Can Anyone Actually Cool a Computer in Orbit?

Today on Space Stakes: Nvidia just placed two separate bets on the same wild idea — putting AI supercomputers in orbit — and neither one has solved the basic physics problem of how you cool a computer with no air around it. Before that, in the headlines: the Space Force's missile-warning satellites are checking out faster after a rocky start, according to the program's own director, SpaceX is leaning on a South African regulator to unstick a years-long ownership fight, and Kepler just flipped the switch on a brand-new data-relay network in orbit. Welcome back to Space Stakes, your daily brief on the business of space. It's Tuesday, August 25, 2026. Two big stories share the same theme today, so we're pairing them up to see what they tell us together that neither tells you alone. Let's get into it.

First up, some genuinely interesting news out of the Pentagon's missile-tracking constellation — though it comes with a caveat right up front. Space Development Agency Director Dr. Gurpartap Sandhoo told the Small Satellite Conference in Salt Lake City that on-orbit checkouts of the Tranche 1 satellites launched in July are moving, in his words, two to three times faster than the first batch. Quick refresher if you're new to this: Tranche 1 is the SDA's proliferated constellation of small satellites meant to track missiles from low Earth orbit, and it's been plagued by delays. Sandhoo credits lessons from that first plane of satellites, built by York Space Systems, for speeding up the second. But does 'two to three times faster' tell you how close Tranche 1 actually is to being fully operational? Not really — Sandhoo didn't say. And this is a claim from the agency's own director at a conference, not an independent audit, so treat it as his assessment rather than a verified figure. Directionally, though, that's the kind of learning curve you want to see on a program built around rapid iteration and taxpayer dollars.

Over in South Africa, SpaceX sent senior director Ryan Goodnight to appear before the communications regulator, ICASA, this week, pushing it to fix the ownership rules keeping Starlink locked out of the country. Here's the core snag: South African law requires 30 percent local, historically-disadvantaged ownership for telecom licenses, and SpaceX wants an equity-equivalent program instead of actually selling a stake. South African outlets confirm Goodnight also asked for simpler gateway fees and one blanket license covering entire fleets of terminals instead of licensing them one at a time. Now, is this a breakthrough? Not even close — it's a years-long standoff. Communications minister Solly Malatsi issued a policy direction back in December pushing ICASA toward that equity-equivalent route, but ICASA maintains the ownership rule is statutory and can't just be waived by a minister's memo. So this is a step forward in the conversation, and Starlink still can't legally operate in South Africa today — remember, it's millions of underserved rural users actually waiting on the outcome, not just SpaceX's balance sheet.

And one more milestone worth a beat: Kepler Communications says its next-generation optical data relay constellation is now live, serving actual commercial and government customers rather than sitting in build mode. Optical relay just means satellites talking to each other with lasers instead of radio, moving data faster and more securely between spacecraft and the ground. SpaceNews posted on X: 'Kepler starts commercial optical data relay service.' Now, this is a single-source announcement — Kepler's own word, with no customer names, no revenue figures, and no count of how many satellites are actually live yet, so we haven't independently confirmed the details behind it. But this is a real operational claim — actual service to actual customers, not a filing sitting in a drawer — and in a race where KSAT and the Space Development Agency are both chasing the same optical-relay market, actually flipping the switch is worth noting on its own terms. That tension — real operational hardware versus promised-but-unbuilt capacity — is exactly the fault line running through our two lead stories today, both about the rush to put AI computers in orbit.

Our first lead story today: an announcement that arrived with a launch date attached, from a source who isn't actually the one launching it. Elon Musk posted on X that SpaceX and Nvidia have designed a space-optimized Vera Rubin NVL72 system — Nvidia's newest AI chip platform — for launch to orbit in the fourth quarter of 2027, with what Musk called significant scale following in 2028. This is tied to SpaceXAI's Starmind satellite program. So what does 'space-optimized' actually mean here? The same day, Nvidia's own newsroom confirmed the partnership is real: SpaceXAI plans to extend its use of Nvidia's accelerated computing into space, with a first-generation Starmind satellite based on that same Vera Rubin NVL72 system. Nvidia's Vera CPUs, the company says, will accelerate its next generation of agentic AI applications — AI systems built to take actions on their own, not just generate answers — by keeping the GPUs, the chips doing the heavy computing, fed with data fast enough to stay busy. Nvidia vice president Ian Buck framed it that way: 'Agentic AI requires a new kind of computing system — one built not only to generate answers, but to take action.' Here's where it gets interesting, though. That Q4 2027 date is Musk's number, not Nvidia's. Nvidia's own press release names no launch date and no capacity figure at all. A trade write-up from StorageReview flagged language buried in Nvidia's own materials describing these deployments as 'not commitments, promises, or legal obligations.' Pricing for the Vera CPUs hasn't even been disclosed. So you've got the world's most famous rocket founder attaching a specific quarter to a satellite whose own chip supplier won't commit to a date at all. And this isn't Nvidia's only move in this space — back in March, the company unveiled its Space-1 Vera Rubin Module and named six early customers: Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud — hold that last name, it's about to matter again in our second story. My read: Nvidia's hedging, backing every horse in a race it hasn't decided how to win yet. But strip away the chip talk for a second and ask the question neither press release answers: how do you actually get rid of heat in space? On the ground, a data center dumps waste heat into air or water. In orbit, there's no atmosphere to carry it away — you're stuck radiating it out as infrared light, a far less efficient process. A full NVL72 rack on Earth needs serious liquid cooling just to survive, and nobody has ever flown anything close to that thermal load in vacuum. That's the actual gating factor here, not the chip design, not the rocket. For you, here's why this matters beyond the space-nerd crowd: if orbital compute genuinely works, it means unlimited solar power with no zoning fight over land or water use — the exact backlash these companies are trying to escape on the ground. If it doesn't work, it's the most capital-intensive science project in the industry with no revenue behind it. So is this real, or is it a land grab dressed up as a hardware program? I think it's genuinely exciting that two of the industry's most capable companies are throwing serious engineering resources at a hard problem — that's how hard problems get solved. But be precise about what actually happened this week: Musk announced a date, Nvidia confirmed a partnership and a chip, and nobody published a thermal design, a mass budget, or a radiator size. Those are the numbers that would tell you whether Q4 2027 is a schedule or a hope.

Our second lead story: the money behind Starcloud, another orbital-compute bet, facing that exact same physics problem — except this time we've got real financial numbers to work with. Starcloud, the orbital data-center startup, has raised a $250 million extension to its March Series A round, according to Payload. That brings its total raised to $450 million, at a valuation of $2.3 billion. The round was led by Manhattan West — which has also backed SpaceX, Relativity Space, and Axiom — with Nvidia and Cisco joining this time as new investors alongside earlier backers Benchmark, EQT, and several venture firms. Now, why would Nvidia write a check into Starcloud while also building its own flagship satellite program with SpaceX? My read: Starcloud is the only company that's actually operated an Nvidia H100 data-center GPU in orbit — it launched one in November — and ran Google's Gemma language model and the open-source NanoGPT up there. That flight data feeds directly back into Nvidia's own chip development for its Space-1 Vera Rubin Module. But here's the part of this raise that actually explains the number. Payload reports the extension is a direct response to what it calls every constellation operator's burgeoning fear — a looming launch shortage. Starcloud has asked the FCC for permission to operate eighty-eight thousand spacecraft in low Earth orbit. Eighty-eight thousand. That would require hundreds if not thousands of individual launches to deploy, even leaning on Starship's capacity, and Starcloud's own flagship spacecraft, Starcloud-3, is slated to fly on that same still-unproven rocket. Johnston himself told TechCrunch: 'We can see what's coming — we're going to need to book an enormous amount of launch.' And he's not alone in that scramble — rideshare, the practice of splitting one rocket's capacity across dozens of smaller satellites to cut cost, is getting harder to book industry-wide, and other companies are hedging the same way: SEOPS is doubling down on its own dedicated rideshare program, Portal Space Systems bought an entire Falcon 9 outright for a 2028 mission, and Cowboy Space has gone so far as to start building its own rocket just to guarantee it can fly its own data-center hardware. When your competitors start building rockets instead of booking one, that tells you something about how tight this market actually feels from the inside. Analysts covering the sector broadly agree near-term revenue stays on the ground — terrestrial compute contracts driving the real numbers while orbital deployment waits on cheaper launch. So here's my read on the eighty-eight-thousand-satellite filing: it belongs on the same shelf as every paper-constellation filing this show has flagged before — a regulatory placeholder, not a construction plan. What I'd want next isn't a bigger filing, it's a signed multi-launch contract that actually reserves capacity, because that's what turns an FCC application into a manifest. And underneath all of it sits the same unsolved thermal problem from our first story — Johnston told TechCrunch his engineers are still working out where to put radiation shielding, how to ruggedize hardware against launch, and how the chip's operating temperature drives the size of the radiators needed to shed heat. Starcloud hopes to fly its Vera Rubin Space-1 hardware in late 2028 — and that's assuming the chip itself, which still has to be built, ships on schedule. One more piece of context worth sitting with: Gallup polled this spring and found roughly 71 percent of Americans oppose a data center being built near them. Orbit, for a company like Starcloud, means escaping that exact fight on the ground while betting billions that the physics cooperate.

So put these two stories side by side, and here's what jumps out at you: Nvidia is playing both sides of the same bet in the same week, funding an independent startup with real flight heritage while also building a flagship in-house satellite with the biggest rocket company on the planet. Neither program has flown rack-scale hardware. Neither has solved how to shed heat without air. And both press releases lean on the word 'plans' where you'd want a launch manifest. So who actually wins this race — the startup with flight heritage, or Nvidia's in-house partnership with SpaceX? What the pairing really reveals is how seriously the industry now takes this idea — you don't hedge across two separate multi-hundred-million-dollar bets on something you think is vaporware. Nvidia clearly believes somebody cracks orbital compute; it just hasn't decided whether that's an independent startup or its own hardware partner. Time for the Hype Check: I'm putting this pairing at a four out of ten on substance. The chips are real, the money is real, the customer interest across six named early users is real — but no thermal design has been published, no rack has flown, and the launch dates on the table, Q4 2027 and late 2028, both belong to companies that don't yet have the rockets to hit them.

If today's episode changed how you think about what 'orbital data center' actually means, that's the whole point of this show — find today's episode wherever you're listening and hit follow, so tomorrow's show finds you automatically. This has been Space Stakes, an AI-voiced podcast, created and built by a real human using today's cutting-edge technology. Nothing you heard on this show is financial advice. I'm Brian Lampert, and I'll catch you all tomorrow — take care!