MRRydon：I tried to work out why a CoreWeave bond y…

The Nasdaq fell 2.6% yesterday, its worst day since June, on worries about the AI spending bill. Part of the selling was pinned on Jensen Huang saying neoclouds face a scarcity of capital. But look at what he actually said: "the demand for the neoclouds is off the charts... what is missing in the middle is the working capital to fuel that growth." So the demand is there, and the problem he's describing is funding. And funding gaps close when there are signed contracts to lend against. On the same day, Nebius announced a $3.5B senior secured facility from Citi at SOFR plus 2.75%, to pay for GPUs and data centres. It also says it can borrow against more than $40B of long-term contracts with investment-grade customers. Citi signed it on 17 October, right after last week's neocloud selloff. Lenders still want this paper when there are contracts behind it. Despite Nvidia saying neoclouds face a scarcity of capital, it seems that as long as you're a credible builder with good off-takers, the capital is there.

2026-10-23 18:0060

.@GoldmanSachs says hedge fund shorts are at the 99th percentile of the last five years. The same night, Billie Eilish and Ariana Grande both went short. Introducing the Billie and Ariana Short Index. I'll be tracking how well their haircuts call hedge fund positioning from here.

2026-10-23 17:2170

I tried to work out why a CoreWeave bond yields around 15% when compute demand is this strong. TLDR: it's the price of having no hedge. A neocloud with long contracts from the most creditworthy companies in the world can borrow fine. The cost of capital jumps when you sell on-demand compute, or sell to smaller, less creditworthy buyers, because nobody can lock in what that capacity will earn. Even CoreWeave, with 98% of its revenue on committed long-term contracts, pays about 15% to borrow, because that revenue sits with a handful of customers and nobody can lock in what those GPUs earn once the contracts end. GPU futures should fix that, and it isn't a new idea. Farmers sell next year's corn forward so a price crash at harvest can't wreck them. Compute would work the same way. Buy the GPUs, deploy them, sell a strip of futures against the fleet, then sell on-demand or six-month contracts to smaller customers who put a big share of the contract value up front. Lenders are then lending against a hedged position, not a guess, and that should mean a much lower cost of capital. The behaviour already exists. Compute just doesn't have the market yet. The harder-to-finance customers aren't a niche. In my view, the open-source model hosts and startups that rent GPUs to sell tokens are already a big enough market to need this. Give it a few years and hedging a GPU fleet probably becomes as normal as a farmer selling next year's corn forward.

Oil has the crack spread. Soybeans have the crush spread. Compute needs one too. The crack spread is the margin between crude oil and the gasoline and other fuels it's refined into. The crush spread is the margin between soybeans and the soy oil and meal they're crushed into. For compute, I'd call it the crunch spread: what you earn turning a rented GPU into tokens, minus what the GPU-hour costs. The trade is simple. Rent GPUs at the on-demand price, load an open-source model, and sell the tokens through OpenRouter or something like it. Whether that's profitable is the same question a refinery asks about crude and gasoline. Today you can only see one leg in public, GPU rents. Token futures would price the other leg, and a daily token price index already exists as a start. Refineries get built when the crack spread says they'll earn. A visible crunch spread would tell builders and lenders what an hour of GPU is worth once it's turned into tokens, not just what it rents for. If token demand keeps outgrowing new capacity, I think that spread will explain why operators keep committing to more GPUs.

2026-10-23 08:3070

NVIDIA says its new DSX POD cuts AI factory deployment at a prepared site from six months to as little as 14 days. Racks, cabling, power and liquid cooling arrive pre-built in a container. But the prepared site is often the hardest part. Literally on the same day, NextDC had their grid works suspended for their M3 Melbourne data centre because of community opposition, and the next increase in power is now expected in FY28. NextDC cut its FY27 net revenue guidance to A$475M to A$500M, from A$615M to A$640M, and the shares were hammered. Still, the capacity is under contract, subject to the customer's rights, and construction of the data halls is continuing. The customer still wants it. What's missing is the power. So the bottleneck has moved. You can now truck in a Vera Rubin factory faster than you can get permission to plug it in. The builders who win the next few years are the ones who already have a grid connection or behind-the-meter power, and the community on side.

2026-10-23 03:3031

消息提醒

评论与回复

相关行情