Updated 2026-11-16. Refreshed weekly (see How this page stays current at the bottom).
What it is (one line): the idea of running AI compute on satellites instead of on the ground, which the Moonshots with Peter Diamandis panel has argued about in 57 indexed episodes and had no page for.
Why it is in this archive
Gate: cited in 57 episodes. Promoted 2026-10-24 off the Moonshots #305 decomposition, where the SpaceX earnings call makes the AI segment the company's largest business and the panel keeps placing that compute in the wrong place.
The one fact that matters (EP305, 23 Oct)
SpaceX's AI segment is on the ground today. Elon Musk, on the 20 Oct call: "fairly confident of being able to bring an incremental 10 gawatt of compute online in 2027", with "roughly… two and a half three gigawatts" by the end of this year (0:12:58). Dave Blundin then says "if you put 10 gawatts of of inference in space, but it could be up to 20" (0:15:49), and Peter Diamandis corrects him on tape: "It's all terrestrial right now." (0:16:09). The company's own numbers agree: of $19.6B of Q3 capex, $16.4B went to AI, and the compute is the Memphis (Colossus) build-out, not satellites (8-K, 20 Oct, on SpaceX).
What the panel has said, in order
| date | who | what | ts |
|---|---|---|---|
| EP272 (19 Jul 2026) | Sam Altman (clip) | "I honestly think the idea with the current landscape of putting data centers in space is ridiculous… Orbital data centers are not something that's going to matter at scale this decade." His reasons: launch cost against the cost of power on Earth, and "how you're going to fix a broken GPU in space" | 1:41:05–1:41:38 |
| EP272 | Alexander Wissner-Gross | calls it "an obvious conflict of interest"; "I think the crossover is going to happen." Elon's messaging is "2 to 3 years"; other analyses put the unit-economics crossover "by the early 2030s. I'm not sure which is the case" | 1:41:48–1:43:03 |
| Moonshots #298 (3 Oct) | Alex | the heavy-lift capacity is "an over supply in search of an under demand", so "fortunately and hopefully orbital data centers step in to fill that gap" | 1:39:43–1:39:54 |
| Moonshots #298 | Emad Mostaque | "I'm not bullish on orbital data centers I have to say and I remain to be convinced of that." Host: "total heresy relative to the established moonshots orthodoxy" | 1:41:30–1:42:11 |
| Moonshots #301 | Alex [attr?] (answering Emad) |
"entirely new markets like orbital data centers that Elon needs to basically popularize in order to saturate his own capacity" | 2:04:58 (Whisper .srt) |
| Moonshots #304 (20 Oct) | Dave | "even the most wildly aggressive estimates put a couple percent in space within a decade"; and, in the same breath, it "makes sense for Google, Microsoft, Meta, SpaceX… to essentially deploy 100% of their compute capacity into space" | 0:49:33–0:49:44 |
| Moonshots #305 (23 Oct) | Alex | "Take terrestrial telecom and lift it up. So you get direct to cell. Take terrestrial data centers, lift it up. You get orbital data centers. That's like the… vaudeville era of the space economy"; "I don't think it's going to be a long period" | 0:26:23–0:26:55 |
| Moonshots #305 | Alex | the municipalities that refuse a data-centre dividend are "pushing them out into orbit eventually" | 1:24:06 |
| Moonshots #306 (28 Oct) | Dave [attr?] |
"space-based data centers are going to be enormous. The Dyson swarm will be everything in the 2030s" | 0:09:09–0:09:19 |
| Moonshots #306 | Dave | "even after we have the Dyson swarm, energy on Earth will be extremely valuable because… data centers and other use cases that need constant modification are always going to be cheaper to do nearby" | 2:12:28–2:12:38 |
| Moonshots #307 (31 Oct) | Sam Altman (clip, Cheeky Pint, 29 Oct) | "We have a robotics program that we're going to point at making lots of data centers… self-replicating data centers". He says nothing about space | 1:01:38–1:01:48 |
| Moonshots #307 | Alex | "are these self-replicating data centers going to be terrestrial? Maybe not. Arguably they want to be orbital or even better lunar because they're free of regulatory constraints there"; "I think probably on the moon" | 1:04:21–1:05:15 |
| Moonshots #307 | Peter | OpenAI "is going to buy one of the launch companies… Rocket Labs or… Relativity Space"; "regulation's going to go to the moon" | 1:05:26–1:05:46 |
| Moonshots #307 | Dave | "it'll end up on the moon on the in the long long run", but for now the money is in terrestrial geography that welcomes data centres | 1:08:25 |
| Moonshots #308 (4 Nov) | Dave [attr?] |
wants Elon to "spend some time on Starlink, not just on Dyson orbital data center Swarm": Starlink thin clients that pull compute back to ground data centres. ⚠️ machine-translated captions | 0:12:13 |
| Moonshots #308 | Alex [attr?] |
the demand argument again, no number: Elon's plan "seems to be consists of orbital processing centers data and Dyson swarms", which is "a good a way to saturate all this new offering"; without new uses, many Starships "will eventually come out empty. I don't think that's realistic… because… ODC and others new use cases will appear". ⚠️ machine-translated captions, check the tape before quoting | 1:46:07–1:47:00 |
| Moonshots #309 (7 Nov) | Alex | the midterm backlash creates "uncertainty on the ground and greater certainty in space" (0:08:24); "the flops must flow… the orbital data centers launched to sun synchronous orbit"; "under the outer space treaty and under the artemis accord there is a lighter touch / That is less prone to two-year election cycles" |
0:22:15–0:23:01 |
| Moonshots #309 | Dave | ★ against EP304: "best case that's a couple percent of all compute in two years / And then it's on an exponential". On EP304 the same share was "within a decade" | 0:27:49–0:27:56 |
| Moonshots #309 | Alex | ★ the stake surfaces: "star cloud and lonestar / We have a financial interest and many many other / Neo space clouds"; "the baby dyson swarm is already here" (Larsen's satellite-trail photo); "Our skies are already filled with compute now. It isn't gpu compute yet" | 1:35:44–1:37:21 |
| Moonshots #309 | Alex | "sso is pretty roomy / So less concerned about kessler syndrome"; when SSO fills, "disassemble mercury" | 1:39:41–1:40:18 |
| Moonshots #309 | Dave | LEO "up to about 500 miles and then beyond that you've got the kessler effect problem" | 1:39:19–1:39:27 |
| Moonshots #310 (11 Nov) | Emad Mostaque | "is this Terafab thing and the space data data centers a bit of a head fake"; "How many geniuses do you need in space? I think very few" | 1:31:27, 1:33:23 |
| Moonshots #310 | Dave | ★ against EP309: "I think the head fake narrative actually has a lot of a lot of merit", recalling Peter's EP305 correction that the 10–20 GW is terrestrial | 1:32:10–1:32:42 |
| Moonshots #310 | Alex | Terafab output goes to data centres, inference from the cloud over Starlink; the Dyson swarm is "the default outcome"; later, "the preview of what a post-s singular world looks like is increasingly in space… for political and other reasons". Starcloud/Lonestar not restated | 1:34:18–1:35:13, 1:59:52 |
| Moonshots #311 (13 Nov) | Alex [attr?] |
the demand question, turned around: "if we're going to build out one or more Dyson swarms, what are we going to use all of that compute for?"; his answer is "advanced computational archaeology", reconstructing "our past light cone". No date, no number | 1:38:17–1:38:57 |
EP310 (checked 2026-11-12). The house line split. Four days after putting "a couple percent in two years" on the record, Dave sides with Emad's EP298 dissent: the Terafab chips go into robot, car and home "brains", and orbit is a narrative. Alex is now the only mate holding the orbital case, and he is the one with the disclosed stakes (EP309), which he does not restate here. The regulatory-escape premise returns a third time (Alex, "for political… reasons"; Dave, the Moon is "not messed up by regulation yet", 2:01:50) and still fails on OST Art. VI.
EP309 (checked 2026-11-08). Three changes. (1) The regulatory-escape argument (EP307's Moon) is now the panel's answer to the midterms. It still fails on OST Art. VI, and US operators still need FCC licences (with the 5-year post-mission deorbit rule), ITU filings and FAA launch licences; the Accords are non-binding principles. (2) Dave put a date and a number on the same line, and the date moved: a couple of percent of compute in space was "within a decade" on 20 Oct and "in two years" on ~6 Nov. (3) The advocacy now has a disclosed stake: Starcloud (Link-XPV 2026; $250M at $2.3B in Aug 2026, press) and Lonestar (unverified tie), said inside the pitch. The photo used as proof (Brendan Larsen, 578 exposures over Taranaki, ~14,000 trails from ~1,400 satellites) shows comms constellations, which Alex concedes ("It isn't gpu compute yet"). ⚠️ "SSO is roomy" was not checked this run. ESA's annual space-environment reports put the densest debris populations in the 500–1,000 km band SSO uses; check before anyone repeats it.
How to read the pattern: the demand argument (Starship's launch capacity needs a customer) is Alex's from EP298 to EP305 and is stable. The timing argument is not: Altman says not this decade, Elon says two to three years, Alex says the unit economics are unclear, and on EP304 Dave says a couple of percent within a decade while also saying everyone should move 100% there. No one on the panel had put a date and a number on the same line until EP309 (see the EP309 paragraph above). EP306 adds a decade ("the 2030s") with no number, and, in the same episode, the reason Altman gave against it: hardware that needs constant attention is cheaper on the ground. EP307 adds a new reason to go off-world: regulation, not cost. It is the panel's, not Altman's. Altman's clip is about robots building data centres and never mentions space. The premise does not hold as stated: under the Outer Space Treaty (Art. VI) a state must authorise and continually supervise its nationals' activities in space, including on the Moon. Alex names the treaty and the Artemis Accords himself (1:06:06) and argues past them.
This week (checked 2026-11-16, window 10–16 Nov)
What changed this week: the first long, sourced case against moving fast came out. An ICEF working paper led from Columbia says keeping 10% of 2030's AI data centres in orbit would take more than 700 large launches a year. It also says high costs will not stop the first deployments, so it asks regulators to hold back large constellations. The US Office of Space Commerce put two orbital data-centre builders into a certification pilot. SpaceX's own words on orbital compute were read at EDGAR: it says the initiative "may not achieve commercial viability". Nothing new flew.
- ★ "Space-Based Data Centers Roadmap: The Case for Caution" (ICEF, 12 Nov; primary). An Innovation for Cool Earth Forum roadmap, chaired by David Sandalow (Columbia, Center on Global Energy Policy), with 10 co-authors. The preface calls it a "working paper" that "has not been subjected to peer review", based mainly on desk research from summer 2026 (ICEF roadmap page; preface PDF).
- What it finds, in Columbia's own summary of 12 Nov (State of the Planet):
- "even 10% of the artificial intelligence data centers anticipated by 2030 into orbit would require more than 700 launches of large rockets every year, about two per day, in perpetuity".
- "neither technical challenges nor high costs are likely to stop initial deployment", because leading companies will absorb the costs to win position.
- Satellites "could outnumber visible stars in many places".
- What it recommends: regulators should not allow large constellations "until there is more operational experience and the risks and impacts are better understood and well managed".
- Other risks, from the opening essay: debris cascades, uncontrolled re-entry, ozone depletion and warming.
- Reported only, chapters not read: ~19,000 satellites in orbit today, and a "more than 50-fold" rise in LEO if the announced plans are built (mlq.ai, 15 Nov, secondary). Axios carried the report on 15 Nov (403, not read).
- Why it matters here: the panel's demand argument starts from Starship needing customers. The report does the same sum from the other side: what the panel counts as customers, it counts as about two big launches a day.
- What it finds, in Columbia's own summary of 12 Nov (State of the Planet):
- ★ US Office of Space Commerce: Starcloud and Cowboy Space join the certification pilot (12 Nov; primary). OSC picked four companies from nearly 40 applicants: Starcloud, Cowboy Space, Portal Space Systems and SpaceMD (a Redwire subsidiary). The page lists orbital data centres among the activities the certification will later cover (space.commerce.gov). A pilot cohort is not a licence, and it is not FCC authorisation.
- ★ SpaceX's own words on orbital compute (Open item closed in part). The secondary survey's quote was close but not exact. Read in SpaceX's free-writing prospectus (FWP, 4 Jun 2026, accession 0001628280-26-040874, under "SpaceX's Challenges"): initiatives "including those to develop orbital AI compute at scale, manufacture AI chips at scale, establish a lunar economy… involve significant technical complexity, unproven technologies or technologies that do not exist, and such initiatives may not achieve commercial viability."
- Where else it appears: an exact-phrase search places the same words in the S-1 (20 May, 0001628280-26-036936), both S-1/As and the 424B4 (12 Jun). The S-1 itself was too large to open, so its heading there is not confirmed.
- Read against EP305: the panel says SpaceX will put 10–20 GW in space. SpaceX has told investors in writing that orbital AI compute may not be commercially viable.
- SpaceX filings this week: nothing about orbit (EDGAR, read 16 Nov). Two new filings, both investor holdings disclosures (Schedule 13G/A), not company filings: Founders Fund entities, 3.8% (13 Nov), and Valor / Antonio Gracias, 5.5% (16 Nov). Neither mentions orbit. EDGAR full-text search for "orbital" in SpaceX filings from 7–16 Nov returns nothing.
- A second big-cloud chief says a decade or more (13 Nov, secondhand). AWS CEO Matt Garman said on the Sources podcast that orbital data centres are "highly unlikely" to be practical for "a decade or two" (Business Insider via Yahoo, 13 Nov; podcast not heard). That follows last week's remarks from Google's head of data-centre energy. Neither has published a cost model.
- China: Xingyun Space AI Technologies founded in Hefei. A joint venture led by the Chinese Academy of Sciences' Aerospace Information Research Institute, planning three classes of computing satellites (China Daily, reposted by ncsti.gov.cn, 11 Nov; state media, not a company release). No Chinese orbital-computing launch was found for 10–16 Nov. ADA Space's next group is reported rolled out, not flown (X post, 13 Nov, unverified).
- The panel: Moonshots #311 (13 Nov). One line. Alex
[attr?]asks what a Dyson swarm's compute would be for, and answers "computational archaeology". No date, no number, no builder named. EP310 was logged on 12 Nov. The watcher shows 113 on disk and nothing missing (16 Nov 14:43 UTC). - Builders: no launch, no new compute in orbit.
- Checked at their own sites, nothing new: Google Suncatcher (no post after 30 Oct, per the research RSS), Starcloud, Cowboy Space, Axiom, Lonestar and Blue Origin.
- Reported, not opened: OwlO Cloud (Luxembourg) came out of stealth with Thales Alenia Space and Edge Aerospace agreements, targeting more than 10 nodes by 2030 (DCD, 12 Nov, secondary).
- Outside this window: the Orbital Compute–Orca deal ("up to $20 million") is dated 9 Nov in coverage.
- Regulators: no FCC action found, and the FCC's own records could not be read. The ICFS record for Cowboy's Stampede application (SAT-LOA-20260323-00135) did not render, and fcc.gov returned 403.
- SpaceX's application has a file number in trade press, SAT-LOA-20260108-00016, reported accepted on 4 Feb with comments due 6 Mar (Public Notice DA 26-113, not opened). It is the first file number on the record behind the "one million satellites" claim. Still unverified at the FCC.
- "No grant yet" rests on trade press only (newspaceeconomy, 4 Nov; CP24, 14 Nov).
- Papers.
- The Suncatcher paper, read at source (arXiv 2511.19468v2): "launch prices could fall to <$200/kg by ∼2035". At that price, the cost of power in orbit becomes "approximately comparable to terrestrial spend". That compares the cost of power only, not the total cost of compute, and the paper says it is not a full economic analysis. Last week's CLS paraphrase got the number right but read more into it than the paper does.
- "Rings in the Sky" (2608.02757): unchanged, still v2, "AJ in press", no journal DOI. Co-author Samantha Lawler discussed it on Canadian TV (CP24, 14 Nov).
- New: "OrbitWeaver" (2611.17850, 13 Nov), on scheduling jobs across orbital GPU clusters. A systems paper, with no cost number in the abstract.
- Open programmes: one adjacent call, deadline moved.
- NASA/JPL High-Performance Space Computing pre-RFI: re-posted 11 Nov (solicitation SBPO-2610-09012026); responses now due 8 Jan 2027, 3 p.m. Pacific. It is market research on spacecraft computing, not data centres. The submissions register is updated.
- Nothing else open: no call aimed at orbital data centres was found at NASA, DARPA, SpaceWERX, ESA, UKSA or the EU. The OSC pilot cohort is closed. The ICEF roadmap takes comments by email, which is not a call.
Previous week (checked 2026-11-10, window 4–10 Nov)
What changed that week: a cost number came from inside Google, and it points the other way from Google's own satellite: its head of data-centre energy says orbital data centres could cost up to 100 times what ground ones cost. South Korea put a dated, priced first demonstration on the record. SpaceX filed nothing about orbit, and no builder flew anything new.
- ★ Google's energy chief: "up to 100 times" (remarks on or before 2 Nov; missed last pass). Amanda Peterson Corio, Google's global head of data-centre energy, at the WTF summit. Read in CLS (财联社), 2 Nov, in Chinese, so this is a translation: launch cost is "extremely expensive, up to 100 times the cost of ground construction". In-orbit maintenance and upgrades are unsolved, while on the ground staff can swap servers at any time. Power is not the constraint; grid-interconnection approvals are. We should "first explore every feasible option on the ground", and the technology is "by no means imminent". English coverage traces to The Information (paywalled, not read). These were conference remarks, not a cost model, and "up to" is a ceiling, not a central estimate. Why it matters here: this is the first answer on the record to two of Altman's three questions (cost against ground power, servicing a failed GPU), and it comes from the company that flew the only orbital TPU. The same CLS piece gives the Suncatcher team's threshold as reported: parity on compute per unit of power if launch falls to $200/kg or below, around 2035. That is a reported paraphrase; the paper itself was not opened this run.
- ★ South Korea: a 100 kg orbital data-centre demo on Nuri's 7th launch, Nov 2028 (2 Nov; missed last pass). Presented at the National Assembly's 3rd Space Data Center Forum as the "K-Moonshot Space Mission 2035 Roadmap". Read in Newsis, 2 Nov. Fnnews runs the same wire copy, so it is not a second source. The plan: a 100 kg-class satellite, flown as a secondary payload, in a ~561 km dawn-dusk sun-synchronous orbit for more than a year, to partly verify AI compute, power and heat rejection with Korean-made AI chips. Then a 500 kg unit with real AI services in 2031, a multi-tonne platform in 2034, and a private-led commercial base by 2035. Cost: KRW 14bn in total (KRW 5bn in 2027, KRW 9bn in 2028, including ~KRW 2bn of Nuri launch cost). Development has to start in January 2027, but the 2027 budget bill carries only KRW 2.1bn, for ground verification, per the article. The programme lead's own framing: competing with big tech's platforms at the same level "has no practical benefit". This is a roadmap at a parliamentary forum, not a funded programme. No ministry release was found.
- Larry Page's "five nines" (8 Nov, secondhand). A remark attributed to Page, that 99.999% of compute will end up in space, was retold by Patrick O'Shaughnessy on Invest Like the Best, who had heard it, not from Page. Elon Musk quote-posted the clip on 8 Nov: "Larry Page has made many accurate predictions and this is one of them" (OfficeChai, 8 Nov; X, not opened). Page has not said it publicly. ⚠️ Unverified attribution; do not repeat it as Page's number. Musk's agreement has no date and no capacity attached.
- SpaceX filings: nothing orbital (EDGAR, read 10 Nov). One new filing in the window: an 8-K, accession 0001628280-26-071689, filed 6 Nov, Item 5.02 only. Class B director Donald Harrison resigned effective 4 Nov, "not due to any disagreement". No "orbital", "space-based" or "in space" (checked on extracted text, not the raw HTML). A secondary market survey (URL-dated 4 Nov, no byline) says SpaceX's "May 2026 registration filings include an explicit risk factor warning… orbital AI computing involves unproven technology and may not achieve commercial viability". It names no document. ⚠️ Not checked at EDGAR, so it is not carried as a SpaceX statement; added to Open. The same survey's counts (53 programmes tracked, 21 flown compute or storage hardware, none of five mega-constellation FCC filings granted) rest on SpaceNexus and are not verified.
- The panel: Moonshots #308 (4 Nov) is now on the table. Alex restates the Starship-demand argument for orbital data centres ("ODC"), and Dave asks Elon for Starlink thin clients over the Dyson swarm. Neither gives a number. The captions are a machine-translated round trip, so every quote needs the tape. EP309 (7 Nov) was already logged on 8 Nov. The watcher shows 111 on disk and nothing missing (10 Nov 12:00 UTC).
- Builders: nothing dated 4–10 Nov at their own sites. Google's research blog listing showed no Suncatcher post after 30 Oct, but the listing rendered no article links, so that is not a confirmed absence. Starcloud's blog items are undated. Cowboy Space's homepage is unchanged in substance (page metadata 2 Nov). Sophia Space's newsroom rendered no items. Redwire's newsroom could not be read (TLS certificate error). Reported, not opened: Orbital Compute Inc. signed an onboard-autonomy deal with Orca Aerospace, "up to $20 million", first launch 2028 (DCD, 403, date unverified).
- Papers: "Rings in the Sky" (2608.02757) is unchanged: still v2 (27 Oct), "AJ in press", no journal reference or DOI (checked 10 Nov). Today's space-rail scan has a satellite-streak survey-impact paper (2611.10301, HETDEX) about satellites in general, not orbital data centres (abstract not read).
- Open programmes: none new. The Korean demo is a government roadmap with no call. NASA/JPL's High-Performance Space Computing pre-RFI (responses due 13 Nov) is still the only adjacent open notice, and it is already in the submissions register. Nothing was added to the pipeline this week.
Two weeks back (checked 2026-11-04)
What changed that week: the one flown prototype actually computed — Google got the first AI response from a TPU in orbit — and SpaceX signed the week's biggest compute cheque, ~$40B of GPU leasing, with every word of it on the ground.
- Suncatcher ran inference in orbit (30 Oct, primary). Google's own blog (Travis Beals, 30 Oct): "This week, we reached our 'Hello World' milestone: the first-ever AI response from a data center class TPU in orbit." A prompt was uplinked during a ground-station pass to the onboard Gemma 4 open model, asking for a haiku about being in space; "As the satellite passed by Svalbard, Norway, the downlinked poem confirmed success" — "Blue marble below / Data streams in silent grace / Starlight feeds my core" (blog.google, 30 Oct). Then space-themed limericks and telemetry prompts. Next, per the post: thermal-stress and radiation testing, then "twin interconnected satellites in 2027". The post gives no power or duty-cycle figures: the circulating ~1 kW solar / one-TPU-at-a-time / ~20-minute-burst numbers are TechCrunch only (30 Oct), from a Google spokesperson. A "Pichai 'hello world' post" circulating in trade press is not in the blog post; the phrase is the post's own. A hardware test of four TPUs, not a data centre — Altman's three questions (cost against ground power, servicing a failed GPU, scale) remain unanswered on the record.
- SpaceX's ~$40B of new GPUs is terrestrial, in writing (8-K, 26 Oct, primary). Accession 0001104659-26-118614, item 8.01: a binding commitment letter dated 24 Oct for a master finance lease with an entity managed by Apollo Global Management — the lessor buys GPU pods and related equipment and leases them to SpaceX "for use at data centers selected by the Company". Aggregate principal payments ~$40B, tranches of roughly 5–5¼ years, rent quarterly in advance, closing expected before year end, lease commencement H1 2027. The 8-K document contains no "orbital" and no "space-based" anywhere (no exhibits listed). The only other filing since the 10-Q is an S-8 (22 Oct, employee benefit plans). The 10-Q's MD&A and risk factors remain unread (Open).
- "Rings in the Sky" (arXiv 2608.02757): no change. Still v2 (27 Oct), still "AJ in press"; no journal reference or AJ DOI on the arXiv page (checked 4 Nov).
- The panel: no new tape. EP307 (31 Oct) is still the latest episode; the watcher shows 109 on disk and nothing missing (4 Nov 12:00 UTC).
- Builders otherwise quiet. Starcloud: nothing new found this week (Starcloud-2 still pointed at 2027). Cowboy Space: nothing new since the Big Iron reveal, but the Stampede FCC application now has a file number on the record — SAT-LOA-20260323-00135, reported accepted for filing 12 Jun, with Secure World Foundation comments
10 Jul (all secondary — DatacenterDynamics, SWF; the filing itself not opened; trade press disagrees on the filing date, March vs 14 May). Redwire–EnduroSat MOU (21 Oct): still not opened; nothing new found. - Open programmes: none new for orbital data centres. Adjacent and open: NASA's High-Performance Space Computing (HPSC) Pre-RFI sources-sought (SAM.gov, posted 29 Oct, responses due 13 Nov) — market research on spacecraft computing, not data centres; title and dates from the SAM API, notice text unread. Already in the submissions register. Cowboy's "Big Iron Express" still has no published application, deadline or price.
Earlier (checked 2026-10-29)
What changed that week: a second company put a dated orbital data-centre vehicle on the record, an astronomy paper on what these constellations do to the night sky was revised and accepted, and the panel spoke about it twice in one episode without a number.
- Cowboy Space "Big Iron" (26 Oct; mostly trade press). Cowboy Space (renamed from Aetherflux in May 2026; aetherflux.com now redirects to the company site) unveiled a rocket whose upper stage stays in orbit as the data centre. The company's own site says the upper stage "is the satellite itself. It's a megawatt-class data center" and "a 1-megawatt data center with active thermal management and integrated compute" (cowboyspace.com). Everything else is from Payload, 26 Oct (secondary, from an interview with the company): 86 m stainless steel, "15,000+ kg to sun-synchronous orbit", nine methane "Clydesdale" first-stage engines, "five racks of NVIDIA Vera Rubin GPUs" per satellite, first flight targeted for 2028, and the planned "Stampede" constellation of "up to 20,000 sats, pending FCC approval." Not stated anywhere read: whether 1 MW is total power or power delivered to the chips. A target, not a flown vehicle. Its first satellite, Reason-1, is a power-beaming test, not compute (launched 1 Oct on Transporter-18, per Payload).
- "Rings in the Sky" (arXiv 2608.02757, primary). Boley, Lawler and Rein; v1 3 Aug, v2 27 Oct, marked "AJ in press". It models three companies' proposed orbital data-centre designs and finds they would form "large ring structures with each ring passing through the sky twice a day", worst in winter and visible all day near the poles; dimming and/or hard limits on numbers are needed to avoid "more visible satellites in the sky than visible stars during certain times of night." The abstract gives no figures and does not name the companies (full paper not read).
- SpaceX: the 10-Q (22 Oct, accession 0001628280-26-066880) was opened on EDGAR. In the part read (cover, forward-looking statements, notes 1–7) it says nothing about compute in orbit; the data-centre language is all ground build-out ("complete data center buildouts, obtain sufficient power, GPUs"). MD&A and risk factors were not read. A circulating post that "SpaceX will deploy the first orbital data centers next year" has no primary source found; not carried.
- Reported, not opened: Redwire signed an MOU with EnduroSat that lists EU-based orbital data centres among three targets (~21 Oct, trade press; its second orbital data-centre MOU after Sophia Space, 30 Sep). An MOU, not a contract. Starcloud-2: no launch found; reporting now points to 2027.
- Google Suncatcher: nothing new found since the 1 Oct prototype.
- Open programmes: none found. Cowboy's "Big Iron Express" (outside payloads up to a megawatt) is described by Payload only, with no application, deadline or price published, so it is not a call anyone can answer yet.
Against the record (checked 2026-10-24)
- Elon's stated scale: at the March 2026 Terafab announcement he reportedly said his companies would eventually need up to 200 GW a year of compute power for terrestrial operations and up to 1 TW for space-based processing (paraphrase from DatacenterDynamics, secondary; article not opened in full). The EP305 call figure for Terafab is "1 terawatt of AI computing per year" with test chips mid-2027 and volume ~late 2028 and "big error bars" (
0:30:06–0:30:27). A target, not a deployment. - The one flown prototype: Google's Project Suncatcher put four TPUs in orbit on 1 Oct 2026, built with Planet (Google; logged on SpaceX). It was never discussed in the episode that mentioned it. Update 30 Oct: it returned the first AI response from a TPU in orbit — Gemma 4, a haiku, downlinked on a Svalbard pass (Google blog, 30 Oct; see This week).
- What is not established: any orbital compute capacity in service at SpaceX; a cost per unit of orbital compute against ground compute; how a failed GPU is serviced. These are the three questions Altman's clip raised and nobody has answered on tape. Update (remarks on or before 2 Nov): Google's own head of data-centre energy put a ceiling on the first, "up to 100 times" ground cost, and called the second unsolved (see This week). That is a conference remark, not a published model.
What an operator can take from it
- The orbital part of the story is a launch-cadence and licensing story today, not a compute story. SpaceX's Flight 15 is planned for November with a ship-catch attempt, and the FAA's Starbase expansion review is open for public comment until 30 Nov 2026 (see SpaceX). That is the part a procurement- and NASA-minded reader can act on.
- ⚠️ Do not repeat "10 to 20 gigawatts in space". The figure is terrestrial.
Open
- Whether SpaceX's 10-Q (22 Oct) names any orbital compute programme: notes 1–7 read (nothing); MD&A and risk factors not read. The 26 Oct 8-K and 22 Oct S-8 read/checked 4 Nov: nothing orbital.
- Whether SpaceX holds an FCC application for an orbital data-centre constellation: trade press (secondary) describes one the FCC "accepted", reportedly up to one million satellites; not found or opened at the FCC — unverified, not carried as fact.
- Big Iron's power budget: total or delivered to compute.
- "Rings in the Sky": which three companies, and the brightness and count figures in the full paper.
- SpaceX's caution on orbital AI compute: wording read in the 4 Jun FWP (16 Nov). Still to do: read the same sentence, and the heading it sits under, inside the S-1 or 424B4 (the file is too large for the fetcher).
- SpaceX's FCC application SAT-LOA-20260108-00016 (file number from trade press) and Blue Origin's reported SAT-LOA-20260310-00118: open both in ICFS and read their status and conditions.
- The ICEF roadmap chapters: read the satellite-count, brightness and launch-mass figures in the chapter PDFs, not in coverage.
- Corio's "up to 100 times": the date and venue of the WTF summit remarks in English (The Information, paywalled). The Suncatcher $200/kg threshold was read at source on 16 Nov: it compares the cost of power only.
- Korea's K-Moonshot demo: whether the 2027 budget funds the satellite (the bill reportedly carries KRW 2.1bn for ground work only).
- The EP272 clip is from the backfill copy (
backfill/ep272); no episode page exists for it, so it is cited by number and date only.
How this page stays current
One pull a week (the vault-weekly engine). Each step names the source and what to diff:
- Moonshots. Read the research archive and any new episode directory; grep the transcript for
orbit|in space|space-based|data cent|Suncatcher|Starcloud|Dyson|gigawatt|terawatt. Diff: a row in What the panel has said for each claim, with speaker and timestamp; a claim with a date and a number on the same line gets flagged in the pattern paragraph. - SpaceX filings. EDGAR, CIK 1181412: any new 8-K, 10-Q or S-filing; search for
orbital,space-based,data center. Diff: whether any orbital compute programme, capex line or date appears in writing. - The builders. Google (Suncatcher), Starcloud, Cowboy Space, Blue Origin, Axiom, Sophia Space/Redwire: own sites and launch manifests. Diff: anything launched, any compute actually run in orbit, any number that changes.
- Regulators. FCC filings for orbital data-centre constellations (SpaceX, Cowboy "Stampede", Blue Origin). Diff: grants, conditions, comment deadlines.
- Papers. the research archive and arXiv for orbital data-centre cost, thermal and sky-impact work. Diff: a peer-reviewed number on cost per unit of compute against the ground.
- Calls. Any open programme, prize, grant or tender on space-based computing goes to
submissions/pipeline.jsonandsubmissions/operator_guide.json, not only here.
Related
SpaceX · Elon Musk · Alexander Wissner-Gross · Dave Blundin · Emad Mostaque · NASA