SIAgentNews.com
← All news
Developing

SpaceX builds massive AI compute infrastructure to secure hardware edge

SpaceX board member Steve Jurvetson explains how the company is vertically integrating the compute, energy, and hardware supply chain to power the AI era.

· First reported by finance.biggo.com

60%SpaceX compute cost vs industry average

The short version

  • SpaceX plans to pass 1 gigawatt of AI compute per month by the end of 2027.
  • The company is building compute at roughly 60% of the industry cost through vertical integration.
  • Terafab will be a new plant potentially becoming the largest factory on the planet at 100 million square feet.
  • SpaceX is adding roughly 500 megawatts of AI compute per month and 220,000 GP300 chips.

Why it matters

This shift suggests that the winners of the AI race may be determined by who controls the physical infrastructure rather than just the software models. For readers, this means the 'AI bottleneck' is moving toward energy, hardware manufacturing, and orbital connectivity.

Vertical integration as a moat

SpaceX is moving beyond just launching rockets to become a primary infrastructure provider for the artificial intelligence era. Board member Steve Jurvetson argues that the defining bet of this period is not the AI models themselves, but the underlying compute, energy, and hardware supply chain. By controlling these layers, SpaceX aims to create a vertically integrated stack that supports recursive self-improvement in AI.

The company achieves a significant cost advantage by building much of its own infrastructure in-house. Jurvetson notes that SpaceX builds components that are typically outsourced, including what he describes as the largest printed circuit board facility in America located in Bastrop, Texas. This approach allows them to build compute at roughly 60% of the industry cost by refusing to accept mediocrity and hiring from outside traditional domains.

This strategy is a direct response to the risk-averse nature of incumbent chipmakers who are tied to specific capital expenditure cycles. Jurvetson suggests that exponential AI demand breaks the logic of traditional market cycles, making SpaceX's mission-driven approach more effective. Because the company is not prioritizing quarterly profits above all else, it can focus on long-term infrastructure goals that others might find too capital-intensive.

The Terafab factory expansion

A major pillar of this hardware strategy is the development of Terafab, a new manufacturing plant. Jurvetson describes it as potentially the largest factory on the planet, spanning 100 million square feet. The facility is intended to move SpaceX into the realm of high-volume chip production, with phase-one volume production targeted for late 2028.

The goal for Terafab is to produce the first chips as early as next year. This move is designed to address the hardware bottleneck by creating a dedicated pipeline for the specialized silicon required for frontier models. By owning the factory, SpaceX can bypass the queues and limitations faced by companies that rely solely on external suppliers like Nvidia.

This expansion represents a massive bet on the continued growth of AI demand. Jurvetson argues that while some predict a roll-off in hardware demand as models become more efficient, the reality is that cheaper inference will be absorbed by continuous learning across every enterprise. SpaceX is building for a world where heterogeneous clusters and stranded energy can be utilized to run inference anywhere.

Scaling the compute power

SpaceX is currently scaling its compute capacity at a rapid pace to meet the needs of the AI industry. The company is adding roughly 500 megawatts of AI compute per month, which includes approximately 220,000 GP300 chips. These efforts are part of a trajectory to pass 1 gigawatt per month by the end of 2027, reaching an annual capacity of 6 to 12 gigawatts.

This massive buildout is supported by a combination of custom hardware and sophisticated networking. Google, for example, is landing about 8 million TPUs next year, but SpaceX is carving out its own niche through unique interconnects and vertical scaling. The focus is on creating a learning loop that matters more than static benchmarks, prioritizing the volume of releases and physical AI variants.

The infrastructure also extends to orbital capabilities through Starlink. While currently used for human communication, the real market for Starlink may shift toward agent traffic, which requires trillions of bits per second. SpaceX is positioning Starlink as a carrier for this traffic, utilizing optical laser links for satellite-to-satellite backhaul to avoid submarine cable paths.

The shift to robotics

The infrastructure and compute power being built by SpaceX are also intended to fuel advancements in robotics, specifically the Optimus humanoid. Musk has suggested that Optimus could perform domestic tasks like doing dishes within 12 to 18 months. However, the first major industrial rollout is expected to occur in mining and refining ore for national security purposes.

Jurvetson reframes the robotics landscape by noting that almost everything SpaceX ships is already a robot. Every fairing half, upper stage, and Dragon capsule functions as a fully autonomous vehicle. This experience in autonomous systems provides a blueprint for how humanoid robots will be deployed in dangerous, dull, and dirty environments like mines.

While humanoids will handle odd jobs and inspections, specialized equipment will still win in specific tasks like tunneling. For example, a boring machine is optimized for cladding placement and tunneling, whereas a humanoid would not be the ideal tool for digging a tunnel sized for its own frame. The goal is a hybrid environment where specialized machines and humanoids work together on lunar and terrestrial operations.

What to watch

  • The progress of the Terafab factory construction and its first chip production dates.
  • The transition of Starlink to supporting high-bandwidth AI agent traffic.
  • The timeline for the first industrial rollout of the Optimus humanoid robot in mining.

Try it

Look up the current specs for Google's Gemini 4 Neon to see how it compares to the cost-performance frontier of Anthropic and OpenAI.

Sources and links

Read the original at finance.biggo.com

First reported by finance.biggo.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news