The Casualties, and the Rebuttals

A thesis that only predicts winners is a narrative. A real thesis identifies who gets hurt in the process, names them, and tests itself against the strongest case for the other side. Here, both questions lead to the same place: the counterparty.

Starting in the credit market with Oracle’s five-year credit default swaps trading near 198 basis points, the highest on record for the company, whilst the same company reports the strongest booking growth in its history. Record demand on one axis, but record default insurance on the other.⁸³

Concentration. The US$553 billion backlog is committed by essentially three customers, with OpenAI’s commitment being near US$300 billion.⁸⁴ An RPO is really only as good as its counterparty, and in this case. OpenAI’s 2026 annualized revenue sits at US$25 billion, while contracted infrastructure obligations run into the hundreds of billions, if not trillions.⁸⁵ This backlog is a credit exposure wearing a growth costume, and the credit market has definitely taken notice of this.

Followed with timing. Oracle spends now with revenue that arrives later. Free cash flow ran to the negative of US$24.7 billion, and the company has raised on the order of US$45-50 billion across 2026;⁸⁶ project loans run to US$18 billion in New Mexico and US$38 billion across Texas and Wisconsin.⁸⁷ Completion dates have reportedly slipped from 2027 towards 2028,⁸⁸ and banks providing financing have begun buying credit default swaps on their own client.⁸⁹ Standard risk management, and also the credit market quietly stating its estimate of the tail.

And now the circular financing. OpenAI’s revenue grew from US$3.7 billion to US$13.1 billion in the span of two years, which is extraordinary; however, that’s against infrastructure commitments reported near US$1.4 trillion;⁹⁰ HSBC modeled a funding gap near US$207 billion even on generous assumptions.⁹¹ Nvidia has pledged up to US$100 billion into OpenAI, which uses that capital to buy Nvidia chips.⁹² Nvidia also holds equity in CoreWeave and has agreed to absorb whatever excess capacity it might have, whilst also selling CoreWeave its chips.⁹³ About US$92 billion of partner debt has been raised across the ecosystem.⁹⁴ It’s simple. Investors fund OpenAI, OpenAI signs a contract with Oracle, Oracle borrows against the contract that they have, the backlog appraises both companies, raises both stocks, and higher stocks raise more capital. Every link does make sense and is rational. But the chain as a whole requires capital markets to keep the faith until cash generation catches up.

The neoclouds run the same exact structure but with higher leverage and lower quality. CoreWeave’s quarterly interest expense doubled to US$536 million, nearly half its adjusted EBITDA, against US$24.9 billion of debt and a net loss widening to US$740 million, with its top two customers at roughly 65% of revenue.⁹⁵ Nebius carries around US$10 billion of convertible debt accreting to 120% of principal at maturity.⁹⁶ Just to be clear, these are not utilities with rate bases. These are leveraged companies with concentrated customer books. They are financed at a spread, coupled with racing chip depreciation.

All of this sets up the rebuttal that actually threatens this essay, and it comes in two parts. Part one: Dark fiber. The oversupply in the 2000s was a timing failure, not a demand failure. The demand turned out to be real, and eventually all the excess capacity was consumed. The problem was that the fiber built in 1999 was not needed and was not used until 2010, and the companies that built it went bust by 2002.⁹⁷ If efficiency keeps getting better, and compute demand for each unit of intelligence decreases, today’s gigawatts could become tomorrow’s dark fiber.

My rebuttal to this. Fiber was speculative and empty; it earned absolutely nothing while it waited for buyers who did not exist. However, this capacity is take-or-pay, pre-sold before it is even energized. The question is whether the customers who signed the contract are financially intact and capable of paying, or whether there are even customers at all. This is a much more narrow and specific risk. The problem with dark fiber was not merely due to there being too much capacity, but instead there was too much uncontracted capacity that was built. Efficiency is the only risk, and so far any collapse in inference costs has led to expanded consumption, not less compute demand.⁹⁸

Part two: Counterparty collapsing. What if a customer at the center of the order book cannot pay? A client like OpenAI that has provided backlog worth billions of dollars. But let’s set up a clear distinction. Counterparty risk is not demand risk; they are two separate things. If OpenAI defaults, it’ll be a credit event for Oracle, not a demand event for the world of compute. The gigawatts and computing power don’t disappear; they are physically there. Almost like an empty apartment in a booming city. A tenant defaulting in a full city is a landlord’s problem; in an empty city it is everyone’s. Part II reflects that the city is full. When Microsoft rations compute away from its own products, and AWS has an unserved demand at a US$142 billion run rate, repossessed capacity finds buyers. In this case, Oracle’s capacity will easily be re-let.⁹⁹

All this only holds while scarcity persists – and it won’t forever. If this thesis is correct, the buildout eventually catches up with demand, capacity becomes abundant, and returns begin to normalize. The failure case is different: hyperscaler capacity surpluses flood the rental market even before the demand arrives, collapsing counterparty and demand risk into one.¹⁰⁰ This thesis is falsified by continuous utilization demands across energized capacity, backlog cancellations over renegotiations, GPU rental prices falling below the cost of capital and staying there, or a continuous lapse of counterparties walking back commitments the way OpenAI walked back in February.¹⁰¹ None of which is observable currently.

None of this is really bearish as well depending on how you read it. Railroads went bankrupt in waves while rails stayed and carried an entire economy that compounded over a century. WorldCom died but fiber carried the internet.¹⁰² The buildout can be real, necessary, and even civilization changing but there will always be several entities that are destroyed by the attempt.

Risk asymmetry highly favors the builders. For a hyperscaler, overbuilding is a survivable accounting event. Just write down capacity, compress returns, and send a little apology letter to shareholders. To which most of them can revert back to their business models that the cash to fund this entire buildout in the first place. Underbuilding, however, is a death sentence. It means losing the future of infrastructure that compute and demand will grow upon indefinitely. Every CEO in this essay can do that math. In fact, all of them have. Hence, US$700 billion.¹⁰³

---