250kw Starlinks??? $/Tbps Starship | Episode #21
Insights, discussion points, and conclusions
1. The 250-kilowatt claim raises a question about spacecraft architecture
Discussion · 01:27–08:26. The speakers examine a post from Musk describing satellites with 250 kilowatts of power and NVIDIA Vera Rubin computing hardware. The post also claims 10 terabits per second of connectivity in each direction. A terabit per second is a rate of data transfer. They compare the claim with their understanding of Starlink’s communications spacecraft and Starmind, SpaceX’s proposed orbital computing constellation.
Vlad focuses on the radiators, surfaces that release waste heat into space. Their orientation relative to the Sun affects how effectively they cool the spacecraft. He sees a close fit with a sun-synchronous design, an orbit that maintains a consistent relationship with local solar time. An appropriately chosen orbit near the dawn-dusk boundary can reduce time in Earth’s shadow. Extending a similar computing payload across conventional Starlink orbits would require additional work on heat rejection and power storage.
Conclusion. Both speakers treat the announcement as significant, while leaving its intended spacecraft design unresolved. They agree that the combination of high power and the requirements of different orbits deserves closer examination. The discussion establishes no tested design capable of operating that payload throughout the Starlink fleet.
2. The speakers disagree about how widely the computing payload could be deployed
Discussion · 08:37–13:21. Vlad proposes that the description could apply to a subset of hybrid computing and communications satellites in a sun-synchronous orbital group. Those spacecraft could send computing results directly to Earth instead of routing them through another satellite. Aaron considers a broader future architecture plausible, including shared manufacturing components between communications and computing spacecraft.
They also discuss power per kilogram. Vlad points out that the mass of the newly described communications spacecraft is unknown, so its power density cannot yet be calculated. Aaron cites a roughly four-ton Starmind spacecraft in the filings he has been discussing; pairing that mass with 250 kilowatts would be an illustrative comparison, conditional on both numbers describing the same vehicle.
Conclusion. No common interpretation emerges about which satellites would carry the computing hardware. Aaron distinguishes the ability to produce 250 kilowatts from the harder task of operating that power level across different orbital conditions. Vlad prefers to wait for clarification before treating the claim as a change to the entire V3 communications fleet.
3. Computing revenue depends on how much useful output a rack can sell
Discussion · 13:24–16:17. Aaron sketches an annual revenue calculation for a Vera Rubin computing rack running an open-source AI model. He starts with the rate at which the rack generates tokens, the units of text an AI model produces. Annual revenue then depends on how much capacity customers use and the price charged for its output. He discusses an illustrative range of $25M–$50M per year and cautions that the result depends heavily on those assumptions.
He also argues that this amount of computing hardware could perform substantial inference, the process of running a trained model to answer requests. Vlad remains reluctant to build a detailed analysis around an announcement whose wording and scope are still unclear.
Conclusion. The speakers see a possible commercial use for a large computing payload connected directly to customers. The revenue range is a preliminary calculation, with no demonstrated customer demand or realized revenue established in the episode. It does not include a full assessment of spacecraft costs or operating margins.
4. Repeated fairing recovery illustrates why SpaceX pursues reuse
Discussion · 16:41–18:43. Vlad discusses a Falcon 9 fairing half reportedly reaching its 40th flight. The fairing protects the payload during ascent, then separates from the rocket. The speakers note the repeated effort involved in retrieving it from the ocean and preparing it for another launch. Aaron sees that effort as evidence that recovering hardware can offer enough savings to justify an operational recovery process.
Conclusion. They regard the milestone as an encouraging example of reuse economics. A fairing’s flight history does not establish how often Starship’s more complex stages can fly; that question still requires its own operational evidence.
5. Flight 14 introduces a useful payload while remaining a test flight
Discussion · 19:44–22:25. The speakers describe the upcoming mission as Starship’s first revenue-generating flight, carrying Starlink satellites intended for service. At the time of the September 24 broadcast, they expect a Monday launch subject to Federal Aviation Administration approval. They emphasize that deploying those satellites would begin building the constellation, with customer capacity increasing as enough spacecraft reach their intended orbits.
Vlad also stresses that the first flight’s economics represent a test configuration. Their estimates for a mature operating system depend on later improvements in payload and reuse.
Conclusion. They see the mission as the first opportunity to connect a specific Starship deployment with an estimate of useful capacity and launch cost. Neither regulatory approval nor the flight’s outcome is established during this episode, and its modeled cost should not be mistaken for the cost of routine operations.
6. More bandwidth per kilogram changes the comparison with Falcon 9
Discussion · 22:43–30:28. Vlad compares the modeled costs of deploying Starlink V2 Mini satellites on Falcon 9 with V3 satellites on Starship. For Flight 14, he gives roughly $5,000 per delivered kilogram and later describes that as about 3.3 times the Falcon 9 benchmark. On cost per terabit per second of satellite bandwidth, his estimate is near parity.
The difference comes from the capacity of the satellites themselves: the speakers estimate approximately four times as much bandwidth per kilogram for V3 as for V2 Mini. More service capacity in each kilogram reduces the launch cost assigned to that capacity. Subsequent rocket reuse would provide another improvement.
Aaron then sketches a possible roughly one-year payback of the modeled launch cost if the satellites enter service and earn the annual revenue per unit of bandwidth he assumes. He explicitly treats those revenue inputs as changing estimates.
Conclusion. The speakers expect a bandwidth-based comparison to look favorable sooner than a mass-based comparison. Their payback discussion is conditional on revenue and service assumptions and concerns the launch-cost estimate; it does not establish full investment payback after satellite production and other costs.
7. Fuller payloads and reuse reduce launch cost through separate mechanisms
Discussion · 30:38–32:21. The first deployment is modeled with 26 V3 satellites, totaling about 52 tons. Vlad compares that load with a possible 60-satellite deployment. For the same launch expenditure, a larger payload spreads the cost over more delivered mass. Reuse can reduce the expenditure itself by spreading hardware costs over multiple flights.
He uses an illustrative average of 12 flights per stack, where the stack comprises the Super Heavy booster and the Starship upper stage. With the smaller payload, his chart gives roughly $500 per kilogram; with 60 satellites, it falls below $200 per kilogram. The analysis blends the two stages’ flight histories because the booster and ship can fly different numbers of times.
Conclusion. Their projected cost reductions require progress in both payload capacity and repeat flights. A catch alone would not establish those unit costs because the model also depends on subsequent use of the recovered hardware and the mass delivered on each mission.
8. The cost forecasts depend on an assumed path to operational reuse
Discussion · 25:54–26:31; 32:31–38:54. Vlad presents a base estimate of roughly $650 per kilogram in 2027 and around $300 per kilogram in 2028. He then explains the assumed reuse progression, using Falcon 9’s history as a reference while acknowledging that Starship involves recovering both stages. These are scenarios for future operations, with a range of possible outcomes.
A deployment comparison applies the different cost cases to the mass of 1,000 V3 satellites. Vlad cites roughly $10B using the first flight’s modeled economics and about $3B at the Falcon 9 mass-delivery benchmark. He cautions that V3 satellites would not actually fit that Falcon 9 configuration, making the latter a hypothetical mass-cost comparison. He describes improvements over that benchmark exceeding twofold in the 2027 Starship base case and reaching roughly fivefold in 2028.
Conclusion. The speakers expect the first expendable mission to be a starting point for measuring progress. The lower future costs remain contingent on the assumed payload and reuse path. Falcon 9 provides a benchmark for mass delivery, with physical payload fit limiting what the comparison can mean operationally.
9. The next flights can test the payload and recovery assumptions
Discussion · 39:33–44:00. Vlad identifies payload per mission and the timing of reliable ship catches as evidence to track. His 2027 base case assumes about 50 V3 satellites per flight, compared with 26 on the first deployment. Vlad expects SpaceX to commit to a much larger production effort once it can reliably recover both stages. He presents this as his interpretation of the program.
The speakers also discuss the importance of recording booster and ship reflights separately. Frequent booster reuse could lower costs even if the ship’s reuse develops more slowly. Aaron suggests that rising orbital bandwidth could support stronger Starlink revenue than some forecasts anticipate, but gives no settled estimate and acknowledges uncertainty in the figures he recalls.
Conclusion. They propose updating the model as flights reveal how much payload Starship can deliver and how frequently each recovered stage flies again. More orbital capacity supports their commercial thesis, although the episode does not establish how quickly customer demand would absorb it or what prices would result.
10. Florida launch dates depend on hardware movement and site readiness
Discussion · 23:55–25:17; 44:01–49:56. During the analysis, Aaron notices provisional dates for Flight 15 on October 19 and a first Florida Starship launch on October 30. The speakers treat those as earliest possible dates that could slip. A viewer raises the time required to transport hardware and suggests December as a possibility; the discussion does not establish a confirmed schedule.
Vlad explains that his model accounts for the additional capacity of more launch pads but does not model the logistics of manufacturing and moving vehicles between sites. Aaron speculates that Florida could eventually handle more routine launches while Starbase retains greater flexibility for testing and initial catch attempts.
Conclusion. They regard another launch site as a useful avenue for increasing cadence. The specific dates and proposed division of work between sites remain unresolved. Transport and readiness need to be considered before a provisional launch date can be treated as an operational milestone.
11. The Connectivity Keystone model will compare competing forms of internet supply
Discussion · 50:41–57:41. Vlad previews a new model combining satellite connectivity with terrestrial services such as fiber and cable. It will compare available capacity with demand through 2040. Within each country, the analysis will distinguish population densities from remote areas to dense urban centers. The planned model uses code alongside spreadsheet-style inputs to represent more interactions than the team typically includes in a conventional spreadsheet.
Aaron cites a preliminary expectation that Starlink’s aggregate capacity in orbit could double to quadruple by the end of 2027, reaching roughly 2,000–4,000 terabits per second. He warns that the model is being revised. When discussing comparisons with terrestrial networks, he corrects several dropped thousands and notes that different definitions count different amounts of connectivity.
Conclusion. They plan to test where satellite supply can compete with terrestrial services and how that balance changes by location. The episode provides no completed result showing when most internet traffic would move to space. A useful comparison will need consistent definitions of network capacity and a clear distinction between capacity deployed in orbit and service available to customers.
Editorial note
This recap follows the September 24, 2026 episode, which runs approximately 59 minutes. Timestamps locate the relevant discussions. Company claims and estimates are attributed to the speakers; they have not been independently fact-checked for this recap. Launch dates and cost projections describe expectations at the time of the broadcast. Unresolved design questions and provisional commercial estimates remain conditional.