It was an eventful week. SpaceX went public, the Cursor acquisition firmed up, and Anthropic made news of its own, so the hosts spent the first stretch of the hour on the headlines before handing the floor to a guest who builds the hardware that keeps satellites from cooking themselves. The result was two episodes in one: a market read on a newly listed trillion-dollar company, and a genuinely technical walk through how you cool a rack of GPUs in vacuum.
The IPO finally landed, and it was bigger than expected
The big event since the last recording was the listing itself. The raise came in at $85.7 billion, well above the roughly $30 to $50 billion Vlad had penciled in earlier in the year. That figure matters beyond the headline, because it removes the debt facility the model had been carrying in the near term. SpaceX can now fund the Starship and V3 scale-up out of the IPO proceeds rather than borrowing, at least in the short run, which meaningfully accelerates how fast they can move (with the caveat that the model may be missing something on the cost side).
The allocation process was its own small drama. Fills varied widely across brokers and account sizes, with some participants reporting roughly 60 percent of what they asked for at one venue and others closer to 20 percent at Fidelity, with no clean rule tying request size to fill rate. Aaron's broader point was almost philosophical: a lot of Wall Street is participating in something tangible for the first time in a while, a company doing meaningful physical things rather than another ad engine, and that changes the calculus on a long-term hold. He also admitted to a contrarian discomfort. He feels better arguing that value is being missed than watching everyone pile into a name that has already become consensus.
Cursor, orchestration, and the road to a trillion
The Cursor deal is all stock, so it does not touch the fresh IPO cash, and it was struck at roughly a $1.25 trillion valuation against a SpaceX that is now trading closer to $2.3 trillion. The hosts read it as a win-win: a clean exit for the Cursor team and very little dilution for SpaceX, with a close expected in Q3. The strategic logic is orchestration, the layer that turns a raw model into something useful inside real software workflows. Vlad's candid example was that he uses Claude not because the base model is necessarily best but because the orchestration around it is better, and Cursor is the natural complement to Grok on that front. Expect the two to merge into one platform eventually. The open questions the team wants to work through over the next week or two are what the purchase implies for internal training and inference demand, how the Composer 2.5 cost curve changes the build-out math for Colossus 3, 4, and 5, and how all of that balances against the external hyperscaler and neocloud leasing story.
Two other threads ran through this section. First, Anthropic's new restriction on foreign nationals made waves, and while it is not a SpaceX story directly, it touches a customer and rival, and it raises the talent question. Anthropic retains people unusually well, helped by both a rocket-ship trajectory and a sense of mission that can transcend pay. Can SpaceX AI attract and keep the same talent, and will the Cursor people stay? The team's view is that mission and, crucially, access to compute will matter. Engineers want to play with the biggest clusters, and SpaceX AI is likely to have the most compute available of anyone, which tends to feed everything else (Anthropic is already buying capacity from them). Second, the trillion-dollar gap. Elon floated $1 trillion in revenue by 2030, well above the model's figure, and the team increasingly thinks the difference has to come from the AI side rather than Starlink. Existing AI contracts already represent about $26 billion in annual revenue running out to June 2029, and total run rate could reach the $50 to $60 billion range by year end, but closing the rest requires an order of magnitude more. Linear EWS compute build-out will not get you there. The exponential, if it comes, lives in AI apps and the Composer-style products, the question being whether something like Grok Build plus Composer can go from a roughly $4 billion run rate to $40 billion.
Enter the thermal engineer
To ground the AI satellite analysis, the team brought on McKenzie Wilder-Sandberg, a spacecraft thermal analyst about ten years into his career, five of them at Starlink, where he was around for the launch of roughly 1,700 satellites (many since demised, as he put it, because all things must come to an end). He now works at Advanced Cooling Technologies in Lancaster, Pennsylvania, a well-known maker of spacecraft thermal hardware that was recently acquired by Blackstone Energy Transition Partners. In his words, this is maybe the most interesting time it has ever been to be a spacecraft thermal engineer: everyone has an opinion, and everyone has a problem.
The context for the segment is the analysis Mach33 published this week on the AI satellite, sometimes called the AI1 sat, after SpaceX revealed its new form factor a couple of weeks ago. SpaceX was unusually generous with specs, disclosing everything from peak power to radiator area and the 250 watts per square meter figure you can use to back into solar array area. The team had been modeling the satellite bottom-up from Starlink and the literature, and now that many figures are confirmed, the exercise becomes checking the residual assumptions, namely how mass-dense the radiator and solar array are. Mass is close to everything in spacecraft design, and the lighter those two components get, the closer the design sits to frontier-level technology. The published analysis framed the core trade-off as hitting roughly 2.1 tons at 70 kilowatts per ton: you can carry lighter solar arrays with a heavier radiator, or a lighter radiator with heavier solar arrays, but you are working along a line.
Why the radiator unfolds
McKenzie's first reaction to the radiator was that this looks like the first time SpaceX has flown deployable radiators on a satellite. On Starlink, the radiator is the chassis: heat moves a very short distance, maybe 10 to 50 centimeters, from the source to the radiating surface. This new design instead has hardware that origamis out from the spacecraft and does the radiating at a distance, which means throwing heat perhaps 10 meters from the GPU box in the middle out to the tip of the radiator. That is not exotic in principle, the ISS does it with deployable radiators and a pumped liquid ammonia loop, but for SpaceX it implies a real fluid system: a pump, moving liquid, and careful hydraulic design to make the cooling loop work across that span.
On working fluid, McKenzie expects water rather than ammonia. Water has fine viscosity, so it takes little pressure drop to push through tubes, and an excellent specific heat, so it absorbs a lot of heat without much temperature rise. Its one big problem is that it freezes at 0 degrees Celsius and expands as it does, which can burst pipes (anyone who has winterized a Midwest basement knows the failure mode). You can suppress the freezing point with antifreezes like propylene or ethylene glycol, or alcohols like methanol, or with purpose-engineered fluids, though many of the best of those were PFAs, the forever chemicals, and 3M has pulled a lot of the good ones off the market even as new ones appear. As for why the ISS chose ammonia, McKenzie was honest that he does not know for certain. Ammonia has a wide operating range and is great inside heat pipes for its heat of vaporization, but here the fluid stays liquid, so that benefit does not apply, and ammonia is extremely toxic, which has caused real operational headaches on the station whenever a little leaks into the pressure vessel and the crew has to retreat to a safe section.
Run them hot, and let them degrade
How much chip life do you trade away by running hot in orbit? This is a multidisciplinary question that ground data centers wrestle with too, but it bites harder in space because cooling is a large fraction of launch mass and cost rather than a rounding error. The rule of thumb for silicon, if you squint, is that every additional 10 degrees Celsius roughly halves the lifetime, so 20 degrees is about a 4x reduction. Hotter chips need less radiator but fail more often, and on the ground you simply hot-swap a failed blade, often without powering down the rack. In space that is enormously harder, and McKenzie doubts we will see AI satellites hot-swapping blades, because designing a serviceable spacecraft and a rendezvous-and-repair vehicle is a lot of extra work. The more elegant approach is to let the hardware degrade on a planned curve. Solar arrays in low Earth orbit already degrade steadily as ionizing radiation damages the cheap silicon cells, so if you design the GPUs to lose usefulness on roughly the same timeline, you can build a satellite that simply reaches end of life as a whole in about five years, which he considers a perfectly fine design.
A few related details. Custom silicon, including the AI ASICs Tesla is building, can be designed to tolerate higher temperatures, but there is no free lunch: you make sacrifices elsewhere, such as larger, more spread-out transistors, and everyone building GPUs faces the same physics. Meanwhile chips keep getting smaller and more powerful, so power per unit area keeps climbing, which pushes the cooling problem the wrong way. And the cooling almost certainly goes direct-to-die, with a lidless package so the thermal interface material contacts exposed silicon rather than passing through an extra layer of plastic or metal. That is already common for cutting-edge terrestrial data center GPUs, and there is no reason space would be different, since you always want to cut interfaces to maximize heat transfer.
Which way to point
Orientation is a genuine trade space because these satellites fly in a dawn-dusk sun-synchronous orbit, riding largely over the day-night terminator, with the sun coming from one direction, the residual atmosphere from another, and the Earth from a third. McKenzie laid out three options. Point the broad face of the radiator straight at the sun and you maximize solar heating, the thermally worst case, useful only because solar panels want that orientation and could share hardware (not relevant here, since the radiator and array clearly point different ways). Turn the radiator knife-edge to the sun and it is thermally optimal, but the broad side then plows into the residual atmosphere in a maximum-drag configuration that hurts orbital lifetime, fuel use, and pointing. Or point the broad face at the Earth, which keeps both sun and atmospheric drag off it, at the cost of staring at a planet that radiates at about minus 30 degrees Celsius, roughly 240 Kelvin, filling around 85 percent of that side's view. SpaceX appears to be doing either the high-drag or the Earth-pointing version, and in both cases direct sunlight lands on the thin edge of the radiator rather than the broad face. The punchline is clean: every watt of sunlight or Earth infrared the radiator absorbs is a watt of GPU you cannot run on the same radiator, so you either grow the radiator, run the chips hotter, or fit fewer chips.
Build it or buy it
An audience question repeated a claim that StarCloud's core IP is radiators supposedly 10x lighter and 100x cheaper, which set off McKenzie's alarm bells, but the more interesting half was how SpaceX will source and integrate radiators given its love of vertical integration. His read is that SpaceX already has most of the expertise a liquid system needs. It builds fluid-handling valves in-house for propulsion and may already have them from Dragon. Pumps are the gap, and the space-grade pump market is dire, with units that cost a fortune and take two years to arrive, so he expects SpaceX to start with an upgraded terrestrial pump, or to work with a pump company, and eventually bring manufacturing in-house by acquisition or by growing the capability.
He was pointed about the industry standard, which is that space liquid radiators today are not very good and SpaceX will design a better one. The ISS radiator runs cold, because its ammonia loop ties into an internal water loop that cools the computers and experiments, capping the liquid at roughly 4 degrees at the low end and maybe 40 degrees at the high end. Running his own numbers on the figures shown on screen, a 110 square meter liquid radiator (around 220 square meters once you count both faces) implies an average radiating temperature near 72 degrees Celsius, give or take, which is both reasonable and far hotter than the ISS. That is the point: you should not buy this off the shelf, because off-the-shelf performance would be terrible, but the numbers are not crazy. As an aside, the team's own model had landed on a higher radiating temperature, and McKenzie's gentle warning was that if it came out around 132 degrees, that is probably too high.
On taking a commercial off-the-shelf GPU to flight, his answer was reassuring. Assuming the silicon, package, and solder balls resemble the ground part, much of the thermal hardware can look substantially the same, because hyperscale data centers already run closed liquid loops. The cooling block, the little fluid-filled box you paste to the chip, is the same idea as in a water-cooled gaming PC, except the space version is probably aluminum rather than copper to save mass. Cold liquid in, hot liquid out, fluid as thermally close to the die as possible, then off to get cold again, where the only real difference is that the heat sink at the end is a radiator instead of a swamp cooler or an air conditioner. From the liquid's point of view nothing changes; only the spacecraft-level and external hardware do. The design effort concentrates in the cooling block and the thermal interface material, where the goal is to minimize the temperature gap between chip and radiator. He spent a satisfying minute on interface materials, the gap pads, gels, and epoxies that fill the microscopic valleys between two surfaces that only look flat. On the ground, air bridges those gaps and sometimes you can skip the filler; in vacuum there is nothing, so a high-performance interface material that shaves even a degree off the chip-to-block drop can be worth the cost, and SpaceX likely moves toward the high end.
Wet noodles and thermal cycles
Both the solar array and the radiator share a structural headache: they are big flat sheets cantilevered off one point, and despite stiffening structure they behave like long, floppy wet noodles. As one GNC engineer told McKenzie, a Starlink spacecraft is really a solar array with a satellite on the end, and most of the guidance work is keeping the array pointed. Adding deployable radiators makes the satellite ring like a bell, which demands either more mass in the pointing system or more rigidity in the deployables than the cells alone would need. Which constraint bites harder, solar or radiator, is unclear; both are real.
The rapid-fire round covered the physics nicely. Yes, space is colder but worse for cooling, because vacuum carries no heat by conduction or convection, leaving radiation as your only tool; a hot cup of coffee on Earth sheds heat roughly one third by conduction, one third by convection and evaporation, and one third by radiation, so in orbit you are working with one hand behind your back. On thermal swings, the satellite is built near 20 degrees, ejects from the dispenser and cools fast, then heats quickly the first time the payload powers on, after which the goal is to hold compute loads steady to avoid power-cycling stress. The cautionary tale was the Xbox 360 Red Ring of Death, a thermal fatigue failure where repeated expansion and contraction eventually cracked a solder ball. The orbital analog is going into Earth's shadow: dawn-dusk orbits see far fewer and shorter eclipses than a normal orbit, but not zero, perhaps 5 percent of orbits dipping into shadow for a few minutes, versus Starlink's 80 to 90 percent of orbits taking a full thermal cycle every 90 minutes. Mitigations include raising chips on taller solder balls for flexibility or adding underfill between them, though underfill can cut both ways on fatigue. The culprit underneath all of it is the coefficient of thermal expansion mismatch between chip and board, the same physics as the Xbox, which makes careful reliability analysis essential. McKenzie's only stock tip, half in jest, was to invest in ANSYS, which owns both Thermal Desktop for spacecraft thermal analysis and Sherlock for fatigue reliability.
Parting thought
The closing takeaway was the one the title borrows: the thermal challenges of an orbital AI satellite are hard, but tractable, advancements on existing lines of development rather than violations of physics. McKenzie pointed listeners to Advanced Cooling Technologies (the site is at 1-act.com, and he is reachable at mackenzie.sandberg there), a shop of engineers and PhDs with R&D roots and a focus on cost-effective, higher-volume thermal systems, and he said he hopes to keep pushing toward ever higher-power designs. Vlad called him a wizard. He demurred that one day he will be a true greybeard, standing for now on the shoulders of the ones he has worked with.