RC / LOADING
ORBITAL COMPUTE · AI INFRASTRUCTURE
Nvidia's relationship with Elon Musk's AI operation began inside a Memphis data center. SpaceX now plans to extend that computing ecosystem into orbit, flying a space-optimized Nvidia system on its first Starmind satellites.
SpaceX wants to put a version of one of the world's most powerful AI computing systems into orbit. On August 24, Elon Musk said SpaceX and Nvidia had jointly designed a space-optimized Vera Rubin NVL72 system. The target is a first orbital launch in the fourth quarter of 2027, with significant scale in 2028. Nvidia confirmed the plan the same day. It announced that the first-generation Starmind AI satellite will be based on an optimized Vera Rubin NVL72, and that SpaceXAI will adopt its new Vera CPU for agentic AI workloads.
This is not a new supplier relationship. Nvidia hardware already forms the foundation of the AI infrastructure that came into SpaceX through xAI. The relationship now spans chips, networking, software, joint engineering and an equity stake recently valued at roughly $21 billion. The orbital announcement extends a computing ecosystem the two companies have been building together on Earth since 2024.
01 · THE HISTORY
The relationship began in Memphis. In 2024, xAI built the Colossus supercomputer using roughly 100,000 Nvidia Hopper GPUs, connected through Nvidia's Spectrum-X Ethernet platform and BlueField networking hardware. Nvidia said the system and its supporting facility were brought online in about 122 days, with training workloads starting 19 days after the first racks arrived. Colossus became the training infrastructure behind Grok. It also made Musk's AI operation one of Nvidia's largest and fastest-moving customers.
Capital followed the hardware. In December 2024, Nvidia participated as a strategic investor in xAI's $6 billion Series C. In January 2026 it invested again in xAI's Series E, a round upsized to $20 billion at a valuation of about $230 billion. By early 2026 the two companies were bound by both technology and money.
Then the corporate structure changed. On February 2, 2026, SpaceX acquired xAI in an all-stock merger that valued the AI company at $250 billion. xAI became a wholly owned subsidiary and was later folded into SpaceX as its AI division, SpaceXAI. Nvidia's technology and investment relationship with xAI became, in one transaction, a relationship with the broader SpaceX organization.
02 · THE COMMITMENT
The strategic depth of that relationship became explicit on August 4, 2026, on SpaceX's first earnings call as a public company. Musk told shareholders that SpaceX had decided to build its AI infrastructure exclusively on Nvidia. He called Vera Rubin "the best AI computer" and cited close cooperation between the two companies on many levels. He said SpaceX expects to end 2026 with more than 2 gigawatts of compute online. Cumulative capacity at the end of 2027, he said, may be closer to 10 gigawatts than 5 gigawatts. Both figures are company targets rather than installed capacity; SpaceX ended the second quarter with about 1.4 gigawatts of nameplate compute.
The disclosure marked a change in what SpaceX is. A company known for launch vehicles and communications satellites is now also a hyperscale compute operator, buying AI hardware at a scale comparable to the largest cloud providers. Nvidia's own results reflect the shift. On its August 26 earnings call, CFO Colette Kress named SpaceXAI among the lead partners already receiving Vera CPU shipments, alongside Oracle Cloud Infrastructure and, starting this quarter, AWS.
FIG. 1 · From a Memphis training cluster to a planned orbital deployment in under four years. Solid markers show completed corporate events; the dashed segment and open marker are company targets. RC SPACE INTELLIGENCE
03 · THE LATEST STEP
The August 24 announcement had two parts. The first is terrestrial. SpaceXAI will deploy Nvidia's Vera CPU, the first Nvidia processor built specifically for AI agents, across its next generation of agentic workloads. The division of labor is straightforward. GPUs perform the heavy model computation. CPUs handle the surrounding work: orchestrating tools, executing code, processing data, running simulations and keeping the GPUs supplied with work between model calls. SpaceXAI President Mike Nicolls said Vera lets the company run that orchestration at scale while getting "more useful work from every watt of compute."
The second part is the orbital one. Nvidia said the first-generation Starmind satellite will be based on an optimized Vera Rubin NVL72 system, and that the two companies are adapting the architecture for orbital computing. Musk added the schedule in his post on X: first launch targeted for the fourth quarter of 2027, with significant scale in 2028. Those dates are company targets, disclosed by the companies rather than assessed independently.
The design intent matters more than the dates. SpaceX is trying to avoid building an entirely separate computing stack for space. Nvidia says adapting NVL72 for orbit preserves a common architecture and software ecosystem across Earth-based and space-based infrastructure. The platform family SpaceXAI is deploying behind Grok in its terrestrial AI factories would run, in modified form, on satellites.
FIG. 2 · The upper path operates today inside SpaceXAI's terrestrial facilities. The lower path is the planned Starmind architecture, shown dashed because no orbital hardware has yet flown. The common element is the Nvidia compute platform in the middle of each chain. RC SPACE INTELLIGENCE
04 · THE HARDWARE
Understanding the announcement requires understanding what Vera Rubin NVL72 is. On Earth, it is Nvidia's flagship rack-scale system. It unifies 72 Rubin GPUs and 36 Vera CPUs through the NVLink 6 fabric, alongside ConnectX-9 SuperNICs, BlueField-4 DPUs and about 20.7 terabytes of HBM4 GPU memory. The design goal is for the rack to behave less like 72 separate GPUs and more like one enormous AI computer, with very high bandwidth between every chip. Nvidia said in late August that the platform is now in full production.
SpaceX is not launching an ordinary data-center rack unchanged. Nvidia is explicit that the Starmind version is an optimized Vera Rubin NVL72, and Musk has said the joint design work covers a space-specific variant. Neither company has published a full specification for the orbital system. What is public is the set of constraints the redesign has to satisfy.
Mass is the first. Every kilogram must ride a rocket, so the enclosure, facility hardware and cooling plant that surround a terrestrial rack are candidates for removal or redesign. Power is the second. In orbit there is no grid; Starmind satellites draw on large solar arrays sized to the compute payload. Cooling is the third, and it is the one most often misunderstood. There is no air or cooling water in orbit, and vacuum is an insulator rather than a heat sink. Waste heat must be pumped to radiators and rejected as infrared radiation, which is why SpaceX's design carries large deployable radiator panels. Radiation, launch vibration and physical integration round out the list. Commercial AI silicon must tolerate an orbital radiation environment, survive the loads of a Starship launch, and function as part of a spacecraft rather than a rack bolted to a data-center floor.
FIG. 3 · The compute architecture is intended to carry over; almost everything around it changes. This is what "space optimized" means in practice, and it is the engineering scope of the joint SpaceX-Nvidia design work. RC SPACE INTELLIGENCE
05 · STARMIND
Starmind is the spacecraft built around those constraints. The first-generation satellite, designated AI1, is designed around a compute payload of up to about 250 kW at peak and roughly 175 kW on average, according to SpaceX's published specifications. Both companies say the satellite will be based on an optimized Vera Rubin NVL72, and on the August 4 call Musk described the spacecraft as essentially an NVL72 computer adapted for orbit. The published design stands about 30 meters tall with a 75-meter solar wingspan and generates around 210 kW from its arrays. Waste heat leaves through roughly 160 square meters of deployable liquid radiators fed by redundant pumped loops. The satellites are intended to operate in sun-synchronous orbit and connect through high-bandwidth laser links into the Starlink network. SpaceX plans to mass-produce them at its Gigasat Factory in Bastrop, Texas, with prototype flights targeted for early 2027.
SOLAR POWER → NVIDIA-BASED LOCAL COMPUTE → AI WORKLOAD → LASER LINKS → STARLINK NETWORK → RESULTS RETURNED TO EARTH
The distinction from Starlink is functional. Starlink satellites relay data. Starmind satellites are intended to compute in orbit, running AI workloads locally and sending finished results down. SpaceX's case for doing this rests on assets it already holds. Those include abundant solar energy in orbit, freedom from terrestrial grid interconnection queues, Starship's projected mass-to-orbit, Starlink's optical mesh, and an industrial-scale satellite production system. Each element exists today in some form. Assembling them into an operating orbital compute service is the work of the next two years.
06 · THE EXCHANGE
For SpaceX, the case for Nvidia is mostly about time. Vera Rubin gives SpaceX a leading compute platform without first building a chip ecosystem of its own. CUDA and Nvidia's surrounding software stack matter as much as the silicon. Workloads written for terrestrial SpaceXAI clusters can, in principle, migrate to orbital hardware running the same software and compute stack. Nvidia can also supply at the scale SpaceX's gigawatt targets require. Developing an internal processor and its software ecosystem first could substantially delay orbital deployment; an established platform lets the satellite program start now.
Nvidia's side of the exchange is just as concrete. SpaceXAI is scaling toward gigawatts of compute, which makes it one of Nvidia's largest customers in its own right. Beyond that, SpaceX supplies things no other customer can. It owns the launch system needed to move compute hardware to orbit in quantity, a satellite factory, and an existing laser network to connect distributed orbital nodes. If orbital computing develops into a real market, Starmind would make Nvidia's architecture the incumbent in an entirely new physical destination for accelerators. That is strategic positioning, not booked revenue, and Nvidia has published no orbital revenue forecast.
07 · THE SHAREHOLDING
The relationship is not limited to selling hardware. Nvidia's xAI investments converted into SpaceX stock at a fixed exchange ratio in the February merger. In a 13F filing on August 14, Nvidia disclosed holding about 122.8 million SpaceX Class A shares, worth roughly $21 billion at the end of June. Reporting by the Financial Times and CNBC placed Nvidia among SpaceX's largest outside shareholders. Within Nvidia's own disclosed equity book, only its Intel position was larger at the filing's end-June values. SpaceX's share price has fallen since its June IPO, reducing the stake's market value, but the structural point stands. Nvidia is simultaneously a technology supplier, a joint engineering partner and a significant shareholder in its customer. The three roles align both companies around the same outcome: the successful buildout of SpaceX's AI infrastructure on Nvidia hardware, on Earth and in orbit.
08 · EXCLUSIVE NOW, MODULAR LONG TERM
One nuance is worth stating precisely. Musk said in August that SpaceX will build exclusively on Nvidia. SpaceX's own Starmind materials, meanwhile, describe the satellite's compute architecture as modular and able to accept chips from different vendors. The company has also said it intends to produce advanced AI chips through Terafab, a semiconductor project planned with Tesla. These statements fit together without contradiction. Nvidia is the architecture SpaceX has chosen for the current buildout and for first-generation Starmind. The satellite bus is designed so future compute modules are swappable. And SpaceX, like every hyperscaler, is investing in longer-term internal silicon capacity. This is an evolving technology strategy rather than a signal about the partnership's future, and nothing disclosed to date suggests SpaceX plans to move away from Nvidia.
09 · WHY THIS MATTERS FOR SPACE
The broader significance extends past the two companies. If orbital computing scales, AI infrastructure becomes a new source of high-mass launch demand, and one well matched to Starship's payload class. Satellite design shifts with it. A spacecraft optimized around compute, power generation and heat rejection is a different machine from a communications satellite, and solar arrays, radiators and power electronics become the defining spacecraft technologies. Chipmakers gain a reason to treat orbit as a deployment environment worth qualifying for, the way they treat automotive or industrial environments today. And optical networking becomes load-bearing infrastructure, because orbital compute is only useful if results move quickly between satellites and Earth.
The deepest change is convergence. The supply chains for compute hardware, cooling, power and satellite integration begin to overlap with the terrestrial AI data-center supply chain. The boundary between the space industry and the AI infrastructure industry becomes less distinct, and companies on each side acquire reasons to understand the other.
CONCLUSION
Nvidia helped build the AI infrastructure behind xAI on Earth, starting with a hundred thousand GPUs in Memphis. After xAI became part of SpaceX, that relationship widened into capital, exclusivity and joint engineering. The August 24 announcement extends it to a new destination. What flies in late 2027, if the target holds, will be a heavily modified machine wrapped in solar panels and radiators. But the platform inside it will share its architecture and software with the Vera Rubin systems SpaceXAI is deploying on the ground. That continuity, more than any single specification, is what the two companies are betting on: that AI hardware can become a category of space infrastructure without being reinvented from scratch.