stellar times.

context

One of the ship’s engines shut down early on ascent, and the webcast briefly said the ship would not reach orbit. Engineers then confirmed the three sea-level Raptors needed for later burns were healthy, and the call came: go for orbit.

A single sea-level Raptor fired for about 19 seconds for orbit insertion, a burn Starship had never done.⁴ SpaceQ reports it ran on propellant from the header tank in the nose. Deployment began about 34 minutes after liftoff and ran for roughly half an hour. SpaceX reports contact with every satellite over laser links, ground stations, and user terminals.

“All 26 Starlink V3 operational satellites deployed and operating nominally.”elon musk on x · 09/28/2026

The team cut the planned ~10 hour, six-orbit profile because of the ascent engine issue. The ship flew its first deorbit burn from orbit, then a flip and three-engine landing burn to a splashdown north of Hawaii at 11:57 a.m. ET. As usual for ocean landings, it tipped over and exploded afterward.

why it matters

Flight 14 checked off three firsts for Starship: sustained orbit, an operational payload with contact to all 26 satellites, and a deorbit burn from orbit with an on-target splashdown.

It also reached orbit after an engine loss on ascent, with a go given only once the engines needed for later burns checked out healthy.

the timeline

what to watch

• Satellite checkout. The V3s raise their orbits and run checkout. SpaceX says service starts within a few weeks.
• Ship catch at the tower. Open. Not attempted on Flight 14. Polymarket traders put 13% on a catch on Flight 15 (as of 09/30).
• Ship reuse. Open. Not attempted on Flight 14.
• Ship-to-ship propellant transfer. Open. The lunar and Mars dates depend on it.

context

1. Safe path first. The ship’s first burn stops just short of orbit. Its path dips about 25 km below the surface, so if anything failed it would fall back on its own.

2. Go for orbit. After health checks, one sea-level Raptor fires for about 19 seconds. About 91 m/s lifts the low point from about -25 km to roughly 262 km, for an orbit of about 262 by 275 km.

3. Come home. A roughly 100 m/s deorbit burn drops it back into the air. Belly-first through the heat, then a flip and three-engine landing burn.

why it matters

The orbit burn is small, under 1% of the ship’s speed in our model. It is the step that turns a safe suborbital path into an orbit, and SpaceX checked the health of the engines needed for later burns before committing to it.

the math, step by step

BAA physics model: a two-dimensional simulation in 0.1 second steps with inverse-square gravity, an exponential atmosphere, Earth’s rotation, and engine thrust that rises as the air thins. Vehicle figures are rounded public estimates (about 4,180 t booster and 1,810 t ship fueled, 26 V3s at about 2 t each). Liftoff, the ~T+25 minute orbit burn, the T+34 to 64 minute deploy, and the 11:57 a.m. ET splashdown follow reported times. Other times and numbers are model outputs. The ground track is a flat slice, so geography is illustrative.

context

Each V3 is designed for 1 Tbps down. Flight 14’s 26 satellites carry roughly 26 Tbps of design downlink, and a fully loaded Starship is planned to carry up to 60 V3s. Tesla Oracle put Flight 14’s addition at about 3% of the whole constellation’s capacity, 26 Tbps on top of about 841 Tbps⁶. Starpath’s math lands in the same place, at about 0.11% of network capacity per V3. Both imply a network of roughly 870 to 910 Tbps; our dazzling data model puts it lower, about 735 Tbps (BAA est.).

How a V3 moves your data. Phased array antennas steer 2,048 downlink and 2,048 uplink beams electronically, with no moving dish. V2 had 192 and 144. Six optical links at 400 Gbps each pass traffic from satellite to satellite, so data can cross an ocean without touching the ground. About 1.2 Tbps of radio links across Ka, E, V, and W bands connect to ground gateways and the wider internet.

How we turn terabits into people. Networks sell 100 Mbps plans and plan for far less per home at the busiest hour, because not everyone streams at once. Stifel uses 2 Mbps per subscriber, a 50:1 ratio, and applies it only to sellable capacity: about 30% of the total once oceans, empty land, and closed markets come out. 1 Tbps × 30% ÷ 2 Mbps is 150,000. The same math on 26 Tbps gives about 3.9 million. As a check, Starlink’s actual ratio today, about 12 million subscribers⁷ on roughly 735 Tbps of design capacity, works out to 16,000 to 18,000 subscribers per Tbps, or about 0.4 to 0.5 million for this launch.

why it matters

This partial load of 26 moved about as much bandwidth as ten Falcon 9 launches of the older V2 Mini, per SpaceX. Our own math gives 9 to 11 (BAA est.). A full load of 60 adds more than 20 times a Falcon 9 launch, per SpaceX. Starlink capacity growth now rides on how often Starship flies. Reuters reports analysts see each batch worth hundreds of millions of dollars in projected revenue.

engineering and operations

Replacement cadence. Starlink satellites last about five years. Roughly 10,500 working broadband satellites (McDowell, 09/28)⁸ means about 2,100 a year wear out, which is about 80 Falcon 9 launches a year just to stand still (BAA est.). A 3,500-satellite V3 fleet would need about 700 replacements a year, or about 12 full Starship launches a year before any growth (BAA est.).

Ground stations and backhaul. A satellite can only deliver what it can hand off. Each V3 carries about 1.2 Tbps of gateway radio links plus lasers, so traffic over the ocean rides the laser mesh to a gateway. Gateway sites, fiber, and a license in each country set where the capacity is usable.

Spectrum and beams. 1 Tbps is the total across 2,048 beams, about 490 Mbps per beam on average. At 2 Mbps busy-hour, one fully loaded beam carries roughly 245 active homes. Dense towns fill their beams long before the satellite runs out, and interference limits that protect other satellites cap how hard each beam can be pushed.

Oversubscription. The 50:1 ratio holds only while busy-hour use stays near 2 Mbps a home. Household traffic grows about 20 to 30% a year. At 3 Mbps, the 3.9 million becomes about 2.6 million (BAA est.).

Supply chain and failure modes. Sixty V3s per Starship means building phased arrays, lasers, and solar arrays at a much higher rate. A single Starship also concentrates risk: losing one full load is about 60 Tbps, against about 2.6 Tbps for a Falcon 9. Expect a few satellites per batch to fail checkout or deorbit early.

the numbers

what to watch

• Checkout. Each satellite still has to unfold antennas and arrays, link up by radio and laser, raise and circularize its orbit, and pass checkout. SpaceX says service starts within a few weeks.
• Real service. It depends on where the satellites sit, ground gateways, and demand. Satellites spend much of each orbit over oceans and empty land, so even 3.9 million subscribers is an upper estimate.
• Subscriber count. Kalshi traders put 58% on SpaceX reporting more than 14.5 million active Starlink subscribers for Q3 2026, in thin trading (as of 09/30).

context

The figures come from Anthropic’s confidential IPO prospectus as reported by Reuters on Sept. 29. Anthropic expects to spend at least $518B over ten years on AI infrastructure with six partners. About 80% of that is non-cancelable or owed regardless of use.

SpaceX’s own IPO paperwork had described about $45B, or $1.25B a month for three years. The new figure is up to $84.5B, and most of it can be ended with 90 days written notice. The capacity traces to May 2026, when Anthropic took exclusive use of Colossus 1, a SpaceX facility with more than 220,000 Nvidia GPUs on over 300 MW. In June Anthropic also began scaling GB200 capacity on Colossus 2. SpaceX’s S-1/A says the monthly fee covers about 325,000 GPUs across both sites. SPCX shares were up about 2.4% by midday Sept. 29, with TD Cowen also starting coverage at buy that morning⁹.

How a compute lease works. A data center packs GPUs, power, and cooling. The AI lab pays a monthly rate for reserved capacity, used for training models or for inference, meaning answering users. Take-or-pay deals bill the full amount no matter what. A 90-day cancel clause lets the renter walk with a quarter’s notice.

why it matters

Most of Anthropic’s $518B plan is take-or-pay. The SpaceX line is the flexible piece: it is largely cancelable on a quarter’s notice, so it is the commitment that can shrink fastest.

engineering and operations

Power and cooling. An H100 or H200 board draws about 700 W. A full 8-GPU server draws about 10.2 kW, or about 1.3 kW per GPU, and GB200 racks run about 1.7 to 1.8 kW per GPU. SpaceX’s filing lists about 100,000 H100 at about 130 MW, plus about 110,000 GB200 at about 210 MW and 110,000 GB300 at about 220 MW. That is about 1.3 kW per H100 and about 2 kW per GB200 or GB300, the same as the server math, so the filing’s megawatts read as IT load. Across the roughly 325,000 GPUs that is about 560 MW of IT load. With a PUE of 1.2 to 1.3, meaning cooling and power losses add 20 to 30%, the meters read roughly 670 to 730 MW (BAA est.).

Getting the power. Memphis runs on a mix of grid power from MLGW, which buys from TVA, plus on-site gas turbines and Megapack batteries¹¹. Each extra 150 MW of grid power needs a TVA board vote, since any new load over 100 MW does¹⁰, and the substations and transmission work behind it take time.

Failure modes. At about 325,000 GPUs across two sites, failed GPUs, memory, and network links are a daily event. Inference shrugs off a lost server more easily than a training run does. A cooling or power fault takes out a block of megawatts at once, which is why the cooling plant gets the redundancy.

commitments by partner, and the math

SpaceX’s own filing (S-1/A, 06/03/2026) says the fee covers about 325,000 GPUs across Colossus 1 and 2, which works out to about $5.27 an hour. Per-MW rent uses the filing’s compute power, which is IT load before cooling. Real pricing per GPU type is not public, so treat these as a sense of scale. 730 is the average hours in a month. Reuters cites six partners, all six listed here.

what to watch

• The 90-day cancel clause. Walking away means about three more months of fees: roughly $3.75B at the old rate, or about $7.0B at the new ceiling (BAA est.).
• The filing itself. These numbers come from a confidential prospectus seen by Reuters, not a public document.
• The IPO date. Kalshi traders put 80% on Anthropic officially announcing an IPO before Jan. 1, 2027 (as of 09/30). Polymarket traders put 81% on an IPO by Dec. 31, 2026 (as of 09/30).

context

At the All-In Summit on Sept. 15, asked why Tesla and SpaceX stay separate, Musk said:

“With all this collaboration, on so many levels, who can imagine what action one might take when there’s so much close collaboration in so many areas.”elon musk · all-in summit · 09/15/2026

Gwynne Shotwell said the businesses are integrating “faster than I thought.” On the July 22 Tesla call, Musk described “more and more overlap” and said any combination needs “the appropriate process.” Neither company has announced or filed a merger agreement.

How Tesla’s new voting option works. SEC staff issued a no-action letter on Sept. 29 for Tesla’s voluntary retail voting program¹². Retail holders can opt into a standing instruction to vote with the board, on all matters or on all except contested director elections and mergers. It is free, can be overridden per proposal or canceled anytime, and comes with an annual reminder. Other issuers can use the same framework.

why it matters

The SEC letter does not mention SpaceX. Its effect on any future deal depends on how many holders pick “all matters” over the version that carves out mergers.

market odds and the merger math

Market odds are yes prices: the midpoint of the bid and ask, or the last trade when the spread is wider than 10 cents. Polymarket rows update live on this page. Kalshi rows are saved as of 09/30/2026, 5:05 p.m. ET. Odds move daily and replace the older secondhand odds in our draft. The two premium thresholds sit about one point apart. FinanceFeeds calculation, Sept. 2026. Estimate only.

what to watch

• A proxy filing.
• A special committee.
• A definitive merger agreement.

context

Starship Flight 15. SpaceX has not given a date. An Oct. 19 FAA air traffic planning date was reported, and trackers now show TBD; T-Minus Zero shows December 2026, day TBD¹⁵, and Next Spaceflight lists no earlier than October 2026¹⁶. Ship 42 and Booster 22 are assigned. Musk said the first attempt to catch the ship with the tower arms comes on Flight 15 if Flight 14 went well. The catch needs a full-duration flight and FAA approval.

What the markets say about Flight 15. Kalshi traders put 46% on Flight 15 launching before Oct. 30 and 74% before Nov. 13 (as of 10/01, 12:12 a.m. ET). Polymarket traders put 13% on the tower arms catching the ship on Flight 15 (as of 09/30). Musk has said the odds of catching the ship on the first attempt are at least 50 or 60 percent¹⁷.

Also this week. The Federal Register published a Finding of No Significant Impact for Relativity’s Terran R at Cape Canaveral on Sept. 28. A FONSI is an environmental finding, not a launch license.

the full calendar

what to watch

• A SpaceX date for Flight 15. The Oct. 19 date is an FAA air traffic planning date, not a SpaceX announcement.
• Crew-13 on Thursday, and the V3 checkout through October.