Three doors. One drain. That's the frame that landed in my feed at 4 a.m. Prague time, and it hasn't left me since.
The thesis, stripped down: capital in agentic AI isn't flowing freely. It's being funneled โ through equity concentration, through compute debt, through the commoditization of the inference layer. Three doors, all opening onto the same basement room where a handful of players hold every lever.
I read it the way I read everything in a bear market: one eye on the narrative, one eye on the wallet. Because when a story arrives this clean, my first instinct isn't "wow." It's "who benefits, and what are they burying?"
Here's the part that matters. The source material behind this frame is messy. Really messy. And the mess is the story.
Let me set the table. Over the past week, a report has been circulating that packages the AI capital boom into a tidy metaphor โ "three doors" โ and stuffs it with numbers that read like a fever dream.
Anthropic carrying a $518 billion commitment and a $1.8โ2 trillion valuation. Google shipping something called "Gemini 4 Argon." A company named "TypeSafe AI" raising $870 million in a Series A at a $7.5 billion valuation. A personal AI computer called "Ghost" pulling $11 million in seed.
I've spent nine years watching this industry, and I've never seen a naming convention that looks less like reality. Google's public line runs 1.0, 1.5, 2.0, 2.5. No elements. No "Argon." And TypeSafe was a Scala toolchain company acquired by Lightbend โ not a frontier lab. A Series A at $870 million would be the largest in history. The report itself flags the mismatch: a source labeled "blockchain/Web3 news" with zero Web3 concepts anywhere in the text.
That's not a small thing. That's a smoke alarm.
I cut my teeth on exactly this kind of check. In late 2017, at sixteen, I bypassed the wires and watched Ethereum's chain split live โ block heights, hash rate, the panic in the Telegram rooms. The lesson wasn't about being right. It was about being fast and skeptical at the same time. Speed is the only metric that survived the crash, I've learned โ and speed without verification is just noise.
So I'm treating this report as a narrative sample, not a data set. And in a bear market, that matters more than ever. My readers don't want a lecture on valuation multiples. They want to know which structures are load-bearing and which are cardboard. This report is a load-bearing structure built on cardboard numbers โ and that combination is the most dangerous kind.
Here's what's real underneath the noise.
AI capital is genuinely concentrating. The direction is right even where the digits are wrong. The overwhelming majority of recent AI funding has flowed into rounds for companies that were already winners โ Anthropic, OpenAI, xAI, Mistral. Same names, again and again. That's not a funnel invented by a language model. That's the actual shape of the market.
Compute debt is a real mechanism, and crypto people should recognize it. Instead of buying GPUs outright, labs sign long-term compute commitments, then finance those commitments through special purpose vehicles, vendor financing, and GPU-collateralized debt. Nvidia and Broadcom aren't just selling chips anymore โ they're becoming quasi-banks, underwriting their own customers. Broadcom's custom ASIC deals, reportedly tied to financing arrangements in the tens of billions, are the clearest signal: the chipmaker is now the lender.
I watched this exact pattern in 2022. The difference is that FTX's collateral was a token with no cash flow. Here, the collateral is a GPU โ an asset that depreciates 30โ50% a year, with a resale market that can freeze overnight. That's not a moat. That's a countdown.
Here's the mechanical problem. If GPU residual values collapse โ and they will, as each generation lands โ the collateral backing those loans shrinks while the debt stays fixed. That's a margin call in slow motion. It doesn't need a villain. It needs one new chip generation and one soft demand quarter.
Inference is commoditizing, and that's the real squeeze. The report describes Google pricing cached input at a 95% discount โ from $2 to $0.10. The number is almost certainly inflated; the real figure in Gemini's pricing history sits closer to 75%. But the mechanism is genuine. Prompt caching, quantization, speculative decoding โ all of it drives inference cost toward zero. When inference gets cheap, the layer that used to print money becomes a loss leader.
Stack those three and you get the funnel: equity concentrates, debt piles up, and the price of the product collapses. The players winning the pricing war are the ones with structural cost advantages โ like Google's in-house TPUs โ not the ones with the biggest promises.
And here's the trap the report walks into without noticing. It argues that if the market accepts a $1.8โ2 trillion valuation, the $518 billion commitment becomes "rational." That isn't analysis. That's reflexivity โ using price to justify debt, and debt to justify price. Soros would spot it in a heartbeat. So would anyone who survived 2021.
What's missing is just as telling. No cash runway. No interest rates. No covenants. No named equity investors. A $518 billion obligation against an undisclosed revenue base is not a business plan โ it's a bet dressed as a forecast. The report's own logic admits it can't verify a single number against a primary source.
Notice, too, what the funnel framing erases. It treats AI compute companies as if they were standard software firms โ "a scale that would bankrupt a normal software company." But they aren't. Compute businesses are capital-intensive, depreciating, and rate-sensitive. Borrowing the software playbook and calling it analysis is rhetoric, not rigor.
Now the angle nobody's publishing.
Everyone's reading this as an AI story. I'm reading it as a crypto story wearing an AI costume. And the costume is the tell.
Look at what the funnel leaves out. It never mentions open source. Not once. No Llama. No DeepSeek. No Qwen. But open-weight models are the single strongest force against the "only the biggest balance sheet survives" thesis. A model trained for a fraction of the cost is now nipping at the heels of closed frontier systems. If capital were the only door, open source would be impossible. It isn't.
It also never mentions China. The parallel price war โ DeepSeek, ByteDance, Alibaba cutting API costs to the bone โ is the loudest counterexample to the funnel narrative, and it's simply absent. When a report systematically omits its own best counterarguments, that's not oversight. That's framing.
Then there's the part that touches us directly. Crypto has been running its own three-door funnel for years. Equity concentrates in a handful of L1s and infra plays. "Compute debt" has an on-chain cousin in over-collateralized lending against volatile assets. And "inference commoditization" is just another name for the fee compression that kills every middleware protocol the moment it stops being scarce.
This is also why I've never bought the RWA-on-chain fairy tale. For three years, the pitch has been that traditional institutions will bring their assets on-chain. But institutions don't need a public chain to move a bond. They need a custodian, a lawyer, and a Bloomberg terminal. The chain is the last thing they want โ it's the part they can't control. The funnel proves the point: capital doesn't seek openness. It seeks the shortest path to the lever.
I learned this the hard way. In 2021, I watched social capital outpaced code in the ape arcade โ profile pictures trading on status, not utility. The floor held as long as the vibes held. When the vibes broke, the floor didn't fall. It vanished. Same physics. Different asset.
So when someone tells you AI capital is funneling into an oligopoly, ask the crypto question first: what's the collateral, who holds it, and what happens when it reprices? Liquidity flows like adrenaline, not like water. It rushes in, and it flees faster.
Here's what I'm watching next.
Not the AI valuations. The financing terms. The interest rates, the covenants, the margin calls hidden inside those compute commitments. That's where the next crack shows up โ and it won't announce itself in a press release. It'll surface in a filing nobody reads, on a Friday, when the desks have gone home.
Reading the room while the order book burns is a skill. But the room is never the whole story. The story is always in the plumbing.
Keep your eyes on the three doors. Just remember: every funnel has a drain, and the drain is always where the leverage hides.