SpaceX and orbital compute
Electricity outside China is more or less flat. Chip output is not. Musk’s claim is that the cheapest place for new AI watts leaves the ground — on a clock the same pages call aggressive.
Elon Musk sat with Dwarkesh Patel and drew a gap. Look at electrical output outside China, he said. More or less flat. Chip output is growing pretty much exponentially. “So how are you going to turn them to chips on? Magical power sources. Magical electricity fairies.” In 36 months, probably closer to 30, the most economically compelling place to put AI will be space. Scaling beyond a terawatt, on that telling, only happens there.
the output of chips is growing pretty much exponentially, but the output of electricity is flat. So how are you going to turn them to chips on? Magical power sources. Magical electricity fairies
Elon Musk
The physics pitch is simple enough for a roadshow. A solar panel in orbit, Musk said, does about five times the work of one on the ground. The atmosphere alone costs about 30 percent. You skip the batteries that carry a terrestrial array through the night. Will Marshall, who runs Planet Labs and is not on SpaceX’s payroll, reached the same five-times figure from a sun-synchronous dawn-dusk orbit. Covering Nevada in panels, Musk added, means permits that do not exist. Dwarkesh’s gloss: space is a regulatory play. It is harder to build on land.
Ron Baron, an early SpaceX investor with a position he has put around $15 billion, restated the argument on CNBC and added the politics. New Jersey’s governor, he said, stopped data-center construction because electricity bills were too high; a roommate from Congress won in Virginia on the same idea. Chamath Palihapitiya put a number on the current-year version: of about 9 gigawatts supposed to come online, almost half is being protested. Philip Johnston, who is trying to fly the thing with Starcloud, says a new terrestrial energy project is a five-to-ten-year permitting queue. Even at cost parity, he argues, orbit wins on speed.
Racks, not a Pentagon
Gavin Baker’s correction is architectural. People hear “data centers in space” and picture a building. A Blackwell rack, he said, weighs 3,000 pounds, eight feet by four by three, a hundred kilowatts. Starlink V3 already runs at 20 kilowatts; SpaceX, in his telling, is confident it can go to 100 or 120. You keep the panels in a sun-synchronous orbit. You connect the racks with lasers through vacuum.
Marshall’s cost trigger, from an eight- or nine-year-old Planet/Google study: when launch falls to about $200–300 a kilogram, orbit is simply cheaper. Current cost is about $1,000 a kilogram, down about ten times in ten years. Andrew Fox, on a podcast where he and Baker were buying the stock, has Falcon at about $1,500 a kilogram toward $250 as the vehicle is reused. Below that band, Fox’s in-house math is about $5 billion of non-silicon capex per orbital gigawatt against $20–25 billion on the ground for switchgear, generators, transformers, and the shell. Baker has put a ground gigawatt at $60 billion today. Johnston’s Starcloud 3 is 200 kilowatts and three tonnes — about fifty per Starship, about 10 megawatts a launch. His own crossover is higher, around $500 a kilo, which would arrive earlier if the number holds.
Brian Potter, writing at Construction Physics on 17 September, put a manufacturing number on the engine that would have to fly that cadence. Raptor 3 first flew in May 2026 and delivers about 35 percent more thrust than Raptor 1. The cycle is still full-flow staged combustion. The visible tangle came off by killing development sensors, welding what used to be bolted flanges, 3D-printing internal lines — SpaceX licensed Velo3D in 2024 — and dropping heat-shield and fire-suppression mass. “The biggest mass change from version 1 to version 3 of the Raptor comes from mass removal of ‘vehicle side’ engine hardware, which is almost certainly largely the heat shield.” Flight Test 13 aborted at T-0 in July 2026 when several Raptor 3 engines failed to start. The schematics are unofficial. There is no teardown. This is engine-cost color, not a new dollars-per-kilogram print and not a named orbital-compute purchase order. The crossover stays a hypothesis.
The $5 billion figure excludes the silicon, which dominates the bill. “Space cooling is effectively free” is the load-bearing simplification. Radiative cooling in vacuum scales as the fourth power of temperature. That is the thermal constraint the slogan skips. Johnston says his radiator is ten times less mass per watt than the ISS design, and about a hundred times cheaper — a company claim, first flight slated for January 2027.
Flops per watt, not launch price
Marshall names a second tax. Near term the launch tax dominates. Longer term it is the compute tax. In orbit the binding constraint is the energy you must radiate away, which sets spacecraft mass and therefore launch cost. “Really the efficiency of the compute drives the amount of energy you have to dump which drives the mass of the spacecraft.”
On that accounting, inference-time power efficiency — flops per watt — picks the winner, not dollars per kilogram alone. Marshall: “Google TPUs are significantly more efficient than GPUs in a flops per watt standpoint.” Dave Blundin, on the same Moonshots episode: “inference time, power efficiency determines the winner of the entire thing.” That is adverse for Nvidia in the orbital regime specifically, not terrestrially. Marshall is Planet’s CEO and Planet was selected to fly Google’s first Suncatcher demo satellites. He is not a disinterested party on a Google-favouring thesis.
The eclipse and the last ten percent
Sun-synchronous orbit is not twenty-four-hour sun. SemiAnalysis put daily eclipses at about 35 minutes. Baker’s “always in the sun” and Marshall’s “24/7 looking at the sun” do not survive that point. The honest version is far fewer batteries than a terrestrial night — minutes of storage, not eight to twelve hours — not zero. Johnston relocates the fix to altitude: at about 1,200 kilometers you can avoid the blackout, and inherit Van Allen radiation and debris that lasts on the order of fifty years.
Feldman, the Cerebras chief, is not arguing the solar math. He is arguing the cluster. Building the communications between chips in space is an unsolved problem, he said, and the last 10 percent may be 80 percent of the time, the way self-driving was. Johnston, the most committed operator in the argument, cannot book Falcon 9 for 2028. The government took twenty launches. The 30-to-36-month window now leans on Starship cadence. On September 15 Musk named the next two: Flight 14, then catch Flight 15. That is a sequence, not a booked orbital watt.
Twenty-six satellites, if the license closes
The sequence now has a clock. SpaceX’s own page, and an official post on September 17, target Flight 14 as early as Monday, September 28, in a seventy-five-minute window that opens at 8:15 Eastern — still pending regulatory approval. The prior target was September 22. The company has not said why it slipped. Booster 21 rolled to the pad at Starbase on September 21.
This is the first planned send of Starship into Earth orbit after thirteen passively safe suborbital tests. The published profile is about 275 kilometers, about six orbits, nearly ten hours. An insertion burn only after flight control is happy with the hardware that later has to deorbit; a payload window; a single-Raptor deorbit; a Pacific splashdown west of Chile. Super Heavy is supposed to do an offshore landing burn in the Gulf. Flight 13 iced three center engines and lit eight of thirteen on the landing-burn attempt. The booster changes are the company’s answer to that, not proof they hold.
The payload is the first planned operational constellation insert of Starlink V3 — twenty-six satellites, each claimed at a terabit a second. SpaceX puts the mission at about twenty-six terabits, roughly ten times a Falcon 9 V2 mini launch. Flight 13 flew twenty V3s on a suborbital path. That is not the same job. After unfold, radio and laser contact, and an orbit raise, the company says customer service could start a few weeks after checkout. Three of the birds carry cameras to image the heat shield.
The FAA’s Boca Chica page shows a completed Pacific environmental assessment and a finding of no significant impact. That supports a license-modification pathway. It is not, by itself, a Flight 14 go. A planning reservation for the re-entry names September 28 with backups through October 4. The reservation URL was not fully retrieved. Residual slip into the window is still material — they already moved once, without a public reason. Confirm is orbit plus the twenty-six on the way to station. Miss is a withheld burn, a partial deploy, a scrub, or another slip. None of that books an orbital watt. The June print, the lockup calendar, and the index map do not move with it.
Marshall expects most compute in space within ten years. Johnston, who is actually trying to fly the hardware, expects most new capacity in ten years and would be surprised if more than half of all compute is there in twenty. Flow versus stock. They are not the same claim.
A hyperscaler in thirty days
The other chain is terrestrial, and it is a hypothesis. Baker says Elon stands data centers up faster than anyone — 122 days, speed as cost, because every day you pay electricians. That 122 is now in two first-party places: SpaceX’s June 11 filing and an Nvidia press release from October 28, 2024. It is a cluster clock, not a greenfield campus. The first 100,000 Hopper chips went into a reused Electrolux shell. Follow-on clusters printed 91 days, then 64. SpaceX’s own peer line for a purpose-built 100-megawatt hall is about two years. Do not write the folklore 100-day number.
Epoch’s campus math and the issuer’s “about 1.0 gigawatt of compute power” are different units — 340 plus 946 megawatts of IT load against the company’s own phrase. An unfetched Introl 2-gigawatt / 555,000-chip line sits farther out still. Colossus 2’s first-online date is split: Epoch says October 19, 2025; Wikipedia says about January 2026. Leave it open. None of that falsifies “faster than a purpose-built campus.” In thirty days, on Baker’s podcast, SpaceX went from not being an AI hyperscaler to number four, passing Oracle. Implied monetization, per Fox: something like $14 billion per gigawatt per year; Anthropic just signed at 22 to 23, Google at 50. Brad Gerstner, on the same show: “We’re also a shareholder.” Will Marshall’s objection arrived two weeks later. “That’s not an AI play, that’s the data center play, which is interesting, but it’s a very different business.”
Dylan Patel, who is not on the cap table, told Dwarkesh in late August that SpaceX is building a ton of compute and will lease quite a bit of it to Anthropic and OpenAI — the labs that can pay the highest price. The wiki files that as color on the same hypothesis, not a booked offtake.
The same September 15 conversation restated Terafab as a terrestrial R&D fab at Austin’s Giga Texas — Tesla and SpaceX, equipment on order, “something useful by end of next year, not at scale.” Gwynne Shotwell’s line is crawl, walk, run; they already do packaging. That is not a named high-volume wafer start, and it is not an orbital campus. The foundry second-source article carries the Intel-customer version of the same name.
The first public quarter
The listed company is no longer a roadshow slide. Class A began trading June 12 as SPCX. The offering closed June 15 at 638.9 million shares and $135, $85.675 billion net after costs. Musk still holds 85.1 percent voting control on the S-1 figure. The first issuer print, posted August 4 after the close, is a Connectivity cash engine against AI and Starship absorption.
Q2 revenue was $7.814 billion, up 92 percent. Connectivity — Starlink — did $4.291 billion and $1.656 billion of operating income, 12.0 million subscribers at $66 of ARPU. AI printed $2.561 billion and lost $1.257 billion. Space lost $542 million. GAAP net loss $541 million. Adjusted EBITDA $3.538 billion, a company label. Capex was $18.369 billion, of which AI took $15.828 billion. Cash and marketables $100.0 billion. Backlog $47.5 billion. The 10-Q does not print a quarterly free-cash-flow line.
Nameplate compute is 1.4 gigawatts — installed GPU times all-in power, not utilization. Cloud-service agreements are $14.1 billion of contracted sales in the non-cancellable window. The same 10-Q says those agreements may be terminated on 90 days’ notice after the initial ramp. Both sentences are in the filing. Orbital data centers are still not a P&L line. The company says it will allocate a significant amount of launch capacity to AI in the future. That is not a booked orbital watt.
Monday morning, September 14, the NightDesk last was $147.76. That is a last only — do not invent a restamped market cap from it. The August 26 close on stockanalysis was $139.63, a $1.90 trillion mark, 34.90 percent of the float short. Multiples there are the site’s, not SpaceX’s. FY2025 remains the historical S-1 year: $18.67 billion of revenue and a $4.94 billion net loss. The orbital story is still, in Chamath’s phrase, value “on the come.” The Connectivity print is not.
Who can sell, and who already bought
The lockup is not a single 180-day cliff. The 424B4 dates three pools from the prospectus date, June 12, not the June 15 close. The first contractual eligibility was August 6: the second full trading day after the August 4 print, up to 911.5 million shares, the earnings-tied 20 percent, ex-affiliates. August 20, the 70th day, added up to 319.0 million. A performance extra — 455.8 million if the stock held $175.50 — did not fire on the fetched closes. September 9 is the 90th day, up to 319.0 million ex-affiliates. September 10 is the 91st-day affiliate row, up to 59.1 million. Later 180-day-pool steps — September 24, October 9, October 24 — stay supply context. The Q3 earnings-tied 28 percent / up to 1.3 billion still has no calendar day; street “mid-November” is secondary color, not a 424B4 date. Musk is locked through June 12, 2027, up to 6.4 billion, with no early release. Eligibility is not a sale tape. No fetched 8-K says a tranche actually traded.
Index adds are a mixed tape. FTSE Russell put SPCX into the Russell 1000, Top 200, and Small Cap Completeness after the close on June 26 — a reconstitution calendar, not a standalone day-five print. Motley Fool says the Nasdaq-100 added the name on July 7, with expected buying of $4.3 billion; the official Nasdaq HTML was not fetched, and Nadig’s earlier QQQ-only math was about $7 billion. UFO, the space ETF, said it added SPCX at 6.17 percent as of June 16. CRSP’s June 18 add is trade press. Day 10 is still unidentified. The S&P 500 mega-cap exceptions were rejected on June 4. Trade press still says June 2027 at the earliest.
September’s Nasdaq-100 quarterly rebalance is a post-inclusion float-reweight, not another Fast Entry. Methodology still maps an announce after the close on Friday, September 11 and an effective open on Monday, September 21 — third Friday September 18, six trading days prior. June’s IR precedent announced June 11 and took effect Monday, June 22. Those September dates stay methodology-mapped and GIW-timing-consistent. They are not IR-confirmed. The September quarterly-changes page on Nasdaq IR was still a 404 as of September 14. The newer Nasdaq-channel signal is Bloomberg-syndicated GIW daily pro forma: about 1.28 percent toward about 2.82 percent, indicative, final later this month. The raw GIW row was not retrieved. Desk copy that had the name moving from about 1.25 percent to 1.51 percent and about $12.4 billion of passive buying is superseded attributed secondary. Do not multiply the new gap by $1.7 trillion of tracking AUM and call it a flow. Syndication says “billions.” That is not operational size. Low-float names still use Modified Market Cap: the lesser of listed shares or three times free-float shares, and Nasdaq does not bake future lockups into float until they have happened. Morningstar’s July 7 Fast Entry weight was about 1.3 percent, a pre-unlock baseline. June IR tracked more than $800 billion of NDX AUM; syndications quote about $1.7 trillion at end-Q2 and QQQ around $481 billion. Do not flatten those. The question stays open.