NVIDIA Chips Enter Orbit: SpaceX's Orbital AI Economy Blueprint Emerges
Nashnova编辑部
Nvidia announced that SpaceXAI will deploy its Vera CPU aboard Starmind satellites, extending Nvidia's compute stack into orbit for the first time — giving SpaceX a near-term path to orbital AI without waiting for its own silicon.
What makes this CPU special?
The Vera CPU packs 88 cores with memory bandwidth of 1.2 TB/s.
On agentic AI, reinforcement learning, and data-processing tasks, Nvidia claims it runs up to 1.8× faster than conventional x86 CPUs.
This means → it is not a general-purpose processor but a purpose-built "orbital brain" for AI inference and data-heavy workloads.
Why isn't SpaceX waiting for its own chips?
The first-generation Starmind satellites will carry Nvidia's Vera Rubin NVL72 system directly into orbit.
In plain terms = SpaceX chose "fly Nvidia's off-the-shelf solution now, figure out in-house chips later" — trading time for capability.
On the ground, SpaceX is simultaneously expanding the Vera Rubin architecture inside its Grok infrastructure, building compute on both ends.
What roles do Starlink and Starship play?
Starlink has already built a constellation of thousands of satellites with inter-satellite laser links, providing the networking backbone.
SpaceX plans to convert its completed V3 satellite platform into AI-compute satellites — swapping communication hardware for AI accelerators — with a preliminary design target of roughly 100 kilowatts of compute per metric ton.
Starship V3 aims to deliver 100 metric tons to orbit per launch in fully reusable configuration, ultimately deploying up to 100 GW of orbital AI compute per year.
This means → Starlink weaves the network; Starship hauls the hardware. Both must work together to move compute from the ground to space at scale.
What timeline has Musk laid out?
Musk offered a mid-term roadmap explicitly labeled a "best estimate, not a commitment."
The target: roughly 1 GW of annualized orbital-AI deployment by end of 2027, scaling in stages toward 10 GW and eventually 100 GW.
In January 2026, SpaceX filed with the U.S. Federal Communications Commission for authorization of an orbital data-center system of up to one million satellites — a regulatory ceiling, not a deployment pledge.
This reflects an ambition far beyond any existing on-orbit infrastructure, but the gap between "filing" and "deploying" remains enormous.
What is the Terafab plan?
SpaceX and Tesla jointly announced Terafab in March 2026; Intel joined in April.
The facility aims to integrate lithography-mask design, logic fabrication, and memory manufacturing — pushing vertical integration down to the silicon level.
In plain terms = Terafab means SpaceX doesn't just want to build rockets and satellites — it eventually wants to fabricate its own chips, closing the loop from orbit to fab floor.
Can this blueprint actually deliver?
The Nvidia partnership solves the near-term transition problem: get usable compute into orbit now.
Terafab addresses the long-term autonomy goal: eliminate dependence on external chip suppliers.
Whether the two tracks can advance in parallel — and whether Starmind's on-orbit performance validates the commercial case — are the critical proof points for this entire blueprint.
市场有风险，内容仅供研究参考，不构成投资建议。