Goldman Sachs' latest forecast suggests SpaceX's Starship program could launch over 5,000 dedicated AI missions by 2031, potentially deploying millions of AI accelerators in orbit and creating a memory chip shortage unlike anything the industry has ever faced. While the projection sounds like science fiction, even a fraction of these launches would require memory production on a scale that has never been attempted, making semiconductor suppliers the unexpected winners in the space AI boom.

What Is Goldman Sachs Predicting for SpaceX's AI Ambitions?

Goldman Sachs recently published research outlining an ambitious long-term vision for SpaceX's Starship program that centers on space-based data centers and AI satellites. The forecast estimates 5,288 dedicated AI missions by 2031, with each Starship launch carrying 30 to 50 AI satellites. Every satellite would house roughly one GB300-equivalent AI rack, which is Nvidia's latest computing architecture designed for artificial intelligence workloads.

To understand the scale, consider that Nvidia's Blackwell architecture, and its successor Vera Rubin, relies on eight high-bandwidth memory (HBM) stacks per accelerator. A single AI rack contains many accelerators, meaning every launch could require thousands of HBM stacks before accounting for conventional DRAM and flash storage needed throughout the system. Some analysts extrapolating Goldman Sachs' assumptions estimate those launches could eventually translate into millions of Nvidia accelerators in orbit, with the cumulative installed base potentially exceeding 200 million accelerators by 2031 if every projected mission ultimately flies.

Why Would Memory Chips Become the Real Bottleneck?

The AI infrastructure discussion typically centers on graphics processing units (GPUs) like those made by Nvidia. However, these accelerators cannot function without enormous amounts of memory. Micron Technology, along with SK hynix and Samsung, is one of only three companies capable of manufacturing leading-edge HBM at scale. Micron has already revealed long-term HBM supply agreements extending well into future production cycles, reflecting how constrained supply remains even today.

The memory challenge extends beyond just HBM. Every advanced accelerator needs HBM, and every AI rack requires even more conventional memory around it. Whether those chips sit inside terrestrial hyperscale data centers or eventually orbit Earth, memory manufacturers remain one of the few companies positioned to supply a resource the entire AI industry cannot function without.

How to Understand the Scale of This Demand

• Launch Frequency: Goldman Sachs projects 5,288 AI missions by 2031, which would require SpaceX to achieve routine launch reliability and regulatory approval for thousands of launches over the next five years.
• Memory Requirements Per Launch: Each Starship could carry 30 to 50 AI satellites, with every satellite housing roughly one GB300-equivalent AI rack, translating to thousands of HBM stacks per mission before accounting for additional DRAM and flash storage.
• Total Accelerator Deployment: Some analysts estimate the cumulative installed base could exceed 200 million Nvidia accelerators by 2031 if every projected mission ultimately flies, a figure that would consume memory production on an unprecedented scale.

Some observers have even suggested that such deployment would ultimately consume every advanced wafer Taiwan Semiconductor Manufacturing (TSMC) could produce. Even if that is an exaggeration, it illustrates just how large these assumptions have become. Goldman Sachs' projections represent a best-case scenario, not a roadmap. Everything would need to go right: Starship must achieve routine launch reliability, regulators would need to approve thousands of launches, and orbital AI data centers must prove technically and economically viable.

Early missions during 2027 and 2028 would almost certainly be demonstration projects before any meaningful scaling occurs. There is also an interesting contradiction buried inside the broader investment thesis. Goldman Sachs' estimates assume orbital AI data centers could cost roughly $15 billion to $20 billion per gigawatt, well below the approximately $28 billion to $32 billion per gigawatt often cited for terrestrial AI facilities. However, that cost advantage would necessitate a future SpaceX-Tesla manufacturing effort producing custom AI chips internally rather than continuing to rely primarily on Nvidia hardware.

In other words, the model initially assumes enormous Nvidia deployment, while the long-term economics become more attractive only if Nvidia eventually becomes less central to the equation. This contradiction highlights the speculative nature of the forecast, yet it does not diminish the fundamental insight: even if Starship completes only a fraction of those launches, AI infrastructure demand appears poised to outgrow memory supply for years to come.

Investors should not buy memory stocks because Goldman Sachs predicts exactly 5,288 AI missions. That figure demands nearly flawless execution across launch technology, satellite engineering, manufacturing capacity, and regulation. What matters is the direction of travel. The broader point is that memory manufacturers occupy a critical position in the AI supply chain, one that will only become more important as both terrestrial and orbital AI infrastructure scales. For long-term investors, that is the part of Goldman Sachs' ambitious forecast worth paying attention to.