Space is often imagined as cold. For an AI computer, it may be dangerously hot. That paradox sits at the centre of Elon Musk’s latest technological wager. SpaceX and Nvidia have designed a space-optimised version of Nvidia’s Vera Rubin NVL72 AI system for launch into orbit in the fourth quarter of 2027, Musk said on X. He expects a “significant scale-up” in 2028.
Nvidia has confirmed that SpaceX’s first-generation Starmind AI1 satellite will carry a computing payload based on Vera Rubin NVL72, taking the architecture used by terrestrial AI factories into orbit. It is an audacious proposition. Instead of transmitting every photograph, sensor reading or scientific measurement back to Earth for processing, a satellite could analyse the information itself, decide what matters and send down the answer. In a more ambitious future, constellations of satellites could host parts of the enormous computing infrastructure required to run artificial intelligence.
The Vanguard #
21 Aug 2026 - Vol 05 | Issue 34
BJP Rearmed for 2029
Read Now The Vanguard Space would provide abundant solar energy, room to expand and freedom from terrestrial fights over land, electricity and water. It would also remove the power grid, cooling towers, technicians and repair crews that allow a data centre to survive. The hardware may be ready to fly in 2027. The harder question is whether the economics and physics will permit orbital AI to scale.
What exactly are SpaceX and Nvidia launching? #
The planned satellite is called Starmind AI1. Its computing payload will be derived from Nvidia’s Vera Rubin NVL72, a rack-scale AI system built for training models, running inference and powering AI agents. The terrestrial NVL72 combines 72 Rubin graphics processing units with 36 Vera central processing units, along with high-speed networking and data-processing components. It is not simply a powerful computer with more chips. The architecture is designed to make those chips behave like one giant computing system.
The GPUs handle the immensely parallel mathematical operations that underpin modern AI. The CPUs organise the work around them: retrieving and preparing data, executing code, managing software tools, running simulations and deciding which task should go where. Nvidia calls Vera its first CPU purpose-built for agentic AI. Unlike a chatbot that merely returns text, an AI agent may search databases, write and execute code, call other applications, compare results and repeatedly adjust its approach before completing a task. All those steps can leave expensive GPUs waiting while conventional CPUs struggle with orchestration. Vera is meant to reduce that idle time.
According to Nvidia, the processor has 88 custom-designed Olympus cores capable of running 176 threads. Its LPDDR5X memory system can deliver bandwidth of up to 1.2 terabytes per second. Nvidia claims Vera can complete certain agentic, data-processing and reinforcement-learning workloads up to 1.8 times faster than conventional x86 processors. SpaceXAI President Mike Nicolls said Vera would allow the company to run enormous volumes of code, orchestration and data processing while keeping its GPUs occupied.
But the phrase “space-optimised NVL72” matters. SpaceX cannot bolt an ordinary data-centre rack inside a satellite and send it up on a Falcon 9. Nvidia describes its terrestrial Vera CPU rack as a 48U system that is entirely liquid-cooled. An orbital version must be redesigned around weight, power generation, radiation protection, heat rejection, launch vibration and the impossibility of routine human repair. Starmind will carry the architecture and computing logic of NVL72, not necessarily an untouched terrestrial rack.
Why put an AI supercomputer in orbit? #
The most convincing early use is not replacing Earth’s data centres. It is processing space-generated information before sending it to Earth. Modern satellites generate enormous quantities of data. Earth-observation spacecraft capture high-resolution images. Weather satellites study clouds, storms and atmospheric conditions. Defence satellites monitor movement. Scientific missions collect measurements that may need to be filtered, classified or compared.
Today, much of that raw information is transmitted to ground stations for processing. That creates a bottleneck. A satellite may gather data faster than it can send it home. Onboard AI could examine an image immediately, discard unusable frames obscured by clouds, detect wildfires, identify changes in crops or infrastructure and transmit only the useful result. Instead of sending thousands of images containing nothing unusual, the satellite might send one alert: a fire has appeared at these coordinates.
Nvidia says its space-computing platforms are intended for orbital analytics, autonomous scientific discovery and processing satellite data closer to its source. The company claims its Vera Rubin Space-1 module can offer up to 25 times more AI computing per GPU for space-based inference and orbital data centres, although that remains a company performance claim. The attraction for SpaceX is especially clear. It already operates the Starlink network and possesses the rockets, satellite-manufacturing capacity, orbital communications infrastructure and launch frequency required to attempt a constellation. A company that had to buy launches from somebody else would begin with an enormous cost disadvantage. SpaceX can manufacture the satellite, launch it, connect it through laser links and potentially replace it when newer hardware arrives. That vertical integration does not make orbital computing easy. It makes SpaceX one of the few companies for which the idea is not immediately absurd.
Isn’t space naturally perfect for cooling computers? #
No. This may be the most misleading part of the orbital data-centre dream. Space is cold in the sense that there is almost no matter around a spacecraft. That emptiness is precisely the problem.
On Earth, computers can transfer heat into moving air or liquid. Fans push air across components. Cooling systems carry heat away in water or refrigerant. Ultimately, the heat is released into the surrounding environment. A vacuum has no air through which heat can be carried. A spacecraft can lose heat mainly by radiating it away as infrared energy. That requires large radiator surfaces. The more electricity an AI system consumes, the more waste heat it produces and the larger its heat-rejection system must become.
A recent technical analysis of a hypothetical one-megawatt orbital data centre estimated that it could require approximately 5,640 square metres of solar panels and 2,500 square metres of radiators. The paper, which has not yet passed through the scrutiny associated with a mature commercial deployment, illustrates the scale of the problem rather than prescribing SpaceX’s design. The radiator could become one of the largest and most vulnerable structures on the satellite. It would add mass, complicate deployment and have to survive micrometeoroids, thermal cycling and debris.
Computer workloads also fluctuate. An AI satellite may switch rapidly between intense computation and lighter activity, producing uneven thermal loads. If heat cannot escape quickly enough, processors must slow themselves down to avoid damage. Space offers sunlight. It does not offer free cooling.
Where will Starmind get enough electricity? #
From solar panels, almost certainly. In an appropriate low-Earth orbit, a satellite can receive sunlight for much of its journey. A dawn-dusk sun-synchronous orbit can keep a spacecraft illuminated for unusually long periods, reducing its dependence on batteries. That is one of the strongest arguments for orbital computing. A space-based system does not need to acquire land for a solar farm, connect to a congested electricity grid or consume water for evaporative cooling. Yet sunlight alone does not solve the engineering problem. Solar panels produce limited power for their area and weight. They degrade under radiation. Batteries are needed when the spacecraft passes through darkness or requires power beyond immediate solar output. Every panel, cable, battery and radiator must be launched. AI factories on Earth are increasingly measured in hundreds of megawatts and even gigawatts. A first-generation satellite will operate at a tiny fraction of that scale. Starmind AI1 is, therefore, best understood as an engineering pathfinder. It can test whether high-density AI hardware can survive launch, manage its heat, withstand radiation and deliver useful computation over an extended orbital mission. Calling it an orbital AI factory describes the destination. The first satellite is closer to the opening experiment.
Can ordinary Nvidia chips survive space radiation? #
Not without protection and careful system design. Most commercial data-centre processors are not radiation-hardened. On Earth, the atmosphere and magnetic field protect electronics from much of the high-energy radiation present in space.
In orbit, energetic particles can flip bits in memory, corrupt calculations, crash systems or permanently damage circuits. As transistors become smaller and chips become denser, controlling these errors can grow more difficult.
Spacecraft traditionally use radiation-hardened processors, but those chips are often generations behind the fastest commercial hardware. AI demands make that performance gap increasingly painful. The alternative is to use powerful commercial processors while designing resilience around them. That may involve shielding, error-correcting memory, redundant computers, repeated calculations, fault-detection software and the ability to isolate and restart damaged components.
SpaceX may also accept shorter hardware lives. A satellite built to operate for three or five years does not need the same durability as a spacecraft intended to survive for decades. If Starship eventually reduces launch costs and SpaceX manufactures Starmind satellites at scale, the company could regularly replace old or degraded computers with newer ones. That sounds wasteful, but AI hardware on Earth also becomes obsolete rapidly. The economics may depend less on making every satellite immortal and more on making replacement cheap and routine.
How will the satellite move all that data? #
This could become a bigger constraint than the processors themselves. Inside an NVL72 system, CPUs and GPUs exchange enormous volumes of data through Nvidia’s high-speed interconnects. The system’s internal traffic can operate at terabyte-per-second scales.
Links between space and Earth are far narrower. Researchers studying orbital data centres describe a fundamental mismatch between the vast amount of data exchanged inside AI systems and the limited capacity available for transmitting it to and from space.
That makes some workloads much better suited to orbit than others. If data originates aboard a satellite and can be reduced to a small answer, orbital AI makes sense. Starmind could analyse imagery, sensor streams or network conditions and send down a conclusion rather than the entire dataset. If enormous training datasets must constantly be uploaded from Earth and equally large results returned, the communications bill becomes much harder to close.
SpaceX possesses one possible advantage: Starlink’s optical inter-satellite links. Starmind could potentially move information between satellites through lasers before routing selected outputs to ground stations. But even a global communications network does not erase the bandwidth gap. It encourages a different computing model: send intelligence, not raw information.
Will Starmind train giant AI models in space? #
Probably not at first. Training a frontier model requires thousands of accelerators exchanging information continuously at extraordinarily high speeds. Even small interruptions can reduce the efficiency of the whole cluster.
Creating that tightly connected computing fabric across multiple satellites would be far harder than connecting racks inside one building. The spacecraft would move relative to one another, links could be interrupted and every extra communication step would introduce delay. An early Starmind satellite is more likely to focus on inference, data processing and autonomous decision-making. It could run an already trained model, analyse information and take action without waiting for instructions from Earth.
That may include managing satellite operations, optimising communications, identifying anomalies, processing imagery or allowing AI agents to execute tasks in orbit. Training could become possible later if many large computing satellites are connected through extremely fast optical links. But the leap from one orbital AI payload to a distributed space supercomputer is enormous. Musk’s reference to “significant scale” in 2028 should therefore be treated as a target, not a settled deployment schedule.
Does an orbital data centre make financial sense? #
At present, the case remains difficult.
Every kilogram of computing equipment requires a spacecraft around it: power generation, shielding, structure, communications, batteries, thermal control and propulsion. All of it must survive launch and then operate without technicians. The recent orbital-data-centre analysis estimated that a one-megawatt system could weigh roughly 34 to 59 kilograms for every kilowatt of usable computing power once major spacecraft systems were included.
At around 40 kilograms per kilowatt, the study found that the combined launch and spacecraft-manufacturing cost would need to fall to between $250 and $1,000 per kilogram to compete with a broad range of terrestrial infrastructure costs. That was substantially below publicly available dedicated Falcon 9 launch-price benchmarks even before the cost of building the spacecraft was included. Those calculations depend heavily on assumptions about hardware life, launch prices, solar efficiency, utilisation and the value of the computing service. SpaceX may argue that reusable Starship launches and mass-produced satellites can change them dramatically.
The sceptics are not convinced. Amazon Web Services chief Matt Garman described orbital data centres as “pretty far” from economic reality, citing launch capacity and payload costs, according to Reuters. Both propositions can be true. Orbital data centres may be commercially unattractive with today’s launch economics and still become plausible if launch costs collapse. SpaceX is effectively betting that the rocket curve and the AI-demand curve will cross.
Why does Nvidia want to go to space? #
For Nvidia, Starmind extends its central business strategy. The company does not merely want to sell GPUs. It wants its architecture, networking, software and programming ecosystem to become the standard platform on which AI runs. If the same software stack operates in a terrestrial data centre, at a ground station and aboard a satellite, developers do not need to rebuild applications for every environment. That makes Nvidia’s technology harder to replace. Its official space-computing platform already lists partnerships with satellite operators and space-infrastructure companies. SpaceX is the most dramatic addition because it can combine computing with launch and connectivity. For Nvidia, “one architecture, everywhere” means Earth, the edge and now orbit. For SpaceX, the Nvidia partnership avoids having to invent an advanced AI-computing ecosystem from scratch. Each company supplies what the other lacks.
What would success look like in 2027? #
Success will not mean that Earth’s data centres suddenly become obsolete.
A credible first mission would demonstrate that Starmind can launch safely, deploy its power and cooling systems, run demanding AI workloads, withstand radiation errors and communicate useful results to Earth over a meaningful period. It would also have to show that the computer delivers enough useful work to justify the mass and money placed into orbit.
The most promising near-term market may be “space edge computing”: analysing data where it is produced instead of attempting to recreate a terrestrial hyperscale data centre above the atmosphere. If Starmind can do that reliably, SpaceX could add computing payloads to future satellites, distribute workloads across a constellation and gradually test more ambitious applications.
If it cannot control heat, radiation damage or communication bottlenecks, the project may remain an impressive demonstration attached to an uneconomic machine. The announcement is, therefore, significant without yet being revolutionary. Nvidia has built computing architecture powerful enough for SpaceX to consider taking it off the planet. SpaceX has made launches frequent and satellite production industrial enough for the idea to be tested seriously. But orbit does not repeal the rules of infrastructure. It intensifies them. On Earth, a failed server summons a technician. In space, it becomes a very expensive shooting star.
*(With inputs from ANI)*