperscalers assume 4–6 year server lives. GPU generations turn over every 1–2 years. Shorten the schedule from six years to four and annual depreciation expense rises 50%. For a firm running $80 billion in annual capex, that is tens of billions in operating profit erased by an accounting election alone. A large share of the "$3.7 trillion requirement" is plausibly revenue needed to cover an extended depreciation path — not genuine economic demand.
The fourth is the circular structure. Chip vendors take equity in customers. Customers spend that capital on chips. That becomes revenue. Model labs sign multi-year compute commitments, some publicly reported in the hundreds of billions, which hyperscalers cite to justify new capex, which converts into fresh chip orders. Data center SPVs move the assets and debt off balance sheet into private credit. Every link confirms revenue before cash flow and prices commitment above verified ability to pay.
Now map who bears the loss when revenue undershoots. Hyperscalers sit at the top — self-funded, running proprietary silicon, able to absorb a shortfall by extending depreciation and routing demand internally. Chip and system vendors hold long order visibility but concentrated customers and rising receivables. Neoclouds carry the sharpest exposure: high leverage, GPU asset duration of 1–2 years against debt duration of 5–7 years, and a debt-service coverage ratio that reprices violently against a falling rental rate. Converted bitcoin miners face customer concentration and counterparty credit risk on their long-term hosting contracts. The financiers — private credit funds, asset-backed securitization buyers, and single-purpose vehicles — sit at the bottom of the waterfall and take first loss.
The scale of that financing is not marginal. Industry estimates suggest roughly half of AI infrastructure investment across 2024–2028 requires debt, implying on the order of $1.5 trillion in issuance. That is among the largest single-sector exposure expansions in the history of private credit. When a market of that size is priced by private vehicles with limited secondary liquidity, the pricing spread and the funding availability become the circuit breaker for the whole chain.
Strip the headline and look at what actually pays today. The verifiable, scaled AI revenue streams in 2025 are narrow: coding assistants and customer-service automation. Combined, they do not yet sustain a hundred-billion-dollar annual run rate, let alone a trillion. Every credible revenue pool — cloud AI services, AI-native applications, advertising and subscription uplift — sums to roughly $200–400 billion. Against $3.7 trillion, that is a ten-fold gap. The figure therefore describes a demand-side obligation that must materialize, not a demand-side observation that has.
There is one technical variable the article never mentions, and it cuts both ways. Inference efficiency is deflationary for compute demand: per-token cost has fallen over 99% in two years, and MoE sparsity, INT4/FP8 quantization, speculative decoding, and prefix caching are production-ready, not research-stage. But test-time compute and agentic workflows push the other direction — reasoning models stretch output tokens 5 to 50 times, and agents multiply calls per task from one to dozens. The net compute curve is the product of agent penetration, tokens per task, and inference price. That product, not any model leaderboard, decides whether $3.7 trillion is reachable.
I have seen this pattern before. In 2017 I built a checklist-based due diligence framework for an ICO portfolio and flagged two tokens as structurally unsound from explorer data alone — before launch. The tell was identical: revenue recognized ahead of cash, promises stronger than capacity. Code is law only if the audit trail is unbroken, and the AI data center trail currently runs through three sets of books that never reconcile on the same day.
Everyone reading the $3.7 trillion headline reaches the same conclusion: AI is overbuilt, the bubble is priced. I reach the opposite conclusion from the same arithmetic.
If depreciation holds at six years and gross margin at 50%, then even hitting $3.7 trillion in cumulative revenue delivers close to zero excess return. The number is not a ceiling on demand. It is the floor beneath which the capital structure breaks. The operative question is not whether $3.7 trillion is achievable. It is whether $3.7 trillion is sufficient. Consider the hyperscaler trajectory: the four largest players already run $380–400 billion a year, and at 20–30% growth, cumulative 2025–2030 spend lands at $2.5–3.0 trillion. That is the same order of magnitude as the reverse-engineered $1.85 trillion.
Which produces the central judgment of this analysis: the industry has already committed to roughly this scale of construction. The $3.7 trillion figure is not a prediction of future demand — it is the backward computation of what existing commitments already require. The demand-side promise occurred before the demand-side income.
And the crypto channel is the transmission path nobody is pricing. If a leveraged neocloud or a miner-turned-host must liquidate GPU assets, secondary market liquidity is thin. Realized prices may clear at 30–50% of book value. Through tokenized debt, stablecoin lending desks, and SPV exposure, that markdown transmits into on-chain credit with no circuit breaker. Code is law only if the audit trail is unbroken — and here the trail leaves the chain at the exact moment the collateral is revalued.
This is why I treat the $3.7 trillion figure as a stress-test input, not a verdict. It does not tell us whether AI has value. It tells us that the timing of value realization and the timing of debt service may not match. History is patient with technology and merciless with financing structures. The fiber buildout of 1999 delivered its long-term value — and still destroyed a generation of telecom investors.
Over the next 18 months the central variable in the AI narrative shifts from model capability to revenue verification and financing cost. Four quarterly indicators will decide whether $3.7 trillion stays a distant concern or becomes an immediate one: hyperscaler depreciation policy changes, H100 and H200 spot rental prices, data center ABS issuance spreads, and neocloud debt-service coverage ratios. If all four deteriorate in the same direction through 2026, the gap between far risk and near risk closes inside a single earnings season. Watch the audits, not the headlines.
