SpaceX spent $2.35 of capex for every dollar of revenue — and its own share price just pushed a bigger unlock into December

Valuation — an unusual problem, because there are no earnings to divide by

At the $115.98 after-hours print the market values SpaceX near $1.65T on ~13.17B shares. There is no P/E: the company lost $541M in the quarter and $9.36B over the trailing twelve months. So the multiples are all revenue-based, and they are extreme by any standard.

The forward multiple is the interesting one. At 32.4× the market is already underwriting roughly $51B of forward revenue — a 63% increase on the current run rate. That is not a sceptical price. It is a price that assumes the cloud contracts ramp on schedule and the Starlink subscriber curve holds.

Context matters on the drawdown, though, and it cuts both ways. The stock is 48.6% below its $225.64 June high and 14.1% below the $135 IPO price — so a great deal of enthusiasm has already come out. It also rose 9.4% on the day of the print and gave back 7.5% after hours, which tells you the market liked the revenue and disliked the capex, in that order.

The capex — the number that moved the stock

The market liked the revenue and disliked this, in that order. Capex of $18,369M against $7,814M of revenue is a ratio of 2.35×, and it beat the ~$13.2B consensus by roughly $5B — which is why a stock that rose 9.4% during the session gave back 7.5% after hours.

Almost all of it is one segment. AI capex went from $749M to $15,828M in a year — a 21× increase — while AI revenue went from $737M to $2,561M. At $15.8B of spend against $2.6B of quarterly revenue, the build is running years ahead of the business it serves. That is a deliberate choice, not an accident: nameplate compute went 0.4 GW → 1.0 GW → 1.4 GW across three quarters, and management is buying capacity for contracts that ramp later.

Some of that asset isn't even depreciating yet. Construction-in-progress went $4.6B → $12.6B, and the 10-Q says it is "primarily comprised of ongoing construction and expansion of the facilities and equipment as well as AI infrastructure that has not yet been placed in service." So forward depreciation rises sharply as that lands — which is precisely the line Adjusted EBITDA adds back.

The other two segments are, by comparison, ordinary. Connectivity spent $1,367M to generate $4,291M and $1,656M of operating income. Space spent $1,174M against $962M of revenue, which is heavy but is a launch business building the vehicle that carries its own constellation. Neither is what the market reacted to.

Some of that asset isn't even depreciating yet. Construction-in-progress went $4.6B → $12.6B, and the 10-Q says it is "primarily comprised of ongoing construction and expansion of the facilities and equipment as well as AI infrastructure that has not yet been placed in service." So forward depreciation rises sharply as that lands — which is precisely the line that Adjusted EBITDA adds back.

The customer

Note 3 of the 10-Q discloses two customers above 10% of consolidated revenue. Customer A at 18.3% spans all three segments and is long-standing — it was 20.9%, 24.2% and 25.2% of revenue in 2025, 2024 and 2023, so it predates the cloud business entirely. Customer B at 19.5% is new: it did not clear 10% in any prior period, it relates only to the AI segment, and at ~$1.52B it is roughly 95% of the $1.6B of incremental cloud revenue the release attributes to the new agreements.

Put plainly: essentially all of the AI segment's revenue growth this quarter came from one counterparty that was not material three months ago.

The company doesn't name it in the 10-Q, but it isn't a secret either — the IPO prospectus describes an Anthropic arrangement of roughly 325,000 GPUs at about $1.25B/month running to May 2029, and the CFO named Google and Anthropic on the call, with the Google agreement ramping from October. Note 17 lists only Tesla and Valor as related parties, so the cloud customers are arm's-length — a reading worth ruling out explicitly, and it is ruled out.

The tension the filings don't resolve. The risk factors say cloud agreements "may be terminated by either party upon 90 days' notice." The prospectus describes a term running to 2029. Both are SpaceX's own words. The honest position is that the contracted term is long and the cancellation right is short, and only one of those gets tested if the AI capex cycle turns.

The financing — a director's firm holds a third of the debt

This is the disclosure most worth reading slowly. Of $39,364M of total debt, $13,329M — 33.9% — is owed to Valor Equity Partners, whose founder, CEO and Chief Investment Officer Antonio Gracias serves on SpaceX's board. It comes from three equipment lease agreements for AI infrastructure hardware (November 2025, January 2026, April 2026, the last through CTC Property LLC), all of which failed sale-leaseback accounting — meaning the assets stay on SpaceX's balance sheet and the proceeds book as debt. SpaceX guarantees the lessees' payments. The balance nearly tripled in six months, from $4,507M, and it carries $327M of the quarter's $629M interest expense — 52%.

None of that is hidden; it is all in Notes 9 and 17. And the terms are not obviously off-market: the 10-Q's Other Financings paragraph states these arrangements carry an average fixed interest rate of 5.9%, against 5.855% weighted average on the public investment-grade notes.

There is, however, a gap the filing does not reconcile. $327M of quarterly interest against an average related-party balance of roughly $8.9B annualises to the mid-teens, not 5.9%. Failed sale-leaseback accounting can produce interest expense that exceeds a stated coupon, so there are innocent explanations — but the filing gives balances and interest and a rate that don't sit together, and does not explain why. It is the single thing I would most want asked on the next call.

Risk — each isolated, do not blur

1. The capex/funding gap (dominant). −$25.0B of H1 free cash flow, capex nearly doubling sequentially, ~$105B of liquidity — roughly 18 months at the Q2 exit rate before the company needs the capital markets again. It has investment-grade access today. That is a condition, not a guarantee.
2. Customer concentration. 37.8% of revenue in two customers; ~60% of the AI segment in one that appeared this quarter.
3. The depreciation assumption. 90% of Adjusted EBITDA is D&A and stock comp, $12.6B of asset is not yet in service, and the 10-Q publishes no useful-life table. If GPU lives are shorter than assumed, the adjustment does the damage on the way back.
4. Starlink unit economics. ARPU −22% and revenue per subscriber −28% while subscribers doubled. Management attributes it to international mix and cheaper plans; the filing gives no like-for-like or US-only ARPU, so a price cut in mature markets cannot be ruled out.
5. The 2027 obligation wall. $22,244M of the $27,955M of non-cancelable commitments falls in 2027.
6. Governance. A third of the debt and half the interest expense run to a board member's firm, with an unreconciled implied rate.
7. Operational and legal. The NAACP is seeking a preliminary injunction against the gas turbines powering Colossus II under the Clean Air Act; separately, multiple class actions over Grok's image generation, with a $354M litigation accrual.

The lockup — the stock's weakness made the overhang bigger, not smaller

The mechanics here are genuinely non-obvious and worth getting right, because they are two days away.

On August 6, up to 911.5M shares (excluding affiliates) come free — about 1.43× the entire 638.9M-share IPO float, roughly $106B at the after-hours price. That much was widely reported and is correct.

What was less noticed: a second tranche of 455.8M "Additional Release Shares" was also scheduled for that date — but only if the stock closed at least 30% above the $135 IPO price ($175.50) on five of the ten trading days ending on the earnings date. Over that window (July 22 – August 4) the closes ran $108.37 to $118.24. The highest close was 32.6% below the trigger. Zero of ten days qualified.

So those shares do not release now. Under the prospectus's alternative branch they roll to December 8, 2026 — where up to 797.6M shares come free instead of 328.4M. The weak share price didn't avoid the dilution. It deferred it, and made the December block 469M shares larger.

Two caveats: every figure is an "up to," assuming no option exercises or RSU settlements after March 31; and the lockup, price condition included, is waivable with Goldman Sachs' written consent.

Time horizon & position sizing (kept separate)

Horizon. The next 90 days are dominated by mechanics rather than fundamentals — the August 6 unlock, the October ramp of the Google agreement, the Cursor close in Q3, and a Q3 print that will show whether Customer B's revenue is a step or a spike. The actual thesis resolves on a 3–5 year horizon: does 1.4 GW of compute and a doubling Starlink constellation generate returns above a depreciation base that is compounding faster than revenue?

Sizing considerations (not a recommendation). SpaceX is the highest-beta expression of the same AI-capex cycle that drives Amazon, Microsoft and Nvidia — it does not diversify an AI-infrastructure basket, it concentrates it, and it adds two risks the others do not carry: a single-customer AI revenue base and a funding requirement that returns in roughly 18 months. Against that, Connectivity is a real, growing, profitable business that would be worth something substantial on its own, and the Space franchise has no competitor at its cadence. A position here is not one bet; it is three businesses with one balance sheet, two of which lose money at the operating line.