SpaceX is planning to commence in-house fabrication of its own graphics processing units. This insight emerges from leaked details within a confidentially submitted S-1 filing, valued at $1.75 trillion (~140 trillion rubles), which sheds light on the company’s strategic business objectives. The documentation explicitly highlights an intention to channel billions into the production of select processors needed internally, a move necessitated by the absence of long-term supply agreements with silicon chip vendors. Notably, the chips being manufactured internally are termed GPUs, rather than specialized ASICs intended for AI acceleration, as reported by Reuters. Nevertheless, the precise terminology remains a subject of ongoing debate. This revelation follows yesterday’s significant announcement that Elon Musk will utilize Intel’s 14A process technology at a new TeraFab chip fabrication facility, with SpaceX set to manage the operational aspects of the production site. The S-1 form, which Reuters has reviewed, cites “in-house GPU production” as a driver for the “significant capital expenditures” the company anticipates incurring in the future, stemming from what appears to be a lack of “long-term contracts with many of our direct chip suppliers.” While the confirmation that SpaceX would build and operate its own high-performance semiconductor fabrication plant for manufacturing silicon chips designed by Tesla was provided by Elon Musk, CEO of both Tesla and SpaceX, last night, the company’s explicit plan to create “in-house GPUs” seems to be a novel development. Reuters speculates that various companies employ different nomenclature for their AI accelerators. While AMD and Nvidia explicitly label theirs as “GPUs,” Google refers to its AI accelerators as “TPUs,” Microsoft calls its Maia devices “accelerators,” and SambaNova terms its AI chips “RDUs.” The majority of hyperscalers and independent hardware vendors (IHVs) designate their AI accelerators as Application-Specific Integrated Circuits (ASICs). Given that SpaceX does not specifically mention the Tesla-designed AI-ASICs, Reuters suggests that SpaceX aims to develop and manufacture what they term “in-house GPUs,” implying something distinct from Tesla’s existing AI processor series. It remains unclear whether SpaceX—or its xAI division—possesses the capability to engineer sophisticated custom chips capable of competing with the AI GPUs/accelerators from AMD, Nvidia, Rebellions, or SambaNova. However, Elon Musk himself previously referred to Tesla’s AI5 processor as a “GPU,” even though this chip lacks the dedicated hardware for traditional computer graphics processing. “In AI5, we removed the legacy GPU, the traditional GPU that was in AI4, but AI5 doesn’t have a legacy GPU because it is a GPU itself,” Musk stated during an earnings call. “We also removed the image processing processor. There is a long list of very important removals.” Considering the inconsistent naming conventions for AI accelerators at Tesla—largely because they are optimized for highly specific tasks—it is difficult to ascertain whether SpaceX’s plans involve manufacturing Tesla AI5 and/or AI6, or if they are indeed pursuing their own distinct “GPUs.” As the S-1 form was filed confidentially, Tom’s Hardware has been unable to independently verify its contents. Rutab.net
About the Author
