According to a Nikkei exclusive dated November 9, Terafab, the semiconductor manufacturing initiative led by Elon Musk, has presented chipmaking equipment suppliers with a long-term plan. From 2032 onward, it would secure wafer-processing capacity of around 1 million wafers per month, with about 700,000 of them devoted to memory production. Earlier disclosures described a concept of making logic chips and memory at the same site. This report reveals a concrete allocation of capacity centered on memory. However, the number of wafers processed is an indicator of a factory's production scale. It is not the same as the memory capacity available for AI computing or the number of finished products shipped. The start date of mass production, the product mix, and yields will greatly change what the figure of 700,000 wafers a month means as real supply capacity.
About 70% of the 1-million-wafer plan is allocated to memory
Under the long-term plan Nikkei reported, monthly wafer-processing capacity would be split between about 700,000 wafers for memory and at least 300,000 for logic semiconductors, which handle computation and similar tasks. If the total is around 1 million wafers a month, memory would account for roughly 70%. But this is a ratio based on wafer counts. It does not mean that 70% of capital expenditure or revenue would be memory-related. In addition, the overall capacity is given as a rough figure and logic is described as "at least 300,000 wafers," so the split should not be treated as a precise number.
What had been officially presented until now was a concept of consolidating different semiconductor manufacturing processes at a single site. On August 6, the Texas Governor's Office announced that capital investment in the first phase of SpaceX's expansion in Grimes County, Texas, would exceed $16.8 billion. In Musk's comments included in that announcement, he explained a policy of consolidating not only logic and memory manufacturing but also advanced packaging, the process of connecting and assembling chips, at the same site. However, $16.8 billion is the investment plan for the first phase. It is neither money already spent nor a sign that all the funding needed to expand to 1 million wafers a month has been secured.
According to the latest Nikkei report, Terafab's procurement officials visited Japan in early November and explained the plan to semiconductor manufacturing equipment makers. It also reportedly said they asked for delivery of some equipment within a few months. This suggests that procurement negotiations with equipment makers are moving forward as the concept takes shape, but the total order value and the quantities of firmly confirmed orders have not been disclosed.
If about 70% of production capacity can be directed to memory as planned, Terafab would become a large-scale site that makes not only computing chips but also the memory that supplies the data those chips need to process. However, planned equipment capacity and actual supply capacity are not the same. To assess Terafab's impact on the semiconductor market, it is necessary to look not only at the scale of equipment installed but also to consider, separately, when each product enters mass production and how capacity expands afterward.
The 2027 prototype start and the post-2032 capacity target are separate things
At SpaceX's earnings call on October 20, Musk said he wants to get the first SRAM prototype chips running at the research fab in Austin in May or June 2027. These prototype chips are meant to validate the manufacturing line. As for the main Terafab facility to be built at Gibbons Creek, the aim is to start mass production around the third quarter of 2028, with the possibility that it slips into the fourth quarter. Musk himself acknowledged substantial uncertainty around these targets, describing them as extremely ambitious and saying it is unclear whether they can be achieved. These remarks can be confirmed in the official transcript of the earnings call released by SpaceX (page 9).
At the same call, Musk also expressed gratitude to equipment makers such as ASML for moving up delivery schedules. However, he did not disclose the number of EUV lithography systems ordered, the models, or delivery schedules. Even if manufacturing equipment arrives ahead of schedule, that does not immediately mean a mass-production line will run stably, so equipment procurement and the actual ramp-up of production need to be evaluated separately.
Terafab's development plan can be organized into three stages: confirming that prototype chips work, starting mass production, and expanding capacity to the scale of 1 million wafers a month.
| Timing | Goal | Status of the information |
|---|---|---|
| May–June 2027 | Run SRAM prototype chips at the Austin research fab and validate the manufacturing line | Target Musk presented on October 20 |
| Q3–Q4 2028 | Start mass production at Terafab in Gibbons Creek | Ambitious target presented by Musk; achievement not assured |
| 2032 onward | Secure capacity of about 1 million wafers a month, with about 700,000 allocated to memory | Long-term plan reportedly presented to equipment makers, per Nikkei |
Distinguishing these three stages shows that even if the goal of starting mass production in 2028 is met, it does not necessarily mean capacity of 700,000 memory wafers a month will be secured that year. In the first place, the monthly processing capacity envisioned for 2028 has not been disclosed. Also, successfully confirming that SRAM prototype chips work would move validation of the manufacturing line forward, but it would not by itself mean that mass-production technology for DRAM or stacked memory has been established.
The gap between when a factory produces its first wafers and when market supply actually increases is also evident in existing manufacturers' plans. For example, according to Micron's earnings materials published on September 30 (pages 8–9), the company plans to produce its first wafers at its ID1 fab in mid-2027. It also explains that bringing up new DRAM and NAND fabs takes time, and that their contribution to supply starts to become significant several quarters after the first wafer output. Micron's fab ramp-up period cannot simply be applied to Terafab, but it is a concrete example showing that starting manufacturing and establishing a large-scale supply system are different stages.
For HBM, wafer counts cannot be converted directly into supply capacity
A representative example showing that wafer-processing counts do not necessarily match the supply of finished products, when considering the supply of AI memory, is HBM (high-bandwidth memory). HBM is a technology that stacks multiple DRAM chips vertically and connects them with through-silicon vias (TSVs) and other means to secure high memory bandwidth. According to Micron's product explanation, HBM is placed close to a CPU or GPU and built into the same package. Therefore, even after DRAM circuits are formed on a wafer, steps such as stacking, connecting, and testing the chips are needed, and only after these does it become a product that can be used in combination with a processor.
The number of wafers HBM needs also differs from that of ordinary DRAM. According to Micron's 2026 annual report (Form 10-K, page 28), when producing the same number of bits of memory at the same technology generation, HBM typically requires more wafers and more cleanroom space than standard DRAM. What matters here is that the comparison conditions are "the same technology generation" and "the same number of bits." In other words, even with the same number of wafers processed, the capacity that can ultimately be supplied changes depending on the type of memory produced.
Micron's report also cites raising the yield and quality of stacked HBM to sufficient levels, and developing advanced packaging technology, as technical challenges. Yield refers to the proportion of manufactured chips or products that meet quality standards. Even if a factory can process a certain number of wafers, the memory capacity it can actually supply varies with the number of good chips obtained from them. For HBM, quality standards must also be met in the stacking and assembly processes, so the supply of finished products cannot be judged from wafer-processing capacity alone.
For the reported Terafab memory plan of 700,000 wafers a month, no official product mix has been confirmed, such as devoting all of it to DRAM or HBM. HBM is therefore not a product Terafab has been confirmed to make. It is a concrete example used to explain the difficulty of estimating actual supply from wafer counts. Once the product mix, the technology generation adopted, and yields become clear, it will be possible to estimate the memory capacity that can be produced as good output and to assess the supply capacity of finished products, including the stacking and assembly steps. Until then, the figure of 700,000 wafers a month should be treated as a target indicating the scale of the factory's production, and distinguished from memory capacity that is actually usable.
Strengthening supply for its own use is different from the impact on the whole semiconductor market
In its 2026 IPO prospectus, SpaceX cited preparing for future semiconductor shortages and designing chips suited to the space environment as purposes of Terafab. At the same time, it indicated that it expects to continue sourcing a considerable portion of its computing hardware from external suppliers. Both of these are policies as of the time the prospectus was prepared, and they do not amount to a plan to stop external procurement altogether once Terafab is complete.
Given this policy and the production capacity allocation now reported, the emphasis on memory manufacturing can be seen as aimed at securing a stable supply of the components that support the company's own computing demand. If the plan is realized, the first thing likely to change is the relationship between the quantity of semiconductors that SpaceX and others buy from outside manufacturers and the quantity that can be covered by in-house production. If products made in-house can replace what was previously procured externally, existing manufacturers' production capacity could gain some slack, which might indirectly increase the quantity available to other customers. However, the products and quantities Terafab would sell directly to the external market have not been disclosed, so its entire planned production capacity cannot simply be counted as an increase in supply for the semiconductor market as a whole.
Furthermore, by 2032 and beyond, when Terafab aims to expand to 1 million wafers a month, existing manufacturers will also have upgraded their production facilities and manufacturing technology. According to Micron's earnings materials (pages 7–9), the company plans mass production of next-generation DRAM and NAND technologies in the second half of 2027. As for fab expansion, it plans to produce first wafers at ID2 in the second half of 2028 and at its first New York fab in 2030. These are also future plans as of now, but by the time Terafab reaches full-scale operation, competitors' product generations and production capacity are likely to have changed. For that reason, calculating Terafab's market share after 2032 based on current manufacturing capacity would not give a reliable comparison.
In the same earnings materials (page 4), Micron also forecasts that supply and demand in the memory and storage market in fiscal 2027 and fiscal 2028 will be tighter than in fiscal 2026. However, there is a gap of several years between the period this outlook covers and 2032 and beyond, when Terafab aims for a scale of 1 million wafers a month. The existence of a long-term plan for large-scale production increases does not necessarily mean it will ease the supply shortage expected in the near future.
In judging Terafab's capabilities, what will matter going forward is not only the number of wafers the equipment can process but also how many good products can be produced for each product type. For stacked memory, the supply capacity of products that have completed packaging will be the question, and the market impact will further depend on how output is divided between in-house demand and the external market. Only once the product mix, yields, and assembly-process capabilities become concrete will it be possible to evaluate how much AI computing demand the long-term target of 700,000 wafers a month can actually support.