In SpaceX’s S-1, the document filed ahead of what would be the largest IPO in US history, that sentence reads roughly as there being no assurance that the SpaceX Terafab project will meet its objectives within expected timelines, or at all. It is a frank acknowledgement of what building a leading-edge semiconductor fab from scratch actually involves, and why even a company that has commercially launched rockets and built a global satellite internet network considers this its most difficult engineering challenge yet.
Terafab is SpaceX’s plan to manufacture one terawatt of AI compute hardware per year at a megafactory in Austin, Texas, targeting chips for Tesla’s Optimus robots and vehicles as well as SpaceX’s orbital compute infrastructure. Initial investment is $55 billion, with potential scaling to $119 billion across all phases. Here is what the filing actually says about why this is hard, and what SpaceX plans to do about it.
Why SpaceX Terafab Chips are needed in the First Place
The goal is to supply security. The demand for AI chips from Tesla, SpaceX, xAI, and other Musk-affiliated enterprises far exceeds existing supply chain capacity. During Tesla’s Q1 2026 earnings call, Musk framed the choice directly: “Either build Terafab, or lose the future.”
SpaceX’s Q1 2026 AI spend of $7.7 billion implies annualised spending of more than $30 billion, up sharply from $12.7 billion in fiscal 2025. It currently operates two AI clusters:
- Colossus 1 is a roughly 130-megawatt cluster running approximately 100,000 H100 GPUs.
- Colossus II is a 430-megawatt site running around 110,000 GB200 and 110,000 GB300 accelerators, with additional planned capacity targeting 1 gigawatt total.
A company spending $30 billion per year buying chips from NVIDIA has an obvious motivation to make its own. That motivation is not competitiveness, as Musk explicitly rejected the notion of “fighting Nvidia,” describing Terafab as “survival infrastructure” to solve his own compute shortage.
Challenge 1: Manufacturing at 2nm and 14-Angstrom Is Genuinely Hard
Producing large-die GPUs at 2-nanometer and 14-angstrom scales requires mastery of extreme ultraviolet lithography and advanced packaging. Early production runs at these nodes historically face low yields; even modest defect rates can generate tens of millions of dollars in scrapped wafers per month until processes stabilise.
TSMC spent years optimising yields at 3nm before reaching commercial viability. Samsung’s 3nm rollout in 2022 was marked by lower-than-expected yields that pushed major customers to TSMC instead. At 2nm and 14 angstroms (the nodes Terafab is targeting), the margin for error shrinks further.
Terafab’s mitigation is hardware-intensive. The plan calls for deploying more than 50 High-NA EUV scanners from ASML, each costing approximately $380 million with a throughput of 200 wafers per hour. These will be paired with atomic layer etch techniques to achieve precise profile control at the scale of individual atomic layers, the level of precision required for transistor gates and interconnects at advanced nodes.
Fifty High-NA EUV scanners represent an investment of roughly $19 billion in lithography equipment alone, before fab construction, utilities, or any other tooling. That figure alone explains why the $55 billion initial commitment is a starting point.
Challenge 2: Hardware Without Software Is Unusable
A GPU that runs fast is worthless if developers cannot program it. NVIDIA’s dominance in AI compute is not primarily about silicon performance but about CUDA, the software platform that has been accumulating developer investment for almost two decades. Every AI framework, every research codebase, every inference optimization that exists in the world has been written with NVIDIA’s software stack in mind.
Building a competitive software and driver stack is a multi-year engineering effort. Intel required months, if not years, to fine-tune its Arc GPU drivers and architecture to deliver a noteworthy alternative to NVIDIA’s GeForce RTX and AMD’s Radeon RX offerings. Intel entered that effort with decades of driver engineering experience and a massive existing developer ecosystem. Terafab starts from zero.
The S-1 filing does not detail a specific software mitigation strategy beyond acknowledging the dependency. The filing’s language notes that manufactured chips have “limited practical utility for AI workloads” without a complete software stack. It is an honest characterisation of how far this challenge extends beyond the fab itself.
Challenge 3: The Intel Partnership Is Not a Contract
SpaceX has referenced Intel’s 14A process node as the manufacturing foundation for Terafab. Intel confirmed the partnership directly, and Intel’s stock rose more than 3% on the announcement, meaningful for a company whose foundry business lost $10.3 billion in 2025.
The S-1, however, is careful about language. SpaceX’s own filing describes the Terafab and Tesla collaboration as in “very early stages,” with no financial terms, no intellectual property rights, and no binding commitments finalized. A partnership described as non-binding in a company’s own IPO filing, a document where understatement carries legal consequences, is a partnership that investors should not treat as settled.
The S-1 also notes the lack of long-term supplier contracts more broadly. A leading-edge fab depends on specialized inputs and equipment, and supply chain uncertainty compounds the execution risk of a project already operating at the frontier of what semiconductor manufacturing can do.
Challenge 4: $119 Billion Is a Number That History Treats Unkindly
Semiconductor facility projects have a documented pattern of cost overruns and timeline extensions. Intel’s Ohio fab, announced in 2022 with a $20 billion initial commitment, has faced repeated delays. TSMC’s Arizona expansion has been subject to similar schedule revisions. These are companies with decades of fab construction experience.
SpaceX is considering spinning off Terafab as a separate publicly listed entity to independently raise funds, a structure that would keep capital pressure from contaminating the core SpaceX balance sheet while Musk retains a controlling interest. That contingency plan, disclosed in the same filing, is an implicit acknowledgement that $119 billion across all phases is not a commitment SpaceX can self-fund.
The Yield Management Bet: ML Compression of a Multi-Year Problem
The most technically ambitious claim in the filing’s mitigation section is the timeline for yield improvement. Traditional semiconductor fabs take two to three years to move from initial process runs, where yields might sit at 40–60%, to commercially viable production above 90%. Terafab’s target is to compress that ramp to months.
The mechanism is a real-time metrology and machine-learning system. Partnerships with KLA and Camtek will provide high-speed inspection using broadband plasma, laser diagnostics, critical-dimension scanning electron microscopy, and X-ray photoelectron spectroscopy. A dedicated ML infrastructure will process defect and electrical test data to predict failures and identify yield limiters in near-real time. The stated goal: move from initial yields near 60% toward 95% at a speed no conventional fab ramp has achieved.
This is the most speculative element of the entire technical plan. KLA and Camtek are credible metrology partners, and ML-driven process control is an active area of semiconductor R&D. Whether those tools can deliver a yield ramp that compresses years into months on a brand-new process node, in a brand-new facility by a team with no prior fab operation experience, is a question the filing raises but does not answer.
What to Watch Next
SpaceX’s IPO roadshow is targeted for the week of June 8, 2026. Press consensus points to a $1.75 trillion valuation and a $75 billion raise. If these figures are confirmed, it would more than double Saudi Aramco’s $29.4 billion record for the largest IPO ever.
The S-1’s full public disclosure will be the first time outside investors can read the complete filing. Terafab’s progress, specifically whether the Intel partnership moves from early-stage to contractually binding, will be the most important technical milestone to watch in the months that follow.
Comments are closed.