SpaceX CFO Just Put $100B ARR and Orbital Compute on a Clock

SpaceX's CFO just told Goldman's Communacopia conference the quiet numbers that turn the IPO story into a schedule. Bret Johnsen said SpaceX believes it is on track to hit $100 billion in annualized recurring revenue by the end of this year by annualizing December. Earlier this month, he said, SpaceX closed another hosting deal — customer unnamed — that translates into about $1.11 billion a month starting December 1, roughly another $13 billion of ARR from that date. He said almost all of the external compute deals, including this one, are 90 days with a 90-day out, roughly six-month commits, so treat that $13 billion as run-rate, not a locked decade. Then he stacked the next clocks on top of that. End this year with a little over 2 gigawatts of terrestrial compute deployed. Next year, a range of 5 to 10 gigawatts. First orbital compute satellites next year. Huge amounts of compute in space going into 2028. Direct-to-device satellites flying next year, service on in the first half of 2028. Farzad posted the cut Friday, September 11, under the title SpaceX CFO Makes Insane Predictions. The Goldman talk was Thursday, September 10. The source is a Goldman conference conversation with Bret. Treat every figure below as his live framing unless a separate filing confirms it.

$100 billion ARR is not a vibe. It is a December run-rate

Johnsen has been SpaceX CFO for 15 years. He tied the $100 billion ARR target to annualizing the December number. The update he gave Goldman is why conviction went up, not down. Earlier this month, he said, SpaceX closed another hosting deal — customer unnamed — that translates into about $1.11 billion a month starting December 1, roughly another $13 billion of ARR from that date. He said almost all of the external compute deals, including this one, are 90 days with a 90-day out, roughly six-month commits, so treat that $13 billion as run-rate, not a locked decade. That is Bret on stage, not a 10-K line. If the December run-rate lands, SpaceX is already pricing itself as an AI infrastructure company that also launches rockets, not a rocket company that dabbles in AI.

On the earnings call, he told Goldman, SpaceX talked about $30 to $50 as next year's monetization range, and that the company was at the high end of that range. He did not attach a unit to those two numbers on the Goldman mic. He framed the model as compute from a wattage perspective and compared peers on dollars per watt. On the August 4 earnings call, Musk's public unit was monetization per watt, “somewhere between $30 and $50.” Do not write $30 to $50 billion.

The compute deals are short on purpose

Almost all of the external compute deals, he said, are essentially 90 days with a 90-day out, roughly six-month commits. The deal he just closed earlier this month uses the same structure. The reason is not that demand is soft. The reason is internal conviction. SpaceX does not want to lock forever compute away from its own products while Grok, Grok Bot, and the Cursor team are still ramping.

Sheridan asked him to compare the cost of standing up this terrestrial compute to Colossus 1 and Colossus 2. Bret did not answer with those cost comps. He named strong relationships with Anthropic and Google, and he kept the exit ramp short so SpaceX can take capacity back for its own products. Less than a one-year payback on these investments, in his words, collapses the usual capex scare story. Old Nvidia generations still holding pricing, he said, is why residual value and GPU financing suddenly look like an asset class instead of a write-off risk.

He also said SpaceX is Nvidia exclusive on the allocation side. SpaceX does not control chip supply. It controls standing the power and the buildings up once the allocation lands. That is the vertical-integration pitch again: GC on the building, power on site, model to the end customer, orbital next.

2 GW this year. 5 to 10 GW next. Orbit after that

Terrestrial is the bridge. Ending this year a little over 2 gigawatts deployed. Next year 5 to 10. Confidence, he said, comes from line of sight to power, including power SpaceX stands up itself in existing locations. Elon called power the constraint. Bret said the IPO process a couple months back still met skepticism on that point. He thinks the industry sees it now. Permitting, cooling, and community pushback make orbital the cleaner long answer. Terrestrial pays the bills and teaches the ops while Starship gets cheap enough to move the stack upstairs.

Orbital compute, in his framing, is not a 2035 science project. SpaceX is targeting first orbital compute satellites next year. The bus is largely the same V3 Starlink bus flying production satellites on the next Starship flight, with a different payload and larger solar. Everyone else in AI, he said, will admit orbital is the future and then put it many years out. SpaceX controls launch, satellites, power story, and customer path. That is why his timeline is shorter than the consensus timeline. The misunderstood piece, he closed with, is exactly when orbital brings gigawatts or tens of gigawatts. His claim is that outsiders are late.

Cost parity with terrestrial, he said, is about reusability. Falcon 9 first-stage recovery in December 2015. Bret said they started reflying boosters the year after that, and that they have now reflown, he thinks, over 500. First commercial reflight was March 2017 (SES-10), not 2016. Over 500 is his count of reflown missions, not 500 unique cores. Starship first stage already recovered and reflown. Second stage soft precision splashdown on Flight 13, towed back so engineers could crawl the heat shield. When both stages are reflying, hopefully as soon as next year, the deployment cost curve drops while terrestrial power, cooling, buildings, and real estate keep getting more expensive. He floated parity maybe as soon as next year. That is a CFO aspiration tied to reuse milestones, not a signed offtake curve.

Starship Flight 14 is the production tell

Flight 13 delivered demonstration V3 satellite payloads, relit a Raptor, and gave the softest second-stage splashdown of the program so far. Flight 14 is later this month. For the CFO, the excitement is revenue. Production V3 Starlink satellites to orbit. After that, shots at bringing back both stages later this year. Starship is the shared chassis under broadband, direct-to-device, and AI sats. Without the flight rate and reuse, the rest of the stack is a slide deck.

Cursor, Grok Bot, and why the software side suddenly matters

SpaceX closed the Cursor deal a few weeks ago, he said. A huge capable team is now inside. Product cycles are already improving. Grok Bot is out and, in his word, viral. Grok 4.6 was a notable jump from 4.5. They are talking about 4.7 shortly. He said that for 15 years as CFO the company stayed almost entirely organic, and that this year it did some M&A — AI and Cursor so far. Time-to-frontier in AI is the binding constraint. The bet is simple. If SpaceX is even close to tier-one on the model, and token cost is lower because the infra stack is cheaper, customers follow. Bret's line after the IPO was that intensity went up, not down. Friends told him to relax. He said that is a joke.

Direct to device is the same V3 bus with a different payload

Version one is already up. About 600 satellites in a year. In the US that shows up as T-Mobile T-Satellite text-and-light-voice path, disaster and dead-zone coverage, not full 5G. Next version flies next year. Full 5G quality from space. Service turn-on targeted for the first half of 2028. SpaceX bought mid-band spectrum from EchoStar, with direct-from-space and terrestrial capability already FCC-approved, he said. US first. Overseas repeats the Starlink broadband regulator slog. Bret said Starlink broadband is now in over 170 countries. Starlink's own site still prints 160+ countries, territories, and other markets. Mobility backlog is real on aircraft. He said almost every major airline, three of the four largest, with the missing one left unnamed on purpose. Cruise, yacht, train, any dead zone becomes inventory.

Physical AI makes that connectivity non-optional. Humanoids, autonomous cars, anything that cannot drop a link will not live on terrestrial-only networks. That is Bret connecting Starlink to the robot wave, not a separate consumer ISP story.

What to actually watch

The near checklist is public and brutal. Does Flight 14 loft production V3 birds on schedule. Does December annualize near $100 billion ARR with the new host deal in the run-rate. Does terrestrial compute exit 2026 a little over 2 GW and print inside 5 to 10 GW in 2027. Do first orbital compute sats fly in 2027. Does direct-to-device service actually turn on in 1H 2028. Those are clocks, not slogans. Bret just put them on a Goldman stage. Farzad's cut is the briefing. The market still prices SpaceX like a launch company. The CFO is pricing it like the company that sells power, watts, tokens, and coverage on the same vertically integrated stack.