The SpaceX included in-house manufacturing of GPUs in the document’s list of “substantial capital expenditures” S-1 filed with the U.S. Securities and Exchange Commission (SEC) in preparation for a IPO valued at US$1.75 trillion scheduled for summer 2026. Registration has been reviewed by Reuters and presents the first time that the Elon Musk publicly formalizes this intention before investors.
Disclosure is directly linked to the project Terafaban industrial chip manufacturing complex for artificial intelligence applications being jointly developed by SpaceX, xAI and Teslawith facilities planned for Austin, Texas.
The project was officially announced by Musk in March 21, 2026during an event at the former Seaholm Power Plant, and the Intel joined the initiative in April 7, 2026.
What is Terafab and what does it intend to produce
THE Terafab is described as a vertically integrated factory: unlike the conventional model of the semiconductor industry, in which steps such as design, lithography, manufacturing, packaging and testing are carried out by separate companies around the world, Musk’s project intends to concentrate the entire process under one roof.
The technical goal is bold: the installation aims to work with 2 nanometer process nodesthe most advanced currently transitioning to commercial production, with an initial capacity of 100,000 wafer batches per month.
For reference, the TSMC it takes decades and hundreds of billions of dollars to build equivalent capacity, and its first 2nm fabs in the US are not expected to reach full production until 2029.
The estimated cost of the project is between US$20 billion and US$25 billion. Musk has already indicated that when Terafab is at scale, he intends to use Intel’s next-generation process, the 14Adescribing the technology as “probably quite mature or ready for use in large-scale production” at that point.
Two types of chips are in initial scope: inference chips for Tesla vehicles and for those Optimus humanoid robots (including chip AI5whose small volume production is scheduled for 2026 with larger scale in 2027), and the chips D3developed for AI satellites in orbit.
The motivation: chip shortages and dependence on third parties
In own document S-1, SpaceX recognizes a strategic vulnerability: the company “does not have long-term contracts with many of its direct chip suppliers” and wait “continue to source a significant portion of its computing hardware from third-party vendors”. In-house manufacturing of GPUs is listed as a direct response to this risk.
Musk was more direct in his tone when presenting the project: he stated that existing factories on the planet produce only about 2% of what Tesla and SpaceX will need for all your projects combined. The conclusion was categorical: “Either we build Terafab, or we won’t have the chips. And we need the chips, so let’s build Terafab.”
The broader market outlook also weighs in on this equation: memory shortages RAM has caused instability in consumer electronics prices throughout 2025 and 2026, and growing demand for hardware for training AI models increases pressure on the entire semiconductor supply chain.
What exactly are these “GPUs”?
An important point of ambiguity permeates the announcement: the S-1 document uses the term “GPU” to describe the chips that SpaceX intends to manufacture, but does not specify whether this designation refers to conventional graphics processing units or broader AI accelerators.
This distinction is very important. THE Teslafor example, has already called your processor AI5 of “GPU” in previous communications, even though it is a dedicated inference chip, with no purpose in traditional graphics processing.
THE NVIDIA dominates the AI GPU market with general-purpose architectures, while Google operates with its TPUs (Tensor Processing Units), tuned for specific model training tasks.
If Terafab follows a similar path to TPUs, or develops proprietary architectures for inference in vehicles and robots, direct overlap with NVIDIA may be smaller than the project title suggests.
Musk’s recognized risks and track record of promises
The IPO document itself carries an explicit caveat: “There is no guarantee that we will be able to achieve our objectives with respect to Terafab within the expected timeframes, or at all.” This boilerplate language in SEC filings gains weight when placed alongside the Musk ecosystem’s recent history of ambitious projects.
The battery cell program Tesla 4680announced in September 2020 with promises of massive production within a year and 50% cost cuts, ran years over schedule, underwent six or seven revisions to the dry electrode process, and is still far short of its original goals.
THE Tesla does not have experience in semiconductor manufacturing. Building a single 2nm fab with a capacity of 50,000 wafer starts per month costs approximately US$28 billion and it takes about 38 months just to build the structure in the USA.
The announcement also came at a specific time: Tesla vehicle sales fell for the second consecutive year in 2025, with a significant retraction in Europe and the first annual drop in China, while SpaceX prepares to go public with one of the highest valuations in history.
Analysts like Dan Levyfrom Barclaysrecognize the strategic logic of the project, but point out that the volume of capital involved and the technical challenges of wafer yield leave the financial viability of the venture open to question.
“We are very grateful to our existing supply chain, to Samsung, TSMC, Micron and others. But there is a maximum rate at which they are comfortable expanding. That rate is much lower than we would like… and we need the chips, so let’s build Terafab.”
Elon Muskin a meeting with Tesla analysts
Terafab as a vertical integration bet in a market dominated by giants
If the project gets off the ground within the deadlines, Terafab would represent a relevant structural transformation: for the first time, a company outside the traditional semiconductor segment would try build a complete chip manufacturing chains for own use on an industrial scale, integrating design, lithography, testing and packaging.
The closest parallel is the path taken by Apple with its M series chips and Graviton processors from Amazonbut these efforts start from design and outsource manufacturing to TSMC or Samsung.
SpaceX and Tesla aim to do what no technology company has done to date: control the knife and cheese in the silicon process, from mask to tested chip. The ambition is monumental. The execution, for now, is still on paper.
Sources): Investing.com (Reuters)
TeraFab: Elon Musk’s new project promises to surpass all chip factories in the world.
Source: www.adrenaline.com.br
Source link
