SpaceX’s $10 Billion Bet on Nvidia: The Computing Power War Spreads from Ground to Space

On August 5, 2026, SpaceX dropped a bombshell during its first earnings call since going public: the company will “exclusively” adopt Nvidia GPUs to meet all of its artificial intelligence computing needs. SpaceX CEO Elon Musk stated bluntly on the call that Nvidia’s latest Vera Rubin architecture is “the best AI computer available today.” The news sent Nvidia’s stock up 3.4% that same day, marking its fifth consecutive trading session in the green, with a cumulative gain of 15.4%.

This is far more than a simple customer endorsement—it is a strategic escalation of the computing power landscape. When a company whose core business is rocket launches and satellite internet suddenly becomes the most prominent cheerleader for an AI chip giant, the industry’s underlying logic is being rewritten.

The 10 GW Computing Ambition: The Staggering Numbers

According to SpaceX’s published compute expansion plan, the company expects its online computing capacity to exceed 2 gigawatts (GW) by the end of 2026, with a further goal of ramping total capacity to 10 GW by the end of 2027. What does 10 GW mean? It could translate into approximately $400 billion to $500 billion in revenue for Nvidia (roughly 2.7 trillion to 3.4 trillion yuan). For context, Nvidia’s projected revenue for the full fiscal year 2026 is about $130 billion—meaning a single customer like SpaceX could potentially generate orders equivalent to more than three years of Nvidia’s total revenue.

Wolfe Research analyst Chris Caso ran a more detailed calculation: based on an estimate that every 1 GW of new chip deployment could bring about $35 billion in revenue for Nvidia, SpaceX alone could exceed the $160 billion incremental growth that Wall Street currently projects for Nvidia’s data center business in 2027. Caso noted in his report, “At a time when the market is worried about whether AI spending can be sustained, Musk’s comments provide evidence that the growth momentum of AI spending may well continue.”

It is worth noting that SpaceX’s computing demand is not a one-time explosion but a phased rollout. The leap from 2 GW to 10 GW means that in just one year—from the end of 2026 to the end of 2027—compute capacity would need to quintuple, a rate of exponential expansion rarely seen in industrial history. Achieving this goal requires stable chip supply, adequate power infrastructure, and efficient system integration—exactly the kind of full‑stack solution that Nvidia can deliver.

From Ground to Orbit: The Ambition of Space‑Based Data Centers

SpaceX’s computing expansion is not limited to terrestrial facilities. The company is actively extending its AI compute capabilities to low‑Earth orbit. Nvidia has already introduced the Space‑1 Vera Rubin module, designed specifically for orbital environments. This module integrates four Rubin GPUs with two Vera CPUs and delivers up to 25 times the AI inference performance of the previous‑generation H100 GPU. Musk also revealed that SpaceX’s Starmind AI satellite “will essentially be a modified Vera Rubin NVL72 computer,” scheduled for launch next year. The Vera Rubin NVL72 is a rack‑scale system comprising 72 Rubin GPUs and 36 Vera CPUs.

The strategic move to place computing infrastructure in space aims to overcome the physical bottlenecks of terrestrial data centers—power supply, land acquisition, and cooling. Ground‑based data centers are constrained by grid capacity, floor space, and water resources, whereas space offers virtually unlimited solar energy and an extremely cold environment, naturally suited for high‑density computing. SpaceX has previously announced a partnership with AI company Anthropic to jointly build a gigawatt‑scale orbital AI data center, a collaboration seen as a prelude to moving large‑model training into space.

However, this vision faces enormous challenges. Industry analysts point out that beyond the supply of AI chips themselves, the sheer logistics of launching tens of thousands of GPUs into orbit using commercial space transportation is a monumental engineering task. Although SpaceX’s Starship has significantly reduced per‑launch costs, repeatedly delivering thousands of tons of payload to low‑Earth orbit still requires high‑frequency, low‑cost launch capabilities. Moreover, radiation exposure in space, thermal management of satellites, and orbital maintenance all demand novel engineering solutions. Still, Musk has built a reputation for turning the seemingly impossible into reality, and space‑based compute may simply be a matter of time.

The Computing Race Enters an Era of Massive Expansion: Data from Across the Industry

SpaceX’s blockbuster move is by no means an isolated case. Over the past week, multiple signals from the computing industry paint a picture of broad‑based expansion that is deeper and wider than market expectations.

First, global cloud providers are collectively doubling down on AI infrastructure. Based on the latest plans of nine major cloud vendors, TrendForce raised its 2026 global AI server shipment growth forecast to nearly 31% on August 3. The combined capital expenditure of these nine cloud providers in 2026 is expected to exceed $886.7 billion, representing year‑over‑year growth of close to 90%. Among them, the four U.S. tech giants—Microsoft, Google, Amazon, and Meta—account for more than 70% of total cloud investments, while Chinese players such as Alibaba, Tencent, and Huawei are rapidly catching up. At such scale, AI computing has clearly shifted from “exploratory spending” to “strategic infrastructure.”

Second, compute rental prices continue to rise, reflecting persistent supply‑demand tightness. According to third‑party tracking platforms, rental rates for Nvidia H100 GPUs have increased by about 50% since the beginning of the year, while the newer B200 GPUs—still in the early stages of production ramp‑up—have seen similarly climbing prices, with some cloud providers reporting waiting times of several weeks for B200 instances. This “hard‑to‑get” situation precisely illustrates the “Jevons paradox” recently highlighted by Morgan Stanley: as open‑weight models (such as Meta’s Llama family) lower the cost of AI usage, making individual inferences cheaper, companies tend to deploy AI across more edge cases and long‑tail tasks, ultimately driving up total compute demand rather than suppressing it. In other words, efficiency gains fuel consumption, and the compute market is far from hitting its ceiling.

Finally, at the manufacturing end, capacity constraints remain acute. Nvidia’s GPU orders have pushed TSMC’s advanced packaging lines to their limits. TSMC has decided to outsource more CoWoS packaging capacity to OSATs like ASE Group, while simultaneously accelerating the expansion of its own packaging facilities. Even so, the industry expects that supply of high‑end GPUs will lag behind demand by roughly 10%–15% through 2027. This supply gap not only pushes chip prices higher but also forces cloud vendors and large enterprises to lock in orders two years in advance, further intensifying the “arms race” for computing resources.

Conclusion: Computing Power, the Ultimate Strategic Resource in the AI Era

When Elon Musk declares, “For GPUs, I only buy Nvidia,” when a rocket company plans to send data centers into space, and when global cloud providers are poised to spend nearly $900 billion a year on compute—the sheer intensity of this computing race has surpassed most people’s imagination. It is no longer just a Silicon Valley capital game, but a multi‑front campaign involving aerospace engineering, energy revolution, chip manufacturing, and global supply chains.

Computing power is becoming the scarcest strategic resource of the AI era. And the endgame of this race may not be found in the data centers of Silicon Valley, but in low‑Earth orbit, in the vast expanse of space unbound by the physical limitations of the ground. In the next decade, whoever commands the high ground of computing power will hold the key to AI development. Whether SpaceX’s space‑based compute blueprint will materialize on schedule remains to be seen, but at the very least, it has opened a window—showing us an alternative direction for compute expansion: upward, not inward.