# SpaceX Q3 2026 Earnings: The AI Segment Turned a Profit Before Starship Did — InvestMoat Research

_SpaceX's AI segment made its first operating profit in Q3 2026 while launches fell, which undercuts my October argument that it needed Starship cadence. I think the profit comes from renting scarce GPUs at shortage prices on chips anyone can buy, and that margin fades when the shortage does._

_Published: October 20, 2026 · Last reviewed: October 20, 2026 · Canonical page: https://www.investmoat.com/research/spacex-q3-and-the-ai-profit-before-the-rocket_
_Tags: AI Infrastructure, Space, Catalyst, Business Models_
_Covers: SPCX, CRWV, NBIS, GOOGL_

## Summary

SpaceX's Super Intelligence segment, the AI business, reported $431 million of operating income in Q3 2026 after a $1.26 billion loss in Q2, while launches fell to 35 from 44 a year earlier. On October 3 I argued that the AI story could only borrow the rocket's moat through Starship flying often, and said that a profit on ground compute before that happened would mean the compute business stands on its own. It has, for one quarter. The profit came from revenue tripling, from Cursor, and from renting compute in a shortage, with management saying it rents at effectively the spot rate. Against CoreWeave, Nebius and Google Cloud, SpaceX's compute margin sits above the GPU landlords and below the one cloud that owns its chips, and I read the gap over the landlords as pricing that reverses when the shortage eases. The October falsifiable claim stays as written, and this piece adds a narrower one about the segment's margin through 2027.

## What would prove this wrong

If SpaceX's Super Intelligence segment posts an operating margin of 20% or more in every quarter of 2027, computed from each 8-K as segment operating income plus acquisition-related share-based compensation, divided by segment revenue, while Cursor stays at or under 15% of segment revenue wherever SpaceX discloses it, then I'm wrong that the Q3 2026 profit is a shortage price that fades as GPU supply catches up.

_Status as of October 20, 2026: **holding**._

---

SI operating income, Q3 2026: $431M · Launches, Q3 2026: 35 · Q4 2026 capex guide: $35–40B · SPCX price: (live) · SPCX composite: 83

SpaceX reported its third quarter after the close on October 20, and the number that moved the stock was in the AI business, which it now calls Super Intelligence, or SI. The segment made $431 million of operating income on $7.56 billion of revenue, a quarter after losing $1.26 billion. Group revenue was $14.6 billion and net income $4.4 billion, and management guided Q1 2027 revenue to $33 to $37 billion. SPCX rose about 6.6% after hours, to $157.59 from a $147.79 close.

## What I said on October 3, next to what Q3 showed

[I wrote about SpaceX on October 3](/research/spacex-and-the-cadence-under-the-ai-story), after Starship Flight 14, and argued that the AI segment could only borrow the rocket's moat through Starship flying often. In that piece's counter-case I wrote that if the AI segment reached an operating profit on ground compute before Starship flew at cadence, the compute business would stand on its own, and that it would then be valued like the rest of the compute group. Seventeen days later the first half of that has happened, for one quarter, in a quarter when SpaceX flew fewer rockets than a year ago.

**The October 3 article's claims, set against the Q3 2026 release and call.**

| Said on October 3 | What Q3 2026 showed | Where that leaves it |
| --- | --- | --- |
| The AI segment reports an operating loss once depreciation is charged (Q2 2026) | Operating income of $431M on $7.56B of revenue, after $1.2B of Cursor-related stock compensation | Wrong on timing: it turned profitable on ground compute |
| If the AI segment profits before Starship flies at cadence, the compute business stands on its own | Launches fell to 35 from 44 a year earlier; two were Starship flights | That condition has now happened once |
| Starship's 2027 target is weekly or twice weekly | Musk: weekly at some point in the first half of 2027, twice weekly or more by year-end | Unchanged; still the number I'd count |
| Own silicon through Terafab would give the compute a moat the way TPUs do for Google | Musk: a working test chip around May or June 2027, volume around Q3 2028, as internal targets | Now has dates, still two years out |
| The falsifier: two straight profitable AI quarters in 2027 with fewer than 15 Starship flights | Q3 2026 is outside the window | Holding, and the first condition now looks plausible |

_Figures as of Q3 2026 — SpaceX 8-K and earnings call, October 20, 2026. Source: [SpaceX Reports Third Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm); [SpaceX Q3 2026 Earnings Call Transcript](https://www.benzinga.com/news/26/10/62484605/spacex-q3-2026-earnings-call-transcript); [Earnings call transcript: SpaceX tops Q3 2026 estimates, shares jump after hours](https://in.investing.com/news/stock-market-news/earnings-call-transcript-spacex-tops-q3-2026-estimates-shares-jump-after-hours-93CH-5643891)._

## The AI segment went from a $1.26 billion loss to a profit in one quarter

**SpaceX's AI segment by quarter, as reported. SpaceX renamed it Super Intelligence in Q3 2026 and presents the earlier quarters under the new name. Cursor was acquired on August 14, 2026.**

| Quarter | Revenue | Operating income | Adj. EBITDA | D&A | Capex |
| --- | --- | --- | --- | --- | --- |
| Q3 2025 | $813M | $(1,722)M | $(428)M | $922M | $4,760M |
| Q1 2026 | $818M | $(2,469)M | $(609)M | $1,493M | $7,723M |
| Q2 2026 | $2,561M | $(1,257)M | $1,146M | $1,885M | $15,828M |
| Q3 2026 | $7,560M | $431M | $4,953M | $2,473M | $16,444M |

_Figures as of Q3 2026 — SpaceX 8-K, October 20, 2026. Source: [SpaceX Reports Third Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm); [SpaceX Reports Second Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/0001181412/000162828026052515/earningsreleaseq22608042.htm)._

**SpaceX AI (Super Intelligence) segment margins by quarter. The operating line adds back acquisition-related stock compensation, which was zero before Cursor closed in Q3 2026, so it equals the reported margin in earlier quarters; as reported, Q3 2026 was 6%. SpaceX's releases don't break out Q4 2025. This is the margin the falsifiable claim tracks.**

| Period | Operating margin before acquisition SBC | Adjusted EBITDA margin |
| --- | --- | --- |
| Q3 2025 | -211.8% | -52.6% |
| Q4 2025 | — | — |
| Q1 2026 | -301.8% | -74.4% |
| Q2 2026 | -49.1% | 44.7% |
| Q3 2026 | 21.8% | 65.5% |

Q3 2026: Cursor closed Aug 14.

_Figures as of Q3 2026 — SpaceX 8-K, October 20, 2026. Source: [SpaceX Reports Third Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm); [SpaceX Reports Second Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/0001181412/000162828026052515/earningsreleaseq22608042.htm)._

Three things produced the profit, and I don't think the rocket was one of them. Revenue roughly tripled from Q2 while depreciation rose by about a third, from $1.89 billion to $2.47 billion. The segment spent $40 billion on capex in the first nine months of 2026, and my inference, which the release doesn't confirm, is that much of Q3's $16.4 billion wasn't running long enough to carry a full quarter of depreciation. Cursor brought in $608 million between the August 14 close and quarter end, which is software revenue on top of compute; the CFO put the rest of the segment, solutions and infrastructure, at about $7.0 billion. And there's pricing. On the call Musk said SpaceX would rent at "effectively the spot market rate as opposed to sign these long term deals," and said even H100s now rent for more than they did when they were new. The CFO said that "at current pricing, new compute deployments are generating a payback of less than a year."

The rocket side went the other way. SpaceX flew 35 missions in Q3, against 38 in Q2 and 44 a year earlier, and the nine-month total implies about 40 in Q1, so Q3 was the lowest of 2026's three quarters. Two of the 35 were Starship flights, including Flight 14, which reached orbit with 26 Starlink V3 satellites. The space segment lost $7 million. So the profit didn't come from cheaper launch.

## I had the timing wrong, and the question I care about is still open

> **What I got wrong, and what one quarter can't show** — I expected the AI segment to keep losing money until Starship gave it something a rival couldn't buy. That was wrong on timing. In a GPU shortage, a company with power, buildings and chips in hand can profit on rented compute without any help from the rocket. What one quarter can't show is whether the profit comes from a lower cost per unit of compute or from selling scarce capacity at a scarce price. While prices are high the two look alike, and I'd only expect them to come apart once prices fall.

## SpaceX's compute margin sits between the GPU landlords and Google Cloud

**SpaceX's AI segment next to the compute businesses it gets compared with, each in its latest reported quarter. None is a pure compute line: SI includes Cursor and Grok, Google Cloud includes Workspace, and Nebius is reported only for the whole group, which is mostly its AI cloud. CoreWeave, Nebius and Alphabet report their third quarters later this month or in November.**

| Business | Quarter | Revenue | GAAP operating margin | Adj. EBITDA margin | What's disclosed about selling capacity |
| --- | --- | --- | --- | --- | --- |
| SpaceX Super Intelligence | Q3 2026 | $7.56B | 6% (22% before acquisition-related SBC) | 66% | Stated aim to use about half itself and rent the rest at spot; the $63B backlog is for the whole company and isn't broken out by segment |
| CoreWeave | Q2 2026 | $2.58B | (2)% | 59% | Revenue backlog of about $104B |
| Nebius (group) | Q2 2026 | $582M | (30)% | 41% | About $6.0B of deferred revenue at June 30 |
| Google Cloud | Q2 2026 | $24.8B | 36% | Not reported | Runs on its own TPUs as well as NVIDIA |

_Figures as of Q3 2026 for SpaceX; Q2 2026 for the others. Source: [SpaceX Reports Third Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm); [SpaceX Q3 2026 Earnings Call Transcript](https://www.benzinga.com/news/26/10/62484605/spacex-q3-2026-earnings-call-transcript); [Earnings call transcript: SpaceX tops Q3 2026 estimates, shares jump after hours](https://in.investing.com/news/stock-market-news/earnings-call-transcript-spacex-tops-q3-2026-estimates-shares-jump-after-hours-93CH-5643891); [CoreWeave Reports Strong Second Quarter 2026 Results](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx); [Nebius reports second quarter 2026 financial results (Form 6-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/0001513845/000110465926094568/tm2622968d1_ex99-1.htm); [Alphabet Announces Second Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000066/googexhibit991q22026.htm)._

On operating margin, SpaceX's AI segment beats both landlords and trails Google Cloud by a wide margin, even with the Cursor-related stock compensation added back, and it's ahead of CoreWeave on adjusted EBITDA too. I think the ordering has a fairly plain explanation. CoreWeave and Nebius carry large contracted books, much of it signed before this year's prices, and they pay interest on the debt that bought the capacity. SpaceX says it prices its rented capacity at today's spot rate and funded much of its build with Starlink cash and IPO money. Google's margin rests on its own TPUs, a cost position no buyer of NVIDIA chips has, though Workspace also flatters it.

**The four compute businesses on the pillars that separate a cost position from a pricing position. Scale economics is the column that matters for compute margins once prices fall.**

| Ticker | Scale Economics | Transaction Embedding | Bundling | Learned Interfaces |
| --- | --- | --- | --- | --- |
| **Sells compute at spot** |  |  |  |  |
| [SPCX](https://www.investmoat.com/stocks/spacex) | strong | na | intact | intact |
| **Sells compute on long contracts** |  |  |  |  |
| [CRWV](https://www.investmoat.com/stocks/crwv) | weakened | weakened | weakened | na |
| [NBIS](https://www.investmoat.com/stocks/nbis) | weakened | weakened | weakened | na |
| **Owns its chips** |  |  |  |  |
| [GOOGL](https://www.investmoat.com/stocks/google) | strong | intact | strong | strong |

The framework rates SpaceX strong on scale economics, but the stock page earns that rating on Falcon's cost per kilogram and in-house Starlink production, not on compute. CoreWeave and Nebius are rated weakened on the same pillar because they buy the same chips, power and capital as everyone else, and Google is rated strong because of the TPU. My reading, which the framework doesn't score separately, is that SpaceX's compute sits with the landlords on that question until Terafab produces chips. Musk gave a working test chip around May or June 2027 and volume production around Q3 2028, and called those internal targets with large error bars. Until then SpaceX's compute buys NVIDIA parts on the same terms as the two companies it currently out-earns.

## Selling at spot ties SpaceX's margin to the shortage in both directions

A long contract and a spot sale split the same risk differently. CoreWeave locked in prices that look cheap today, and it keeps getting those prices if rental rates fall. By management's own description, SpaceX rents what it doesn't use itself at the going rate, which is why its margin is higher now, and it will take the going rate later too. That's a stated policy rather than a reported mix: the CFO also talked about current customer contracts for SI compute and new cloud agreements signed in the quarter, and SpaceX doesn't say how much of SI's revenue is contracted.

- The capex wave goes in at whatever price holds when it's switched on. SpaceX guided Q4 capex to $35 to $40 billion, more than twice its Q3 revenue, and the CFO tied it to capacity coming online this quarter. That capacity depreciates through 2027, and a one-year payback at today's rental price becomes a much longer one if prices fall while it's running.
- Depreciation keeps climbing. The segment charged $2.5 billion in Q3 against $16.4 billion of capex in the same quarter, and each quarter of this build adds to the charge, so revenue has to keep running ahead of it.
- Any rival with GPUs can charge the same spot price. Meta said in July it would rent out spare capacity, the hyperscalers sell on demand, and the landlords sell whatever isn't under contract. None of that undercuts SpaceX while demand runs ahead of supply. When supply catches up, I'd expect spot prices to fall before contract prices do.

## The shortage may last longer than any of this matters

The bull reading of this quarter is strong, and I'd put it like this. A payback of less than a year means rental prices could halve and SpaceX would still earn back each new rack in about two years, on hardware that runs for several. Musk said a six-month payback next year was "not impossible," while saying he wasn't predicting it. When H100s, a chip from 2022, rent for more now than they did when they were new, the shortage isn't marginal. And compute shortages have run longer than skeptics expected at every point since 2023.

The second part of the bull case is that I'm treating SpaceX's compute as a landlord when it's also its own biggest tenant. Grok and Cursor use the capacity SpaceX builds, so about half the output has a captive buyer, and Cursor's $608 million in under seven weeks is revenue a GPU landlord doesn't have. Speed counts as well. SpaceX reached 1.7 gigawatts of nameplate compute by September 30, and getting power and buildings online faster than anyone else is a real advantage when power is the binding constraint. A bull would say building speed is a moat of its own, and that it needs neither the rocket nor a custom chip.

I think most of that is right as a description of 2026 and early 2027, and it's why I'm not calling the AI segment a bad business. It's earning real money, and if the shortage lasts through 2027 it'll earn a lot more. Where I still disagree is on durability. A faster build helps while capacity is scarce and turns into a cost once it isn't, because the fastest builder owns the most capacity bought at peak prices. A captive tenant helps as long as Grok and Cursor hold their markets, and the SpaceX stock page already records Cursor losing share in AI coding. Faster building and a captive tenant don't lower the cost per unit of compute, and that's what keeps a margin when prices fall. Google has it through the TPU, and SpaceX won't until Terafab ships.

## My October claim stays open, and I'm adding a margin test

The October 3 piece's falsifiable claim was that SpaceX's AI segment posting positive operating income in two consecutive quarters of 2027, while Starship flies fewer than 15 orbital missions that year, would show the compute business earns a return without the rocket. Q3 2026 doesn't count toward it, because it's in 2026, but it makes the first condition plausible. I'm leaving that claim as written, since changing it after one good quarter would be moving the goalposts, and I've added a revision to that article pointing here.

What I'd add is the narrower test at the top of this page. I'd track the segment's operating margin with acquisition-related stock compensation added back, about 22% in Q3, through each quarter of 2027 alongside the Starship flight count. Because Cursor's software revenue could hold the margin up while compute prices fall, the test only counts while Cursor stays at or under 15% of segment revenue, where SpaceX discloses it. If the margin holds at 20% or more all year while the Q4 build depreciates, I'm wrong that it's a shortage price. If it falls as that build depreciates and as the landlords' contracts reset, the October argument holds, and SpaceX's AI business is a well-run compute operation on bought chips that gets the rocket's moat only once Starship flies often enough to make orbital compute cheaper than a data center on the ground. Musk said on the call he expects Starship to fly weekly in the first half of 2027 and twice weekly or more by year-end, so the flight count and the margin should both be visible by mid-2027.

> **What would break this read** — First, the SI operating margin with acquisition-related stock compensation added back holding at 20% or more in every quarter of 2027, with Cursor no more than 15% of segment revenue, which would mean the margin survived the 2026 build going into service. Second, Terafab producing working chips on Musk's timeline, which would give SpaceX's compute a cost position through silicon the way the TPU does for Google. Third, SpaceX moving most of its rented capacity onto long contracts at today's prices, which would lock in the shortage margin for the length of those contracts and shrink the spot exposure I describe here.

For other AI infrastructure names reporting this month, I'd read a profit earned during the shortage the same way: it shows the business can make money at today's prices, and it says little about what the margin looks like when GPUs stop being scarce. For SpaceX's AI segment, my answer to that for now is the same as for the landlords.

## Sources

1. [SpaceX Reports Third Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm) — Space Exploration Technologies Corp. via SEC EDGAR, October 20, 2026 (filing)
2. [SpaceX Reports Second Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/0001181412/000162828026052515/earningsreleaseq22608042.htm) — Space Exploration Technologies Corp. via SEC EDGAR, August 4, 2026 (filing)
3. [SpaceX Q3 2026 Earnings Call Transcript](https://www.benzinga.com/news/26/10/62484605/spacex-q3-2026-earnings-call-transcript) — Benzinga, October 20, 2026 (transcript)
4. [Earnings call transcript: SpaceX tops Q3 2026 estimates, shares jump after hours](https://in.investing.com/news/stock-market-news/earnings-call-transcript-spacex-tops-q3-2026-estimates-shares-jump-after-hours-93CH-5643891) — Investing.com, October 20, 2026 (transcript)
5. [CoreWeave Reports Strong Second Quarter 2026 Results](https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx) — CoreWeave, August 11, 2026 (press-release)
6. [Nebius reports second quarter 2026 financial results (Form 6-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/0001513845/000110465926094568/tm2622968d1_ex99-1.htm) — Nebius Group N.V. via SEC EDGAR, August 12, 2026 (filing)
7. [Alphabet Announces Second Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000066/googexhibit991q22026.htm) — Alphabet Inc. via SEC EDGAR, July 22, 2026 (filing)

---

Scores are computed deterministically from each asset's data file: composite = Moat^0.40 × Growth^0.30 × Valuation^0.30, scaled to 0–100.

Full analyses for every name above: https://www.investmoat.com/stocks. Site index for agents: https://www.investmoat.com/llms.txt

InvestMoat is an open-source research and education framework. Nothing here is financial advice.