This was a deliberately news-heavy week, so Aaron and Vlad spent the first half of the hour on the headlines before turning to the analysis Mach 33 published on the terrestrial side of SpaceX AI. A quick disclaimer up top: it was brutally hot in the UK, Vlad was running warm and a little grumpy, and listeners were politely asked to hold the "Star Cool" jokes. What followed was a tour through three big stories that all point the same direction, which is a company suddenly armed with cheap capital and deciding how fast to spend it.
Investment grade, on the first try
The piece of news that genuinely surprised Vlad was the ratings. Moody's, Fitch, and S&P all handed SpaceX an investment-grade rating straight off the bat, which is unusual given how much chatter has framed SpaceX as a fundamentally risky business. Getting a clean first-ever investment-grade mark from all three at once was a strong signal, and the detail that stood out in the Moody's report was how hard the agency leaned on Starlink. Moody's treated Starlink as the primary cash flow generator, the asset that makes the lending low-risk, even though much of that cash is being routed toward things that do not generate revenue yet, such as Starship development and orbital compute. In other words, the debt providers underwrote the boring, recurring connectivity business and let it carry the bets.
The bond, the bridge, and a $100 billion war chest
The debut bond went very well. SpaceX went looking for roughly $20 billion and drew on the order of $90 billion of interest, then locked in about $25 billion. The purpose was largely housekeeping: pay off the $20 billion bridge loan that showed up in the Q1 financials, the one taken to pull IPO cash forward, and keep the roughly $5 billion left over. Now that the IPO cash has actually landed, SpaceX negotiates from a much stronger position, which is exactly why the terms were so attractive.
Aaron's favorite detail was the maturity structure. Close to half of the $25 billion is not due for ten years, and a further chunk pushes out to 20 and even 30 years across the longer tranches. If you believe the conservative cash flow profile in the Mach 33 model, locking in money you do not have to pay back for two or three decades is something SpaceX takes every day of the week. It also vindicates a modeling choice the team made a while ago: a flexible debt line item that let SpaceX pull forward cash whenever it wanted. The Giga model had long suggested the most they would ever need to draw was around $50 billion, and Vlad's debt facility assumed a pull of somewhere between zero and $20 billion once the Starship and V3 Starlink designs were finalized, on the logic that finalized designs make it economically rational to accelerate as hard as possible. Reality is running a little faster than the original model, so some tweaking is in order, but the instinct was right.
Stack it all up and the cash position is the headline number: roughly $100 billion, with about $85 billion from the IPO and $15 billion from Q1, and closer to $105 billion once the bond proceeds settle. With interest rates nowhere near what SpaceX can earn deploying that capital, taking on debt right now makes complete economic sense, which is also why the bond drew such heavy demand. As Vlad put it, if they had not done this, he would have been upset, because a working Starship means this money belongs in accelerating productive capacity rather than sitting idle.
Where the money goes
So what does SpaceX actually do with a nine-figure-billion war chest over the next three months? The fastest way to spend at scale is a large chip order, the kind of lock-in where a vendor simply ships you silicon. Beyond that, the spend is mostly acceleration of things already underway: more Starlink V3 satellites, build-out at TerraFab including possible ASML purchases, the solar plant, the rumored launch-facility purchase in Louisiana that would mean starting close to from scratch, and expanded V3 production in Redmond. There could be fresh activity at the Cape too, perhaps breaking ground on a manufacturing facility next to the existing Giga Bay. Vlad's framing was that none of this should be read as a surprise, because every one of these figures, even the $100 billion cash balance, looks minimal next to the free cash flow SpaceX should be generating by 2030. One housekeeping note from the news list: the AI satellite now has an official name, Star Mind, a coinage Vlad won out over Aaron's preferred Star Think.
Starfall drops
The capsules are the fun part. Starfall is a fleet of reentry capsules shaped like frisbees, or maybe hockey pucks, with the heat shield on one face. Starship goes up, whips out a batch of them, and they come home. The first test happened Tuesday, and to be clear about one thing the hosts corrected mid-show, the striking image going around is a render, not a photograph, though Aaron has seen the real hardware stacked and says it looks very similar. Confirmation of the landing was thin, but a friend of the show, Starbase Brewing, publicly announced it had a payload of yeasts and other materials aboard, which at least suggests the flight was not a total loss. That kind of "it was in space" novelty product is the trivial commercial edge of this. The real prize is everything behind it.
Vlad's thesis is that Starfall is the first serious, well-capitalized step toward an in-space manufacturing industry. He has been clear-eyed about the early reality: when he posted excitement about the test on X, someone told him not to be naive, that the first uses will be military testing, and he agrees the opening phase is tests and emergencies. But point-to-point delivery, taken alone, has roughly the same unit economics as launching to low Earth orbit and coming back. What makes Starfall genuinely different is the prospect of producing something in orbit that cannot be made on Earth and then returning it, and the fact that the most dominant player in the industry is now the one taking that step.
Why microgravity makes a better product
Someone on X asked Vlad how much extra energy it takes to make these products in orbit, and his answer was that the question is backwards. Manufacturing the same things off Earth does not cost more energy; the absence of gravity simply yields a better product. You get purer alloys that are stronger and lighter at once because they are more perfect at the molecular level. You get near-perfect semiconductor wafers, which lets you squeeze more compute onto the same physical silicon. And, the example Vlad is most excited about, you can grow a human retina from a patient's own stem cells. On Earth that structure collapses into mush under its own weight as it forms, but in orbit it can grow into the correct shape, potentially restoring sight to someone whose retina has failed. Same energy as on Earth, purer result.
Aaron zoomed out to the economic thesis that ties it together. Everything humanity has ever built was made under gravity, and a lot of engineering is just fighting it. Commercially, space has so far meant sending atoms up to create value delivered as bits: Earth observation, routed data, Starlink connectivity. Orbital data centers extend that logic, sending atoms up so that bits can go up, become more valuable, and come back down. The next progression, moving atoms up and bringing atoms back, has historically cost so much that the only viable markets were tiny and high-value, things like retinas or the pharmaceutical crystals Varda works on, where you can pack enormous value into a single kilogram. The number that gates everything is return cost. Today, bringing mass back runs around $50,000 per kilogram, a figure anchored to the human-rated Dragon capsule because that is essentially the only way home right now. If SpaceX drives up-mass cost toward roughly $100 per kilogram, the return trip could carry only a modest premium, perhaps as little as 20 percent more at $120 per kilogram, or maybe a 2x premium at $200. Land anywhere in that range and the economics invert, and entire new markets open up.
Test pods, not factories
Vlad was insistent on a point of calibration. The one-ton Starfall pucks you are seeing now are test capsules, full stop. They do not mean purer drugs, human organs, pure alloys, and perfect wafers are about to rain down. What happens inside those pods is testing of what could eventually be mass-produced in orbit, and the real production will run at a far larger scale, effectively space stations or entire Starships configured as orbiting factories pulling fibers or growing silicon crystals. He pegs the meaningful industry as a 2030s story, perhaps late 2030s, and reminded listeners that the whole point of the podcast is to look past what already exists today.
Aaron added nuance in the other direction, arguing some of it could arrive sooner. There is hypersonics testing for defense that needs exactly this kind of reentry heat, and point-to-point logistics where the military pays a high markup to land cargo near troops on a moment's notice, both small markets but real. More tellingly, the puck shape is designed to fit a PEZ-style dispenser rather than a clamshell. Every year SpaceX does not build a clamshell Starship implies the large-structure use cases keep slipping, but designing for a PEZ dispenser suggests in-space manufacturing testing is being pulled forward, possibly scaling around 2028 or 2029 even if it is still testing then. On the competitive question, the hosts figure SpaceX could eat Varda's lunch if it chose to, Varda having processed a pharmaceutical in orbit and returned it a couple of years ago, though SpaceX might instead prefer to be the transport layer underneath everyone. And on materials, crystal growth is a standout: compounds like gallium arsenide and the nitrides are heavier, grow smaller, and yield poorly under gravity, which is part of why that chemistry is so expensive. Grown in space they get bigger and purer, and a single meter-scale crystal, sliced up, can be worth millions.
A Martian forge
Vlad let himself nerd out past the investment horizon with a Mars scenario. If microgravity gives you alloys that are simultaneously lighter and stronger, then the best place to eventually mass-produce them might be Mars orbit. The logic runs on the rocket equation. Mars has only about 38 percent of Earth's gravity (corrected on air from his earlier "half") plus an abundance of untapped iron. You mine cheap raw material on a world with a shallow gravity well, lift it to orbit for far less propellant than the same move would cost on Earth, manufacture the alloys in microgravity, and ship them home using minimal delta-v because you are already out of the well. The catch is time, roughly a six-month transit if you are still using chemical rockets rather than nuclear, but the cost per kilogram could be an order of magnitude lower. The vivid illustration: because Mars gravity is so much weaker, a Starship could in principle reach Mars orbit without a booster at all, refuel from tankers, and return, whereas on Earth the booster is most of the stack. A year ago the team had reached for a word to describe this, and it landed here as Mars the forge world, the solar system's foundry.
Drawing a line in the sand on terrestrial compute
The week's published analysis was an attempt to fix a real weakness in the big SpaceX model: the terrestrial build-out of AI compute. Orbital data centers are the long-term future, but the near-term play over the next year or two is 100 percent terrestrial, and now that xAI sits inside SpaceX the two are intertwined for the next 5 to 10 years. The team are space specialists who have spent years building modeling logic for space companies, and they had been leaning on industry benchmarks for the ground side, benchmarks that often turn out to be wrong. So they looked at historical growth rates and projected how much terrestrial data center capacity SpaceX AI could realistically build over the next two to three years.
Aaron flagged a related shift in tone he has noticed: SpaceX holders increasingly trying to get institutional investors to view SpaceX AI through a hyperscaler lens. Part of that is convenience, because you can lift a standard spreadsheet straight from a neocloud investment and drop it onto SpaceX, and Wall Street loves an easy pitch. Part of it is that the terrestrial story is simply the nearer-term, more contractable opportunity; signing a two-year deal is far easier than one that has to be proven out over five.
The central modeling question is how much built capacity goes internal versus external. Internal means Grok, Cursor, and other in-house uses. External means leasing, the Google, Anthropic, and Reflection AI deals. That split defines what kind of company SpaceX AI actually is. Mostly external makes it an emerging hyperscaler. Roughly half and half makes it a hybrid. Mostly internal makes it a cost center, where today's capex eventually feeds apps built on top of the AI or the training of the next model. The team built three cases around different internal-draw assumptions. They were candid that starting from the supply side is not ideal and that the real goal is a demand model, but as everyone in finance keeps telling them, no one knows what AI demand will be, so for now they used deliberately broad growth-rate ranges to bound where the business could land.
The Falcon 9 playbook
The cleanest conclusion is an analogy, and Aaron noted he dislikes reasoning by analogy but sometimes it is the only light in the dark. SpaceX AI looks like it is running the Falcon 9 and Starlink playbook: innovate, iterate, and build capacity as fast as humanly possible, use as much of it internally as you can because owning the end customer captures more margin, and lease whatever is left over to outside players. With launch, the leftover capacity became Starlink, a recurring-revenue machine. The history rhymes on the numbers too. Falcon 9 started around a 50/50 split between external commercial launches and internal Starlink around 2020, drifted to roughly 70/30 in Starlink's favor, and now sits near 75 percent internal. Apply that arc and you would expect SpaceX AI to start near 50 percent internal and iterate toward a future where its own equivalent of Block 5, cheap orbital compute, gets used overwhelmingly for itself. Counterintuitively, you want it to end up as a cost center, because that would mean the company is making most of its money in higher-margin AI applications. One feature that makes the external deals especially flexible: many of them, including Anthropic and most likely Reflection AI, carry roughly a 90-day cancellation window.
The compute you can't lease out
Here is the finding both hosts called the most interesting of the week. Picture all the compute in your data centers. Vlad had assumed day-to-day inference, people using your existing models, would eat most of it. It does not. By the team's estimate only about 40 percent of XAI's compute supply goes to inference. The rest is spoken for in ways you cannot simply rent out. You need a large, dedicated block to train the next model, say Grok 5, running flat out for weeks, and you cannot run that on the same hardware serving inference without degrading the live product that people are paying to use. On top of that sits the research load: the constant internal testing and the models built but never shipped, all of which consumes compute that can neither be handed to external customers nor pulled out of the reserve a training burst requires.
The hard numbers, expressed in H100 equivalents, put a fuzzy ceiling around 3.6 million chips based on power constraints. A snapshot of that has roughly 582,000 held back for training bursts, and that figure sits inside the internal half of a hybrid split, with the other half going to leasing. The practical implication is that bursty training leaves a big chunk of silicon idle at any given moment, and if you want to keep investing in future models you simply have to hold it rather than lease it away. This is why the 90-day cancellation windows matter so much, and why Elon reportedly said almost verbatim not to get ahead of themselves on these deals because they may want the compute internally for a training burst.
Aaron took it one layer deeper. Elon's push to write C directly to the chips, bypassing JAX, points by the team's rough and still-unrefined math toward materially faster training, somewhere in the 1.5x to 3x range. Why that matters here: if hundreds of thousands of chips are being held for bursts, faster training makes those bursts shorter and more predictable, which means you can hold less in reserve and recycle the hardware back into use sooner. Training speed becomes a lever on how much idle capacity the whole system has to carry.
What's next
The plan for the coming weeks is to turn from supply to demand and dig into the revenue side: what the terrestrial data center business actually earns across different use cases over the next few years, which the team considers one of the bigger gaps in the current model. As Aaron signed off, this one ran a little extra geeky, but when the conversation turns to Mars, that tends to come with the territory.