NVIDIA Sends the AI Factory to Orbit: SpaceXAI Adopts Vera CPUs and a Vera Rubin NVL72 Satellite

NVIDIA has taken the AI factory somewhere it has never gone before: orbit. In a press release issued on August 24, 2026, the chipmaker announced that SpaceXAI will deploy NVIDIA Vera CPUs to accelerate its next generation of agentic AI applications — and, more strikingly, that SpaceXAI’s planned first-generation Starmind AI satellite will be based on an optimized NVIDIA Vera Rubin NVL72 rack-scale system, the same architecture that powers next-generation AI data centers on Earth.

The announcement lands on the same day NVIDIA confirmed that its Groq 3 LPX inference racks entered full production with Nebius as first cloud customer. Together, the two stories sketch the full arc of NVIDIA’s 2026 strategy: own the latency-critical decode layer of inference on the ground, and extend the same accelerated-computing foundation into entirely new environments — including space.

What was actually announced

The deal has three distinct layers, each significant on its own.

First, Vera CPUs for agentic workloads. SpaceXAI will use NVIDIA’s Vera — the company’s first fully custom data center CPU, unveiled on May 31, 2026 — to run the CPU-intensive work that surrounds model inference in agentic AI systems: orchestrating tools, executing code, processing data, and running simulations between model calls. As agents moved from chat demos to production workloads in 2026, this “everything around the GPU” layer became a genuine performance bottleneck. NVIDIA’s pitch is that Vera keeps the GPUs fed and fully utilized while making the agent’s non-model steps run faster.

Second, Vera Rubin as the platform for Grok. SpaceXAI is expanding the AI infrastructure behind Grok on the NVIDIA Vera Rubin platform as it scales toward gigawatts of computing capacity — a phrase that until recently was reserved for the largest terrestrial AI factory projects from OpenAI, Google, and Anthropic partners.

Third, the Starmind satellite. SpaceXAI’s planned first-generation Starmind AI satellite will be built on an optimized Vera Rubin NVL72 system — the rack-scale unit that pairs 72 Rubin GPUs with Vera CPUs over NVLink. NVIDIA’s space-computing materials describe the related Space-1 Vera Rubin module as delivering up to 25x more AI compute per GPU for space-based inference, and the companies say they are working to adapt the architecture to the radically different constraints of orbital computing: power, thermal management, bandwidth, reliability, and physical integration.

Vera: the CPU built for agents

The Vera CPU deserves attention in its own right, because it marks NVIDIA’s entry into custom server silicon against Intel and AMD. Vera features 88 NVIDIA-designed “Olympus” cores with 176 hardware threads via Spatial Multithreading, a monolithic die with 164 MB of unified L3 cache, and an LPDDR5X memory subsystem delivering up to 1.2 TB/s of bandwidth with up to 1.5 TB of SOCAMM capacity per CPU.

NVIDIA claims Vera enables up to 1.8x faster task completion compared with x86 CPUs across workloads including agentic AI, reinforcement learning, and data processing. The design philosophy is unusual for a data center CPU: instead of maximizing core count and throughput per socket, Olympus cores target maximum single-threaded performance and minimal thread-to-thread interference — because an agent’s orchestration loop is full of serial, latency-sensitive steps that stall on slow single-thread performance.

That is precisely the workload profile SpaceXAI is buying for. “Vera gives us the CPU performance and memory bandwidth to run enormous amounts of orchestration, code and data processing while keeping GPUs doing what they do best,” said Mike Nicolls, president of SpaceXAI, in the announcement. “That means higher-performance AI agents and more useful work from every watt of compute.”

Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA, framed the agent thesis directly: “Agentic AI requires a new kind of computing system — one built not only to generate answers, but to take action. Vera gives AI agents the CPU performance to act in real time — executing code, processing data and coordinating complex tasks. SpaceXAI is taking this architecture from massive AI factories to the next frontier of computing in orbit.”

Why put an AI data center in space?

The orbital part of the announcement sounds like science fiction, but it answers a real economic problem. Space-based sensors generate more data than can be economically downlinked to Earth: a single Earth-observation constellation can produce hundreds of terabytes of raw imagery and RF data per day, and the downlink is the bottleneck — expensive, latent, and contested. Moving inference to orbit means processing sensor data where it is collected and sending back answers instead of raw bits.

NVIDIA’s space-computing ecosystem, announced earlier in August 2026, already includes Axiom Space, Cowboy Space Corporation (formerly Aetherflux), Kepler Communications, Planet Labs, Sophia Space, and Starcloud — companies using everything from Jetson Orin modules (real-time on-board vision and navigation) to IGX Thor (ruggedized inference) to the Space-1 Vera Rubin module (frontier models in orbit). Planet Labs is building a GPU-native AI engine to turn satellite imagery into real-time planetary intelligence; Kepler is processing data in-orbit across its optical constellation; Firefly Aerospace is integrating Jetson on its Elytra spacecraft for lunar imagery.

The momentum behind orbital compute got a fresh validation just days ago, when Starcloud — the startup building GPU-equipped satellites with giant solar arrays — closed a $250 million funding extension reported by TechCrunch on August 21, reaching a valuation above $1.1 billion. Starcloud’s long-term thesis is that solar power in orbit is effectively uninterrupted, cooling radiates to space, and electricity costs could be roughly 90% lower than terrestrial data centers at scale.

SpaceXAI’s Starmind pushes the same idea further: not just edge inference for one sensor payload, but a general AI compute node in orbit based on the same NVL72 rack architecture used in Earth-side AI factories — one software ecosystem, two very different environments.

The strategic read

Three threads tie this announcement into the broader AI infrastructure race.

One architecture, everywhere. NVIDIA’s explicit goal is a common computing foundation across environments: Vera CPUs for agent orchestration, Vera Rubin for the Grok training and inference buildout, and the same architecture adapted for orbit. For customers, software compatibility between ground and space removes the biggest historical barrier to orbital compute — nobody wanted to port and requalify a parallel software stack for a handful of satellites.

The CPU war heats up. With Vera landing flagship customers like SpaceXAI (and, per The Information’s briefing the same day, Nebius as another new Vera customer), NVIDIA is now directly attacking the x86 strongholds of Intel and AMD inside the AI factory. Agents changed the CPU workload profile, and NVIDIA is exploiting the shift with silicon purpose-built for it.

Compute follows constraints. Terrestrial AI buildouts are colliding with grid, water, and permitting limits — the same week saw Nvidia warning hyperscalers of 15%+ server price hikes driven by DRAM costs, and Cramer declaring the AI trade “currently broken” on data-center-slowdown fears. Orbital compute is still a niche bet, but the Starmind satellite and the Starcloud raise suggest serious money is now hedging the constraint that matters most: energy.

What to watch

The press release is a plan, not a launched product. Key unknowns: the launch timeline for the first Starmind satellite, how a liquid-cooled NVL72 rack survives launch vibration and operates under orbital power and thermal cycles, what “optimized” means in radiation-tolerance terms, and whether the economics of orbital inference genuinely beat downlink-and-process on Earth. NVIDIA’s Space-1 materials say the module is designed for size, weight, and power-constrained environments, but independent validation will only come with the first missions.

What is no longer debatable is the direction of the top of the stack. The AI factory is becoming a design pattern — Vera CPUs, Rubin GPUs, NVLink, Spectrum-X — that can be stamped into hyperscale campuses on Earth, neocloud token factories, and now satellites. When the same rack trains Grok on the ground and runs agents in orbit, “data center” stops being a place.

Sources

• NVIDIA Newsroom — SpaceXAI Adopts NVIDIA Vera CPU to Accelerate Agentic AI at Massive Scale
• NVIDIA — Space Computing: On-Orbit AI & Accelerated Computing
• NVIDIA — Vera CPU product page
• The Information — Nvidia Announces New Customers For Vera CPU, Groq LPX Racks
• The Tech Buzz — SpaceXAI Taps NVIDIA’s Vera CPU for Agentic AI Scale