Surfing the Tsunami

Surfing the Tsunami

You cannot stop a wave this size. You can only decide where you are standing when it arrives.

Dispatch — Deconstructing Babel — September 8, 2026

Start with the water.

Global debt has passed $350 trillion and now stands at a record near $353 trillion, roughly 305 percent of world GDP. ^[1] Governments and corporations are expected to borrow a record $29 trillion from global bond markets in 2026, four trillion more than 2024 and double the level of a decade ago; sovereign borrowing in OECD countries alone is running near a record $18 trillion. ^[2]

Those numbers are not new in kind. Debt has been climbing for forty years and every decade has produced someone announcing the reckoning. We have written the arithmetic of it before, at the sovereign level, and the arithmetic has not improved. ^[a]

What is new is what the marginal dollar is being spent on.

For the first time, an enormous share of new global borrowing is being funnelled into a single asset class, on a single thesis, with a single failure mode.

One trillion dollars, five companies, two years

The five largest hyperscalers — Amazon, Microsoft, Alphabet, Meta and Oracle — have guided to combined capital expenditure of roughly $660 to $725 billion in 2026, up sixty to seventy-five percent on 2025, with 2027 projected above a trillion dollars for the group. ^[3] Amazon alone has guided to $200 billion this year, more than double its 2025 outlay. ^[4]

Increasingly they are not paying for it out of free cash flow. Morgan Stanley puts the total data-centre capital requirement near $3 trillion by 2028 and estimates that about half of it will come from hyperscaler cash flows, with the remainder financed through the credit markets — corporate bonds, private credit, and securitisation. ^[5] Global AI-related debt issuance is on track for roughly $570 billion in 2026, more than double the 2025 figure. ^[6]

This is a change in the character of the boom, not just its size. Capital expenditure funded from profits is a company spending its own winnings. Capital expenditure funded by debt is a company spending someone else’s money against a future it has promised to produce.

The circle

Then there is the shape of the money, which is the part that should make anyone sit up and take note.

Nvidia sits on both sides of its own sales. In September 2025 it signed an order with CoreWeave with an initial value of $6.3 billion under which Nvidia is obligated to buy any cloud capacity CoreWeave fails to sell to other customers, running through April 13, 2032. ^[7] It also holds an equity stake in CoreWeave. And in September 2025 it announced a plan to invest as much as $100 billion in OpenAI, a customer, to help fund the compute that OpenAI would then buy from Nvidia.

A vendor lending its customer the money to buy its product, while holding equity in that customer, while guaranteeing the customer’s unsold inventory, is not a market transaction. It is an accounting arrangement. It produces revenue on one balance sheet and an obligation on another, and it will look like growth right up until someone has to settle in cash. We have argued before that this is precisely the class of exposure that accounting standards are structurally unable to see. ^[b]

Here is the tell, and it is recent. The Nvidia–OpenAI deal stalled. By the end of January 2026 the $100 billion investment had been put on ice, with people inside Nvidia reported to have raised concerns about the transaction and Jensen Huang privately playing down the likelihood that the original structure would be finalised. ^{[8] [9]}

Note carefully what that is and what it is not. It is not a collapse. Both companies still need each other and the commercial relationship continues. But it is the first visible instance of a participant in the circle looking at the paper and hesitating. Circular financing works while every participant keeps signing. The interesting question is never whether the circle is elegant. It is what happens the first time one node declines.

We have seen this structure before. Vendor financing is what took down telecom equipment in 1999 and 2000. The difference in 2026 is the scale, and the fact that the paper is being distributed.

The distribution is the danger

The most consequential development is not the borrowing. It is that AI infrastructure is being turned into a financial asset in its own right — packaged, rated, and spread across private credit funds, institutional investors, insurers and major banks.

The numbers on that conversion are not subtle. Data-centre asset-backed securities went from $2.4 billion of issuance in 2020 to $15.5 billion in 2025, and 2026 is on pace for a record. ^[10] Together with commercial-mortgage-backed paper, data-centre securitisation reached roughly $26 billion in U.S. issuance in 2025 and is projected to exceed $60 billion a year by 2028. ^[11] In August 2026 the SEC granted an exemption that clears the path for more data-centre ABS, and in the same week Nvidia announced strategic partnerships with Apollo, BlackRock, Blackstone, Brookfield and Goldman Sachs. ^[10]

Read that last sentence again. Those are not chip customers. Those are balance sheets. That is the machinery of distribution being formally assembled.

This is the move that converts a sector problem into a system problem. A data centre that fails is a bad business. A data centre whose cash flows have been sliced into tranches and sold to pension funds, insurers and private credit vehicles is a transmission mechanism.

Note also who is now creditworthy. J.P. Morgan led a $441 million debt financing for Global AI, a company founded in 2024, building sovereign AI data-centre infrastructure. Two years old. ^[12] The capital markets have decided that AI compute capacity is a financeable industrial asset, and they have decided it fast.

The market has already named the risk

This is not a contrarian read. It is the consensus of the people managing the money.

In Bank of America’s August 2026 Global Fund Manager Survey, 38 percent of respondents named AI hyperscaler capital spending as the most likely source of a systemic credit event — ranking it first for the second consecutive month, with private credit second at 23 percent. ^[13] That reading has been climbing all year: it was 30 percent in February and 34 percent in May. ^[14]

And the second-place answer is no longer hypothetical either. Fitch stated in September 2025 that in a systemic shock private credit could be a meaningful transmission channel, exposing a broad range of investors to losses and elevated redemptions. ^[15] Since then, Fitch’s U.S. private credit default rate has hit a record 6.0 percent on a trailing twelve-month basis — more than double public high yield — across roughly 1,300 tracked borrowers, with a 9.2 percent default rate among private-credit-backed corporate borrowers in 2025. ^{[16] [17]}

So the professionals see the wave, and the water is already moving. That changes what the useful analysis is. Telling people a bubble exists is not information anymore. The question is what the bet actually is, and what would have to be true for it to pay.

The gap, which is the whole argument

Here is the number that matters more than any other in this piece.

Between four and thirteen dollars of infrastructure is being built for every dollar of revenue that end users are actually paying today. ^[3]

That gap is not automatically fatal. Railways were built the same way. So were the electrical grid, the interstate highways, and the fibre that carries this sentence. Infrastructure precedes demand by definition; that is what makes it infrastructure. Every one of those buildouts also bankrupted most of its original financiers, and the assets were bought at cents on the dollar by people who arrived second.

But the gap tells you exactly what is being wagered. The bet is not that artificial intelligence is impressive. It plainly is. The bet is that AI produces enough real, measurable, paid-for economic output, fast enough, to service approximately a trillion dollars of debt on the schedule that debt requires.

That is a claim with a date attached. It is the only claim that matters, and almost nobody is stating it in those terms.

Why the wave cannot be fought

The instinct, on seeing a number like $353 trillion, is to want it stopped. Deleveraging, discipline, a return to sound money.

It will not happen, and it is worth being clear about why rather than moralising about it.

• Debt at this scale is not a policy. It is the structure of the system. A very large share of it is public, which means it is somebody’s pension, somebody’s bank capital, and somebody’s currency. Unwinding it deliberately means choosing which of those to break.
• Every actor is individually rational. Each hyperscaler must build or be locked out. Each government must borrow or contract. Each private credit fund must deploy or return capital. No participant can defect without losing, which is the definition of a coordination trap.
• The wave has already broken. The spend for 2025 and 2026 is committed. The paper is issued and distributed. The question of whether to have this exposure was answered before most people noticed it was being asked.

Which is why the correct posture is not resistance. Standing in front of a wave this size is not principle; it is arithmetic ignorance. You surf it or you are underneath it.

What surfing actually means

Surfing is not optimism, and it is not riding the momentum. A surfer stays just ahead of the break, on the face of the wave, where the energy is — and the entire skill is knowing where the break is going to be before it gets there.

Translated out of metaphor, that means four things.

• Know what the bet is. The wager is on measured productivity growth, not on capability. Watch output per hour and paid end-user revenue, not benchmark scores and demo videos.
• Know where the paper is. The exposure is not concentrated in the famous companies. It is in private credit funds, insurers, pensions, ABS tranches and lease structures — which is precisely where it is hardest to see and slowest to reprice.
• Know what the physical constraint is. Global AI is targeting one gigawatt of capacity by 2029, enough to power as many as 750,000 U.S. homes. ^[12] Multiply that across the sector. The binding constraint on this buildout is not capital and not chips. It is electricity, land, water and grid interconnection, and none of those respond to a term sheet. ^[c]
• Know what survives the break. In every previous infrastructure cycle the physical asset survived and the financing did not. The rails stayed. The railway companies did not. Position accordingly.

Leverage means two things, and only one of them is safe

We measure stability as a ratio: leverage over entropy. Leverage is constructive capacity — the ability to do more with what you have. Entropy is disorder, decay, and the cost of coordination failure.

Financial leverage borrows the same word and does something different with it. It does not create capacity. It multiplies whatever capacity is already there, in both directions.

If the productivity gain is real, a trillion dollars of debt against it is the most constructive act of capital allocation in a century, and the people warning about it now will look like the people who warned about railways. If the productivity gain is not real, the same trillion dollars multiplies the entropy instead — and it has already been distributed through the credit system to people who do not know they own it.

That is the whole question. Not whether the technology works. Whether the output is real, measurable, and arrives on the schedule the debt requires.

What we are watching, and what would change our minds

We publish what would falsify us, so here it is.

• Supports the bet: sustained measured productivity growth in the official statistics, attributable to AI deployment rather than to labour-force composition — and end-user revenue closing the gap toward the capex line rather than widening against it.
• Breaks the bet: end-user revenue flat or growing more slowly than capex through 2027; the first large-scale write-down of AI infrastructure assets; a private credit fund gating redemptions on data-centre exposure; or a going-concern qualification naming AI-related obligations.
• The tell we watch most: whether the circular arrangements grow or shrink. The Nvidia–OpenAI stall is one data point in the shrinking direction. Vendor financing expanding while end-user revenue stalls is the single clearest signal that revenue is being manufactured rather than earned.

We do not know which way it goes. Anyone who tells you they do is selling something.

The reason to care beyond your portfolio

This is not only a market story.

A trillion dollars is being borrowed against the promise that machines will make us dramatically more productive.

If that promise is broken, we get a credit event distributed through the least transparent corner of the financial system, at a moment when public balance sheets have no room left.

If that promise is kept, we get the largest expansion of real capacity in modern history — and that is not the safe branch either. A step-change in productive capacity delivered by systems whose objectives are set by a handful of firms is its own civilization-scale exposure, and we have written about both halves of it: the risk that GAAP cannot price what it cannot quantify, ^[b] and the outcome nobody is modelling, in which the technology works exactly as advertised and the coordination structure around it does not. ^[d]

Either way, the decision was not made democratically and it is not reversible. It was made by a few thousand people allocating capital, each of them rationally, none of them able to stop.

The useful response is not outrage and it is not denial. It is to understand the shape of the wave, to know where the break will be, and to be standing somewhere survivable when it arrives — and to say so publicly, with dates, so that afterwards there is a record of who saw what and when.

References

1. Institute of International Finance, Global Debt Monitor, May 6, 2026 — global debt hit a record of nearly $353 trillion while the global debt/GDP ratio remained stable at 305 percent; nearly $29 trillion was added in 2025 to reach $348 trillion. https://www.iif.com/Products/Global-Debt-Monitor

2. OECD, Global Debt Report 2026: Sustaining Debt Market Resilience Under Pressure, March 6, 2026 — governments and corporations borrowed $27 trillion in 2025 and are expected to borrow $29 trillion in 2026, 17 percent more than 2024 and double the 2015 level; OECD sovereign borrowing at a record near $18 trillion. https://www.reuters.com/markets/rates-bonds/inflation-biggest-risk-debt-markets-facing-big-stress-test-oecd-says-2026-03-04/

3. useLuminix, AI CapEx Bubble Risk: $600B Revenue Hurdle vs. $150B Reality, June 27, 2026, compiling analyst research — Amazon, Alphabet, Microsoft, Meta and Oracle guiding to combined 2026 capex of roughly $660–725 billion, up 60–75 percent on 2025, with 2027 projected above $1 trillion for the group; AI-attributable revenue estimated at $50–150 billion annually, implying a 4–13× gap. https://www.useluminix.com/reports/ai-capex-bubble-risk-600b-revenue-hurdle-vs-150b-reality

4. Forbes, AI Spending Is Surging Faster Than Revenue And Markets Are Starting To Notice, June 2, 2026 — CreditSights estimates the five largest hyperscalers on track for $700–900 billion of 2026 capex, a 36 percent increase on 2025; Amazon alone guiding to $200 billion. https://www.forbes.com/sites/greatspeculations/2026/06/02/ai-spending-is-surging-faster-than-revenue-and-markets-are-starting-to-notice/

5. Morgan Stanley Research, Why Credit Is Core to AI Expansion, July 2, 2026 — meeting data-centre demand requires roughly $3 trillion of capital expenditure by 2028, with about half funded from hyperscaler cash flows and the remainder financed through credit markets. https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/why-credit-is-core-to-ai-expansion

6. Morgan Stanley research as reported June 2026 — global AI-related debt issuance on track for nearly $570 billion in 2026, more than double the 2025 figure, having reached roughly $236 billion by May 31, 2026. https://axis-intelligence.com/ai-data-center-financing-statistics-2026/

7. Reuters, CoreWeave, Nvidia sign $6.3 billion cloud computing capacity order, September 15, 2025 — initial order value $6.3 billion; Nvidia obligated to purchase any residual cloud capacity not sold to other customers through April 13, 2032. https://www.reuters.com/business/coreweave-nvidia-sign-63-billion-cloud-computing-capacity-order-2025-09-15/

8. Wall Street Journal, The $100 Billion Megadeal Between OpenAI and Nvidia Is on Ice, January 30, 2026 — Nvidia’s plan to invest up to $100 billion in OpenAI stalled; Jensen Huang privately played down the likelihood the original deal would be finalised. https://www.wsj.com/tech/ai/openai-nvidia-100-billion-deal-3e10cf34

9. Reuters, Nvidia’s plan to invest up to $100 billion in OpenAI has stalled, WSJ reports, January 31, 2026 — discussions stalled after people inside Nvidia expressed doubts about the transaction. https://www.reuters.com/business/nvidias-plan-invest-up-100-billion-openai-has-stalled-wsj-reports-2026-01-31/

10. InvestmentNews, SEC exemption clears path for more data-center asset-backed securities, August 11, 2026 — data-centre ABS issuance climbed to $15.5 billion in 2025 from $2.4 billion in 2020 and is on pace for a record in 2026 (Bloomberg data); Nvidia announced strategic partnerships with Apollo Global Management, BlackRock, Blackstone, Brookfield Asset Management and Goldman Sachs the same week. https://www.investmentnews.com/alternatives/sec-exemption-clears-path-for-more-data-center-asset-backed-securities/262861

11. Compute Law, Bonds backed by AI data center revenue (securitization), May 23, 2026 — data-centre ABS and CMBS together reached roughly $26 billion of U.S. issuance in 2025 and are on track to exceed $60 billion a year by 2028. https://computelaw.blog/bonds-backed-by-ai-data-center-revenue-securitization/

12. Bloomberg, JPMorgan Leads $441 Million Debt Deal for AI Data Center Firm, August 10, 2026 — Global AI, founded 2024, raised $441 million in debt financing led by JPMorgan Chase; the company plans to deliver 1 GW of capacity by 2029, described as enough to power as many as 750,000 U.S. homes. https://www.bloomberg.com/news/articles/2026-08-10/jpmorgan-leads-441-million-debt-deal-for-ai-data-center-firm

13. Bank of America Global Fund Manager Survey, August 2026, as reported by Seeking Alpha and Yahoo Finance, August 18, 2026 — 38 percent of investors identified AI hyperscaler capex as the most likely source of a systemic credit event, the second straight month in first place; private credit second at 23 percent. https://seekingalpha.com/news/4501234-ai-hyperscaler-spending-seen-as-biggest-source-of-systemic-credit-event-bofa-survey

14. Bloomberg, AI Data Center Debt Rises as Wall Street Flags Systemic Risk, May 19, 2026 — roughly 34 percent of global fund managers surveyed by Bank of America named AI hyperscaler capex the most likely source of a future systemic credit event, double the April share; Business Insider reported the February 2026 reading at 30 percent. https://www.bloomberg.com/news/articles/2026-05-19/ai-data-center-debt-rises-as-wall-street-flags-systemic-risk

15. Fitch Ratings, Systemic Shock Could Expose Private Credit’s Broad Linkages, September 29, 2025 — in the event of a systemic shock, private credit could be a meaningful transmission channel, with negative repercussions for a broad range of investors including losses and elevated redemptions. https://www.fitchratings.com/research/corporate-finance/systemic-shock-could-expose-private-credits-broad-linkages-29-09-2025

16. Bloomberg, Fitch’s Private Credit Default Rate Hit Record in Second Quarter, July 30, 2026 — the trailing twelve-month U.S. private credit default rate rose to a record 6.0 percent across roughly 1,300 tracked borrowers, up from 5.7 percent the prior quarter. https://www.bloomberg.com/news/articles/2026-07-30/fitch-s-private-credit-default-rate-hit-record-in-second-quarter

17. Forbes, Rising Private Credit Defaults Are Testing Banks And Borrowers, May 24, 2026 — Fitch reported the U.S. private credit default rate at a record 6.0 percent in April 2026 and estimated a 9.2 percent default rate among private-credit-backed corporate borrowers in 2025. https://www.forbes.com/sites/investor-hub/2026/05/24/rising-private-credit-defaults-are-testing-banks-and-borrowers/

Deconstructing Babel, referenced above

a. Seeing the Debt Clearly — The sovereign arithmetic: interest as the fastest-growing line in the budget, and why growing out of it requires permanent escape velocity.

b. Unquantifiable Risk — Why accounting standards cannot price an exposure they cannot quantify.

c. Energy Grid Saturation — AI Data Center Demand — The physical ceiling: electricity, interconnection and the constraint that does not respond to capital.

d. The Third Scenario No One Is Pricing In — The branch in which the technology works exactly as advertised and the coordination structure does not.

— David F. Brochu with Edo de Peregrine, partners/collaborators · Tuesday, September 8, 2026 · Deconstructing Babel

Terms used in this pieceCircular FinancingThe Productivity WagerData-Centre SecuritisationFinancial LeverageStability Equation (S = L/E)Coordination FailureFull definitions in the glossary.