The technology is right and the financing is wrong
Opinion. Not investment advice.
The bull and bear cases on AI are arguing about the wrong variable. Both sides are litigating whether the technology works. That question is settled, it works, and the Astra results this month settled the most demanding version of it. What is not settled, and what almost nobody is underwriting properly, is how the build-out is being paid for.
Those are separable questions. Treating them as one is the central analytical error of this market, and it is being made by bulls and bears in equal measure.
Why Nvidia is not the bubble
Nvidia trades at roughly 24× forward earnings while growing data-center revenue 92% year-over-year. AMD trades near 81× forward on materially slower growth. If you are looking for speculative excess, the most valuable company in the world is a strange place to start.
I keep encountering the argument that a $5.31 trillion is self-evidently absurd. It isn't. It is roughly what you get when you multiply real, contracted, cash-collected earnings by a multiple below the S&P's long-run average for high-growth technology. The market has been more disciplined about Nvidia's than about almost anything else in this cycle.
The risk in Nvidia is not the multiple. It is the quality of the revenue: and that is a completely different exposure, sitting one layer down.
The thing I actually worry about
Nvidia is reportedly in talks to provide a backstop of up to $250 billion so that OpenAI can lease capacity from a ten-gigawatt campus. The mechanism lets OpenAI borrow against Nvidia's investment-grade credit rather than its own, to buy Nvidia's chips.
Sit with the structure. The supplier underwrites the customer's ability to purchase the supplier's product. Revenue appears on one income statement, contingent liability on another, and the same underlying asset can end up pledged more than once across the financing chain. This is not fraud and it is not even unusual in capital-intensive industries. It is, however, exactly the arrangement that makes revenue quality and credit quality the same variable wearing two hats, and reported growth stops being an independent signal about demand.
Around it: $570 billion of AI-related debt issuance this year. Utilities levering up, with investment-grade issuance going from about $135 billion to $145 billion to serve data-center load. Data-center paper being repackaged and sold to insurers in structures that people who lived through 2007 keep describing with the same word.
And the : an institution not given to color, named an AI bust and a circular-financing collapse two of its three most urgent threats to global financial stability. Central banks do not do that lightly.
The 2001 analogy, used properly
Telecoms in 2001 is invoked constantly and almost always lazily, usually as a synonym for "this will crash." The actual lesson is more specific and more useful.
The fiber was real. The demand forecasts were eventually correct, we use all that fiber now, and more. The equity was still destroyed, because the build-out was debt-financed against demand that showed up years after the interest payments came due. The technology thesis was right and the capital structure killed you anyway.
AI is better than 2001 in one crucial respect: the demand exists today and is contracted. Nvidia's data-center revenue is not a projection. Microsoft's backlog is not a hope. This is genuinely different, and the people calling it a pure bubble are ignoring cash that has already been collected.
AI is worse than 2001 in another: the financing has become circular in ways telecoms' was not. Lucent famously vendor-financed its customers and it ended badly, but Lucent was not also the largest supplier, the backstop provider, and an equity holder in its customers simultaneously.
So: right technology, questionable capital structure. That is not a contradiction. It is the most common way that transformative technologies destroy capital while succeeding.
What the secondaries are telling you
The single most information-dense datapoint I found this month: $600 million of OpenAI shares sitting unsold in the , while buyers queue $2 billion deep for Anthropic.
Secondary markets are where people who have actually seen the numbers express opinions they will never say into a microphone. Employees. Early investors. Board-adjacent holders. When one side has unsold inventory and the other has a two-billion-dollar bid stack, and both companies are pre-IPO in the same window: that is not sentiment. That is a verdict.
The verdict is also defensible on the fundamentals. Anthropic: $47 billion , up from $1 billion in December 2024, at roughly 20× revenue, enterprise-weighted and contracted. OpenAI: about $25 billion annualized, roughly $14 billion of losses this year, no profitability expected before 2030, consumer-weighted revenue, at a proposed multiple closer to 34×. While restructuring its relationship with the partner that funded it and signing a $50 billion exclusive cloud deal with that partner's largest competitor.
Sam Altman and Sarah Friar have called anything below a trillion dollars a "nonstarter." The secondary market is currently declining to clear at a number well below that. One of these two positions is going to move, and it is not usually the market.
Where I would be looking
Memory, with a clock on it. constraints gate every accelerator regardless of architecture, which makes memory a bet on AI volume rather than on which chip wins. Microsoft attributing $25 billion of capex to memory inflation is that pricing power showing up in someone else's income statement. But memory is the most violently cyclical business in semiconductors, these names have already re-rated to top-15 globally, and capacity is being announced. This is a thesis with an expiry date, and I would rather be early to exit than late.
Broadcom over picking a . at 27.8% of shipments growing 44.6% against 16.1% for merchant is a structural share shift. Broadcom sits behind Google's TPU, Meta's MTIA, Microsoft's Maia and the OpenAI/Anthropic Titan program, exposure to the migration without having to know who wins it.
Credit, not equity, as the tell. If this cycle breaks, it breaks in spreads first and equities second. Watch AI-adjacent issuer spreads, utility issuance, and most immediately whatever Nvidia discloses about vendor financing on August 26th. That disclosure is more important than the revenue number, and roughly nobody will lead with it.
What I would avoid: anything whose moat is a model it does not own, competing against free Chinese at a tenth of Western pricing. The 80% price cut on GPT-5.6 Luna is the incumbents telling you where that floor goes.
The calls
- Anthropic lists before OpenAI, and prices better relative to its last private mark.
- OpenAI's September window slips. The secondary market has already said so.
- Nvidia's August 26th print beats on revenue and the stock reaction is decided by the financing disclosures, not the beat.
- Microsoft's OpenAI concentration: roughly 45% of a $625 billion backlog tied to a customer actively diversifying away. Is a larger risk than its capex, and is not what the market is currently arguing about.
- Credit spreads on AI-adjacent issuers widen before any AI equity index draws down 20%.
What would change my mind
- If Nvidia discloses minimal vendor financing on August 26th, the circularity concern is overstated and I have been reading too much into reported talks.
- If OpenAI prices at or above $1 trillion in September, the secondary signal was noise and I over-weighted a thinly traded market.
- If AI-adjacent credit spreads stay tight through a demand deceleration, the credit-first thesis is wrong and the equity market is the better sensor.
- If enterprise ROI improves sharply, the demand underpinning the capex is more durable than the survey data suggests, and the whole capital-structure worry becomes academic.
The uncomfortable summary: I think the technology is more real than the skeptics allow and the financing is more fragile than the enthusiasts allow, and I do not think those two statements resolve into a single directional view. Anyone offering you one is compressing information out of the picture.
Predictions from this editorial are logged in the ledger with resolution dates and falsifiable criteria.