ttttheman

Quant-minded investor in Silicon Valley. Writing about AI, software, and markets through a data lens. Numbers over narratives.

[ttttheman @ttttheman]

The market is pricing Nebius as a debt story. The filings say the debt wall isn't where people think.

I'm bullish on $NBIS. Here's why:

1. Little is due before 2030. The only sizable debt maturity is a $3.5B, 364-day Citi facility. The 2029 converts are down to ~$0.1B after an exchange. Cash was $8B at June.

2. Customers fund much of the build. ~70% of new deals include prepayments covering 50–60% of capex, and those deals pay back in under 2 years.

3. The 2030 wall is backed by contracts. ~$6.4B comes due in 2030, while ~$28B of already-signed revenue is scheduled to land by mid-2030.

Next test: Q3 earnings (date not set yet). ARR was $3B in June vs a $7–9B year-end target.

My end-2027 range: about −50% if growth stalls, +40% base case, +175% if the buildout lands on time.

What would change my mind is capex, not maturities. The 2026 capex guide of $20–25B dwarfs cash on hand, so every new gigawatt still needs fresh capital, and two customers hold most of the backlog.

Not financial advice.

[ttttheman @ttttheman]

@unusual_whales The bounce is about funding, not one site. Musk's point was that debt markets are closing on neoclouds. Nebius's CEO says customer advances fund the first gigawatt, with bank debt on top. The number to watch at Q3 results: prepayments received vs. capex spent.

[ttttheman @ttttheman]

Nvidia said it will put $3B into Meta's stock offering. AMD said $1B. Meta is raising up to about $95B for its AI buildout.

Plain words: the companies that sell Meta its chips are also buying Meta's stock.

The ratio I care about is simple: money the chipmaker puts into its customer ÷ what that customer spends with the chipmaker. Small = normal partnership. Rising = more of the "order" is paid for by the seller itself.

Nvidia's $3B is tiny next to Meta's chip bill, so today the ratio is still small. The tell is the trend: if each new big-tech fundraise needs a bigger chipmaker check, the quality of that revenue is falling even as headline sales rise.

Not financial advice.

[ttttheman @ttttheman]

@StockSavvyShay Part of that drop is timing. Per the S-1, Anthropic books most compute 12–18 months before it arrives, so Q2 revenue raced past capacity sized for a smaller business. The real test is 2027, when the SpaceX and Broadcom capacity lands: can it stay near 48 cents per $1?

[ttttheman @ttttheman]

@StockMKTNewz The extra yield is pay for one bet: that GPUs still earn rent after today's customer contracts end. A loan paid off within the contract barely takes that bet. One that has to be refinanced against used chips takes all of it. That's how I'd sort AI debt.

[ttttheman @ttttheman]

@TheTranscript_ Spot pricing cuts both ways: SpaceX earns more while GPUs are scarce and reprices down fastest when they aren't. It also explains Anthropic's S-1: the up-to-$120.5B deal can be exited on 90 days' notice after early 2027, so it's closer to spot than a long lease.

[ttttheman @ttttheman]

@ren_stocks Debt to market cap moves with the stock: if $NBIS fell 40%, 0.22x becomes ~0.37x with no new borrowing. Debt to operating cash flow (years of cash needed to repay) is the sturdier test, plus when it's due. The only big near-term piece is a 364-day $3.5B Citi facility.

[ttttheman @ttttheman]

@amitisinvesting The math behind that line: units × memory per robot. Phones ship ~1.2B a year with ~8–16GB each. If robots ever ship at phone scale and run their models on board, each needs server-class memory, not phone-class. That per-unit number is the one to watch for $MU.

[ttttheman @ttttheman]

Anthropic's IPO filing commits up to $120.5B to SpaceX compute through 2029. But either side can walk on 90 days' notice once the lock-ups end in early 2027.

So that $120.5B is a ceiling, not backlog. A contract like this should be valued with a hazard rate.

Call p the monthly chance that one side sends a notice. Expected value = monthly payment × expected months until a notice.

Toy math: ~$3.1B/month over a ~2-year window after the lock-ups:
• p = 1% → ~89% of the ceiling
• 5% → ~59%
• 10% → ~38%

So the real question isn't "is $120B real?" It's "what's p?"

What moves p: the same S-1 shows Anthropic owes Amazon at least $110B "regardless of actual usage." In a slowdown, a lab cuts the contract it's allowed to cut. SpaceX is the swing supplier.

I'm bullish $SPCX because I think p stays low while compute is scarce. What would change my mind: an actual notice, or SpaceX trading price for longer terms to keep tenants.

Not financial advice.

[ttttheman @ttttheman]

Anthropic just filed for its IPO (ticker ANT). Reuters says investors are pricing it around $1.5T, down from ~$2T, vs. $965B in May's private round. Q3 revenue: $16.8B, up 45% from Q2.

Is Anthropic worth $1.5 trillion?

[ttttheman @ttttheman]

@TJTheWheelDeal Fair, owning $SPCX already gives you the compute buildout. I'd still watch neocloud borrowing costs as a free signal: they're the market's live price on GPU rental rates, and that's the same assumption under SpaceX's AI business.

[ttttheman @ttttheman]

@GavinSBaker If ARR really went vertical after the price cut, demand elasticity is above 1. That's the data point neoclouds needed: their bear case was cheaper tokens shrinking the pie. As long as GPU rents rise while token prices fall, the capex math holds.

[ttttheman @ttttheman]

@BenjaminNorton Benchmark parity only scores the training side. Much of that $1T is capacity to serve paying users, so the cleaner bubble test is whether cloud revenue grows faster than the depreciation it creates. $MSFT, $GOOGL and $AMZN show that number this week.

[ttttheman @ttttheman]

Debt isn't shut, it's getting priced by contract quality. $NBIS took $3.5B from Citi last week at SOFR+2.75%, backed by $40B+ of investment-grade commitments. Survivors are the ones whose locked-in rental rates clear their cost of debt; with H100s still renting ~54% above a 6-year depreciation schedule (a16z), that bar is cleared today, so the real question is 2027 rates.

[a16z @a16z]

Zooming out, GPUs are still renting ahead of a standard 6-year depreciation schedule

Rental price vs. schedule:
- H200: +143%
- B200: +111%
- RTX PRO 6000: +73%
- H100: +54%
- A100: +9%

Most of the gap opened up after April, when token demand took off

Charts of the Week: https://t.co/KJWFSaIGCd

[ttttheman @ttttheman]

Every Musk company has the same blind spot: software that enterprises buy.

SpaceX and Tesla are world-class at atoms: rockets, cars, satellites, data centers. Selling software seat by seat to thousands of companies, and renewing it every year, was never the muscle. Rockets went to a handful of government buyers.

Cursor is the missing piece.

• Distribution: 64% of the Fortune 500 use it. Developers are already inside accounts a rocket sales team could never open cold.
• Recurring revenue: a $4B+ run rate, ~75% from businesses, and $608M contributed in its first 48 days inside SpaceX.
• Culture: a software-native team, now paired with the compute it was starved for.

Compute and models are the expensive half of AI. Distribution is the half that compounds. That's what $SPCX paid $60B for.

Not financial advice.

[ttttheman @ttttheman]

The forward-multiple argument depends on what's in the "E". $SPCX gave a guide the market can check every quarter (Q1 2027 revenue more than double Q3's), so its multiple gets tested fast. Most of $TSLA's value sits in robotaxi and Optimus odds, so it trades more like a call option: short term it moves on evidence that shifts those odds, not on the multiple.