China catches a booster
The first item was technically last week's news, bumped by the Peter Hague conversation: China successfully caught a booster. The vehicle came back smoking, looking to Aaron like a smokestack with a fire going on top, but the catch worked. How fast China got there depends on where you start the clock; the most enthusiastic accounts say a year, while Aaron puts it closer to three years net-net from when the work actually began. Either way it is a quick turnaround, and it confirms that reusability is the path forward and that China, already number two in total up mass to orbit behind SpaceX, is cementing its lead over everyone else in the chase, with US startups, established players, and Europe well down the list. Aaron still gives SpaceX something like a 10-year lead, with Falcon 9 booster reuse effectively old news and Starship coming. Alex was blunter: do not sleep on China in any capacity, because, as he put it, they have powers of capitalism we can only dream of (an audience member offered the correction that these are the powers of socialist capitalism, which the hosts took on board). The tie back to the show's beat is that China has put roughly an order of magnitude more money behind orbital data centers, and cheap up mass and orbital compute go hand in hand.
Launch day, weeks ahead of schedule
Starship's 13th test flight was set for the evening of the recording, and the timing itself was the first story. On historical cadence Aaron would have penciled this flight in for the end of July into August, so it is running a few weeks ahead, with more boosters and ships in the gigabays and coming off the lines. He expects the pace to keep picking up. The flight path is the usual suborbital profile, nothing exotic there.
The flip, the relight, and a bubble in the plumbing
What Aaron is watching hardest is the booster's flip and relight, because that is where last flight went wrong, with Raptor engines blowing up rather than relighting. Peter Hague's theory from last week's episode was propellant sloshing, and it fits: the V3 Raptors are noticeably more powerful, the last ascent looked almost too fast, and the flip maneuver felt too aggressive, and if an engine inhales a bubble instead of fuel it will blow up no matter how well built it is. So the thing to look for is whether the flip comes in gentler. A secondary concern is the upper stage, where one Raptor started and stopped early last time, leaving five of six, though the ship still hit its trajectory and landed right at the buoy. The stakes are larger than the test itself: by Aaron's math, getting just the booster reusable on Starship makes orbital data centers cost effective almost certainly within about three reuses, which makes this one of the more important line items for orbital compute becoming viable at all.
Real satellites this time
The headline cargo is 20 Starlink V3 satellites, and this time they are real. Previous tests flew chunks of steel mimicking the size and weight. These will not enter operational orbit or add capacity, but a successful deployment says the V3 design is ready enough to test, that the larger satellites unfold properly, and that the phased arrays work, with an attempt to hit a ground station over South Africa on the pass. The Giga models have long assumed V3s might start at the very end of 2026 but do not really scale until 2027, so a real deployment test now sits comfortably inside the model's range of error. Aaron was candid about falsification: the team's Starlink capacity estimates are aggressive relative to others (the team considers them conservative), and if no functional V3 batch is up by December, maybe January, it would be time to start questioning the ramp.
Delay math, answered live
An audience question asked what a Starship delay would do to the base case for SpaceX growth. The model has functional satellites starting late 2026 into early 2027 with real scale after that, and revenue anchored to capacity against a falling cost curve, so a delay of a year moves the whole revenue logic back a year rather than breaking it. Aaron then did some live math he cheerfully predicted would be wrong. A Starship launch carries about 20 times the bandwidth of a Falcon 9 launch, and SpaceX currently flies roughly 150 missions a year, call it 10 a month. One Starship launch a month would therefore roughly double the bandwidth SpaceX can already put up with Falcon 9s and V2s, which do not stop flying just because V3s start. Get to two a month and the numbers run away, which is why capacity gets so big so fast.
Thunderstorms, according to competitors
Another audience member, in the American Southeast, reported Starlink degrading substantially during thunderstorms. Aaron, a daily Starlink user in Florida who was on the service for the show itself, and who has had big afternoon thunderstorms every day for two weeks, offered his own data: speed tests do drop, from around 250 down to around 50 in heavy weather, but he has never noticed an actual interruption. The physics is that higher bandwidth spectrum attenuates in weather while lower bands penetrate better but carry less, and the V3s appear to carry multiple spectrum types, confirming a multi-link hypothesis Aaron has floated before, with software picking the right link for conditions. His summary of the lived experience: competitors trying to sell him fiber warn him about thunderstorms far more often than thunderstorms have actually hurt his internet.
How Falcon 9 winds down
A third question asked whether Falcon 9 keeps flying once Starship is operational. The team operates on external information only, and the external signs point to a scale-down: public rumors that the Transporter rideshare program will be discontinued make sense to them, since it earns little and runs almost as a service to the industry, and every Falcon 9 launch consumes a newly built second stage, manufacturing space that could go to Raptors instead. Aaron's hunch is that V2 satellite deployments plateau rather than stop, with a crossover period while the V3 system works out its kinks and terminals catch up to the new spectrum. The next twelve months should show which way it breaks.
Enter the containment tax
The back half of the hour went to the week's analysis on orbital inference containment costs, which Aaron flagged as probably the densest the team has done, spanning LLM serving architecture on one side and cutting-edge spacecraft design on the other (he suggested readers work through it with an LLM for friendlier terminology, noting the team iterated repeatedly just to make it understandable to themselves). Alex, who led the write-up, framed his interest as business first: cutting through the froth, he came away confident that orbital data centers are not some pie in the sky cost center but a revenue driver, and not 10 years out but a year or two. His caution runs the other way on pricing. Serving frontier models is a race to the bottom, much of today's usage is subsidized, and cheap Chinese models keep pulling the floor down (he cited, with the caveat that he could not remember the specifics, a frontier-level model released at around 50 cents per million tokens). Alongside thermal and power, the cost per million tokens is the third big driver in the model, the dominant bar in the analysis's tornado chart.
How inference actually gets served
To follow the analysis you need one mental model. Think of an LLM as a three-dimensional spreadsheet, a tensor of a trillion or more parameters that can be cut into experts. Experts are not topical specialists like coding or finance; the cuts fall more along linguistics and syntax, the way language decomposes. When a token arrives it runs through only a small subset of the model, somewhere between 1 and 5 percent of the parameters, and expert parallelism lets many token streams hit different experts at once. The rule the team kept finding in the published work: the fewer experts hosted per GPU, the more tokens per second you can serve, and tokens per second is dollars per second. DeepSeek and the other efficiency maximalists have pushed all the way to one expert per GPU. The problem is arithmetic. The standard top-of-the-line rack, an NVL72, holds 72 Grace Blackwell GPUs, and a trillion-parameter model might carry roughly 250 experts, so on the ground operators string racks together with InfiniBand until the experts spread out.
One rack per satellite
In orbit the constraint flips. A satellite's power budget supports roughly one rack, so a frontier model has to squeeze all of its experts onto 72 GPUs. That is the orbital containment tax: contain many experts in one node and your tokens per second fall, and with them revenue. The analysis deliberately handed every benefit of the doubt to the skeptic's case and produced a ceiling: at worst, matching the tokens per second of one ground rack takes 1.8 satellites. The team's internal G1 label corresponds to the AI1 satellite SpaceX has revealed; G2 is a two-rack, 144 GPU design further up the power curve.
A 750 meter physics fence
The obvious question is why you would not just spread the model across several satellites with laser links, and the answer is the speed of light. Of the parallelisms running inside an LLM, tensor parallelism is the latency killer: everything must stay in sync within 5 microseconds or it stops working. Light in 5 microseconds gets you 750 meters out and 750 meters back, so the farthest a partner satellite can fly is 750 meters away, orders of magnitude closer than satellites normally operate. Alex's sharper point is that this kills the dream of meshing a giant inference network around the Earth: the 750 meter fence is a physics fence, not an engineering one, and the craziest lens in the world does not go faster than light. Even doubling the sync tolerance, which neither host considers easy work, only moves the fence to 1,500 meters.
That fence is exactly what Google's Suncatcher is engineering against, with smaller satellites formation flying within about 150 meters of one another and pooling GPUs across laser links. A self-described dumb question from the audience, what makes close formation flying hard, got a thorough answer. Everything in orbit moves extremely fast, so packing satellites close is like packing cars on a highway at speed, and one satellite nudged off course by debris can set off a chain of collisions among its neighbors, a thread that runs back to the team's orbital saturation model. Alex added the forces that never show up in ISS footage, drag at these low altitudes, shear, even solar radiation pressure, all constantly pulling a formation apart, plus laser pointing, acquisition, and tracking, where at 150 meters beams can cross and hit the wrong receiver. Aaron offered a side opinion, explicitly not part of the analysis: Google is choosing formation flying because it is not a mass satellite manufacturer, so it builds around hardware closer to what exists today, while SpaceX and StarCloud blaze the unproven path of much bigger, higher power single satellites. Both are attacking the same physics with the strengths they have, and if Suncatcher solves formation flying it can in theory scale to bigger models by adding nodes.
Elon grades the homework
Hours before the show, Elon commented on Aaron's X write-up of the analysis, and Alex spent the news half of the episode researching in the background what the comment did to the model. The short answer is that it confirmed it. The team had predicted the satellite design before he confirmed it: the 72 GPU single-rack package is the correct read. The comment also indicated the AI1 satellite runs a Rubin-class rack, with far higher memory bandwidth than the Grace Blackwell class the model conservatively assumed, which by the team's model raises revenue per node by roughly 2x to 2.5x, and that peak power can reach 250 kilowatts with batteries against the 150 kilowatts previously announced. Aaron cautioned that this was probably not meant as an announcement and is surely only directionally accurate, but the error runs in one direction: the team's numbers were out of date within hours of publishing, in favor of being even more conservative than they thought. On existing AI1 architecture the worst-case tax looks closer to 1.44 than 1.8, and by Alex's read the residual gap between orbit and ground on the chart is now probably a lot less. For SpaceX holders, the picture in the model just improved.
Attacking the tax
The rest of the analysis is a set of levers, stacked and never double counted, that erode the tax. The first is software, meaning the C rewrite covered in last week's analysis: writing to the bare metal is worth about a 33 percent speedup in tokens per second. It helps the ground equally, but against the containment tax it does much of the repair work on its own, and it suggests why the rewrite is happening. The second and biggest single lever is co-design: train the model so its expert count matches the satellite, in the ideal case 72 experts for 72 GPUs, one each, where the efficiency ceiling lives. That is not exotic; an open source model released the day before the show ran only about 20 experts against just under a trillion parameters. Because the team's standing thesis is that SpaceX trains faster than anyone, orbit-specific versions of every model served become an affordable opportunity cost, and the hosts registered a prediction: expect co-designed orbital model runs, likely with SpaceX's own models and Cursor first, and if a ground deal with Anthropic or Google ever translates to orbit, expect a co-design run to come with it. The third lever, rerouting, was included for completeness: host only a subset of experts on a satellite and route only that traffic to it. Stack the levers and the containment tax deteriorates to almost nothing, or at worst something like 10 percent, at which point orbit merely has to be 10 percent cheaper on cost to reach parity, a bar the team expects it to clear over time.
Panels or models
A final audience question asked whether solar panel chemistry, tandem cells or thin film, moves the results more than changes to how the models run. The answer was kind of no. Better cells matter and will play a role, but the decisive variable is matching the model to the GPUs on board. Aaron's oversimplified split: the thermodynamic side, power efficiency and thermal emissivity, reduces the cost side of the equation, while optimizing the inference itself drives the revenue side. Both feed the return; they pull different levers.
What to watch next
The launch, at 6:45 Eastern the evening of the show. Aaron will have the livestream up, expects the V3 Starlinks to be the story, and hopes for a big improvement on the Raptors and a gentler flip. He thanked the audience for bearing with an unusually deep hour, allowed that several listeners had surely fallen asleep, and signed off until next week.