The Moat Moved to the Treasury

A company in Michigan is borrowing against its bitcoin to keep AI servers running — roughly 100 coins pledged at 4.5 to 5 percent variable interest to fund a data-center campus ^[1]. A crypto exchange is lending $100 million in stablecoins against GPU fleets, non-recourse, the chips themselves as collateral ^[2]. A money-losing company just raised $5.75 billion in convertible bonds at a 0.5 percent coupon ^[3]. SoftBank tried to borrow against its OpenAI stake and found no lender willing to price a private company's shares ^[4]. And middlemen are selling access to OpenAI and Anthropic through special-purpose vehicles charging 20 percent fees, while the companies themselves warn such holdings may end up worthless ^[5]. The chief executive of Permian Labs, one of the firms arranging this kind of debt, put a name to it.

Bullish recognizes that compute is becoming a credit market in its own right. — David Choinière

What connects these instruments is that the underlying math no longer works. Morgan Stanley ran the numbers on a fully optimized data center using the latest Nvidia chips: $25 billion a year to rent, $23 billion in output at base-case pricing — and the bank expects prices to keep falling ^[6]. The models themselves are getting cheaper faster than anyone can recoup the cost of building them: OpenAI cut its Luna model's fees by 80 percent, and analysts now describe frontier models as "interchangeable commodities" ^[7][8]. All five cloud giants — Amazon, Google, Meta, Microsoft, Oracle — are posting rising profits alongside falling free cash flow, because the infrastructure spend is consuming the cash operations generate ^[9]. Alphabet just recorded its first-ever negative free cash flow while raising its 2026 capital-spending guidance to $205 billion ^[10]. Convertible issuance hit a record $147 billion this year, up 50 percent, as investors accept near-zero coupons in exchange for a shot at equity conversion ^[11]. Conventional debt and equity markets left this gap because the risk was too opaque to price — so the industry invented instruments that don't require pricing it. If the models are commodities and the economics are negative, the moat is no longer in the lab. The financing is where the competition now happens, and it runs in circles. Nvidia put $2 billion into Nebius for an 8.3 percent stake; Nebius uses the money to buy Nvidia GPUs and rents the compute to Meta ^[12]. SoftBank sold its entire Nvidia stake and $9.17 billion in T-Mobile shares to fund a single private bet it couldn't get lenders to price ^[4][13]. Sovereign wealth funds are sliding into the capital stack where venture equity used to sit — Abu Dhabi's MGX and Mubadala are in talks on a $40 billion data-center deal ^[14]. Each actor is financing the next one's purchase, and the chain only holds as long as someone keeps lending. Sam Altman's defense of the spending is that the risk of running out of compute outweighs the risk of having too much.

We believe the risk to OpenAI of not having enough computing power is more significant and more likely than the risk of having too much. — Sam Altman

The dissent is just as loud. Jim Cramer warns that if the market stops funding data centers and the companies lack the cash to sustain them, the result would be a rerun of 2000; NYU's Aswath Damodaran argues the market hit peak AI months ago, with marginal returns on invested capital already dropping ^[10]. None of this means every player is drowning. Anthropic posted its first operating profit, $559 million on $11.6 billion in quarterly revenue, overtaking OpenAI on the top line ^[8]. CoreWeave sits on a $100 billion contracted backlog ^[15]. Micron's chief executive says data-center demand exceeds supply by 50 percent ^[16]. These are the narrow path — the companies that can cover infrastructure through revenue rather than financial engineering. But even Anthropic's profitability arrives while the broader market's instruments grow stranger, and the price war that made models cheap is the same force squeezing the margins that would pay for the servers. The instruments proliferate because none of them resolves the underlying math — each only extends the runway. A convertible at 0.5 percent doesn't close a multi-billion-dollar gap; it moves the gap a few years out. So the next instrument is stranger, and the next. That someone is borrowing against bitcoin to keep AI servers running in Michigan is not a footnote. It is the tell.