Letter No. 3 · The Snake Eating Its Tail · Part 3 of 3

Where's the Next Tail? The Repricing Arrives

August 2026: the credit market votes, the insiders sell, the landlord owes $134 billion, and the demand math still doesn't close. The snake is looking for its next meal.

By Paul D. Diaz, EA, MBA · 2026-08-03 · Opinions are the author's; every figure is sourced below.

Written in the first week of August 2026. Every number below is a present-tense fact, not a forecast.

Part 1 argued that thirty years of booms share one engine. Part 2 measured who absorbs the damage. This part does something the first two could not: it reads the tape. Because between the drafting of this series and its publication, the repricing started arriving on its own schedule.

The canary was the most credible bird in the mine

SpaceX — government contracts, Starlink subscriptions, commercial launch — priced the biggest IPO in history in June and is down roughly a third from its debut close in seven weeks. Not a rumor-driven flush: a repricing of the one AI-adjacent asset the market was actually forced to price, because a public listing removes the luxury of a stale private mark.

That last clause is the mechanism to watch. OpenAI's IPO is delayed. Anthropic has not priced a new round or issued fresh guidance. Neither fact is neutral. A delayed IPO is information — it means the underwriters saw what happened to SPCX and declined to discover the real bid. And "no new mark" does not mean the old mark is right; it means nobody has been forced to test it. The 2007 lesson applies verbatim: the CDOs were worth par until someone had to sell one. Mark-to-model survives exactly until it meets mark-to-market, and the meeting is never gradual.

The credit market voted first, again

In late July, Nebius fell 10% and CoreWeave 9% in a session — not on earnings, but on rising credit-default-swap costs. Equity investors argue about stories; the credit market prices balance sheets, and it moves first. In 2007 the ABX index was screaming by February while stocks made new highs into October. Watch the spreads, not the narratives.

The insiders have voted too: CoreWeave's co-founder sold roughly $734 million of stock in Q2 2026 and its CEO about $447 million — $1.18 billion out, in one quarter, from the people with the clearest view of the backlog, the leverage, and the customer concentration.

The landlord problem: Oracle

Oracle spent four decades as the perfect business: software licenses, recurring maintenance, near-zero incremental capital — cash-heavy and asset-light. In two years it inverted itself: $55.7 billion of capex in fiscal 2026 against roughly $32 billion of operating cash flow, total debt of $134.6 billion, negative outlooks from both Moody's and S&P, 30,000 layoffs whose stated arithmetic is freeing operating margin to service the debt, and a financing round where US banks stepped back and PIMCO had to anchor $10 billion of a $16.3 billion raise. Asset-heavy and cash-light — the exact opposite of the company Wall Street spent a generation pricing.

And more than half of its $553 billion in performance obligations rests on one tenant: OpenAI, via the $300 billion Stargate arrangement — a counterparty with no profits and a postponed listing. Here is what makes this worse than commercial real estate: an office tower that loses its anchor can be re-let or converted. A purpose-built AI datacenter — power density, liquid cooling, interconnect topology specified for one tenant's training runs — is a single-use industrial asset. When the tenant leaves, the landlord does not own a building with a vacancy. It owns a very specific machine, financed with other people's money, whose next-best use is worth a fraction of its build cost.

This is where Part 1's fiber analogy finally breaks — in the bears' favor. Dark fiber was generic: any carrier could light it, and the glass waited ten years without depreciating. GPUs are a treadmill: a three-to-five-year useful life, obsolete before they fail. The 1990s built an asset. This cycle is renting one from the future, permanently — roughly $5.3 trillion of committed hyperscaler capex from 2025 through 2030 on Goldman's arithmetic, much of it buying silicon with a half-life.

The demand side does not add up — and that is the actual thesis

Set aside the financing loops and ask the retail question: where is the revenue?

The overwhelming majority of AI chatbot users sit on free tiers. Consumer freemium conversion historically runs in the low single digits. Every free query burns real GPU-seconds against zero revenue — which is why the pre-AI ad playbook does not port: Google and Meta could subsidize free users because serving an ad impression costs approximately nothing. Serving a free AI answer does not.

Meanwhile the paying customers with the most money are the most likely to leave. Enterprises with data-sovereignty rules cannot send client information to a cloud model at all. And any CFO can now do the arithmetic that a Chinese open-weights model — or any capable distilled model — delivers the large majority of frontier performance at a small fraction of the cost, running on a machine the business owns. A one-time hardware purchase that permanently deletes a subscription line item. The substitute gets cheaper every year. The frontier bill does not.

Demand for AI is real. That is not the issue. The issue is that demand is migrating to the cheapest adequate provider while the capex bill was sized for the frontier case. Rising usage, collapsing unit economics — the telecoms were right that bandwidth demand would explode, and bankrupt anyway, because they were wrong about who would capture the value.

The concentration nobody says out loud

NVIDIA's fiscal-2026 10-K discloses that four direct customers account for 61% of its revenue — one alone is 22%. Its data center segment is roughly 90% of a $216 billion year; the gaming business that built the company is now about 7%. Run the simple scenario: the top four cut spending in half and a competitor takes a slice of the rest, and NVIDIA is a −30% to −40% revenue company — against a valuation built entirely on compounding growth. Strip the AI datacenter bid altogether and what remains is roughly a $22 billion business. That is not a prediction. It is what the filing says the exposure is.

Where's the next tail?

The snake has always found one. Fiber's corpse fed the web. The dot-com crash's rate cuts fed housing. Housing's collapse fed ZIRP, which fed everything. The pandemic's liquidity fed this.

So the closing question of Part 1 gets its sharper, August-2026 form. The durable third of this buildout — the land, the substations, the transmission queues, the shells — will be repriced and repurposed, exactly as the fiber was. The silicon majority will not; it will be scrapped or run to death. Reshuffling is repricing: the market discovering the difference between what something cost to build and what it actually does.

When that discovery happens — when the compute overbuild finally clears at its real value — the cost of intelligence drops toward zero for whoever is standing there to buy it. The 1990s version of that moment created Google. What does this one create? And the more personal question, the one Part 2 exists to answer: when the repricing reaches your statement, will it find a plan — or a December 31 balance that no longer exists?

The firm behind this analysis: THE TAX CUTTERY® — back taxes, IRS letters, planning, and returns. Start a conversation any time.