re: Nebius - NBIS - AI Infrastructure Company

quote:

---

Starting to feel like another bubble bursting.

---

Software has had a nice couple of days.

maybe when it hits $1200 ill think about it

appears nature is healing with Bloom and STX's reports

SK reports at 7 central and that has potential to send us to the moon or the shadow realm

current contracts expiring (old rates)

current spot prices high (market rates)

next round of contracts will be higher (future earnings)

price is going up

Added today at 165. My first buy in a long time. Also sold a 150 put, if it exercises I won’t be sad at all.

quote:

---

Can you explain this to me like I’m 5 and only used to reading pop-up board books?

---

Claude said this:

THE ONE-SENTENCE VERSION

Bond investors are getting nervous that Amazon, Microsoft, Google, and Meta will have to borrow enormous sums to build AI datacenters. Baker thinks they're wrong, because the price of renting AI compute today is roughly double what those companies locked in on older contracts, and when those old contracts expire and reprice, cash will flood in fast enough to pay for the buildout.

---------------------------------------------------------------------------------------

WHAT'S SPOOKING THE MARKET

"Credit spreads widening" means the extra interest rate a company has to pay over what the US government pays to borrow. When that gap grows, lenders are signaling more perceived risk. Hyperscaler bonds have been widening, which is unusual for companies with fortress balance sheets. The worry is that AI capex (money spent on datacenters, chips, and power) is so large it will outrun the cash these companies actually generate, forcing them into heavy debt.

BAKER'S MISSING PIECE: TWO PRICES FOR THE SAME GPU

There are two ways to buy AI compute. You sign a multi-year contract at a fixed rate, or you rent on the spot market at whatever it costs today. Right now spot is at least 2x contract rates. That tells you demand is far outstripping supply.

The implication cuts two ways. Anyone who signed a cheap contract in 2024 or 2025 is "overearning." They locked in below-market pricing and are effectively sitting on a windfall. The hyperscalers who sold them that capacity are "underearning." They're serving today's ferocious demand at yesterday's prices.

As those contracts expire and renew at current market rates, hyperscaler revenue and cash flow reprice upward without them lifting a finger. Baker says multiple private companies have already said publicly that they expect to pay 2x more per GPU on renewal.

That's why he models operating cash flow growth accelerating from 31% in Q1 2026 to 50% in Q2, and continuing from there. Wall Street consensus has it decelerating in Q3. If he's right about repricing, consensus is structurally too low.

THE MATH ON WHETHER THEY NEED TO BORROW

Datacenters are now measured in gigawatts of power rather than square feet, because power is the binding constraint. The industry rule of thumb is roughly $60 billion of capex per gigawatt, all in.

Capacity added in 2028 (hyperscalers plus "neoclouds" like CoreWeave and Nebius): 25 to 35 GW
Implied capex at $60B per GW: $1.5 trillion to $2.2 trillion
Consensus operating cash flow: $1.3 trillion to $1.4 trillion
Funding gap: $100 billion to $700 billion

That gap is the thing bond investors are pricing. Baker's rebuttal has two layers. First, the cash flow number gets revised up materially as contracts reprice, so the gap shrinks or disappears. Second, even if the gap is real, $100 billion to $700 billion is less than one "turn" of leverage, meaning less than one year's worth of operating earnings in additional debt. For companies this size, that's trivially manageable.

He also mentions Nvidia and Broadcom "credit wrappers." That's shorthand for chip vendors backstopping customer financing, whether through guarantees, vendor financing, or equity stakes. Since both throw off enormous free cash flow, their involvement props up the creditworthiness of smaller buyers.

THE DEMAND EVIDENCE

He points to OpenAI, Cursor, Grok, and open-source inference providers accelerating over the last two months, and Anthropic growing extremely fast while probably generating positive free cash flow. His point is that revenue at the application layer is real and growing, so the compute being built is getting consumed rather than stranded. He's arguing this is what the widely-circulated Bank of America chart leaves out. That chart compares hyperscaler free cash flow against semiconductor free cash flow, essentially showing money flowing from cloud companies to chipmakers.

THE CDS CAVEAT

Credit default swaps are insurance contracts on a company's debt. Their prices are often cited as a clean read on credit risk. Baker's warning is that the CDS market is thin and easy to push around. A classic trade during the 2008 crisis was to short a company's stock, then buy CDS on it to make the credit market look like it was flashing danger, which validated the short. He's saying don't treat widening CDS as independent evidence of anything.

WHAT HE THINKS THE ACTUAL RISK IS

Not money. Electricity. Getting power generated, transmitted, and physically connected to the buildings is genuinely hard, and it's the thing that could push the whole schedule to the right.

WHERE I'D PUSH BACK

Three soft spots worth holding onto.

1. Spot pricing is the most volatile signal in the stack. It's high because supply is short right now. If 2027 to 2028 capacity lands roughly on schedule, spot could collapse toward contract, and the repricing tailwind vanishes with it. He's extrapolating a scarcity premium into a durable one.

2. His argument requires demand to keep compounding through the repricing. Customers paying 2x renewal rates need end-market revenue that justifies it. Anthropic and OpenAI growth is real, but a lot of enterprise AI spend is still experimental budget.

3. He waves off CDS as manipulable. That's fair as far as it goes, but cash bond spreads widened too, and those are much harder to push around. He doesn't really address that.

His core insight is legitimate and underappreciated, though. If you're modeling hyperscaler cash flow off contracts signed at 2024 prices, you're modeling a business that no longer exists at that price. That's a real gap in consensus.

Oil rising

Yeah looks like the beating continues.