Elon Musk is reported to have told investors that Terafab, the chip megafactory Tesla and SpaceX are planning in Texas, will need "tens of billions" of dollars of equipment through early 2028 and will "probably be north of $100 billion" long term, with volume production around Q3 or Q4 2028. The remarks, relayed in posts on October 20 and 21, 2026 and summarized by aggregators, are the latest in a stream of changing Terafab numbers. This post separates what Musk said from what filings and reporting confirm.
We rely on Electrek's August report on the site and first phase, Tom's Hardware's report on the October 20 SpaceX earnings call, search-surfaced coverage of earlier cost figures, and posts relaying the remarks. Tom's Hardware confirms the volume-production target and Musk's own caveat that the timeline is "highly aspirational" with "big error bars." We did not read a transcript, so the dollar figures and the SRAM detail are still secondhand. For background see our explainer What is Terafab?.
TL;DR
| Claim | Status |
|---|---|
| Tens of billions in capital equipment through early 2028 | Reported Musk remark; no filing |
| Long-term cost "probably north of $100 billion" | Spoken estimate; a county filing reportedly outlines up to $119B total |
| Funding split between Tesla and SpaceX undecided | Reported remark |
| Most equipment for first fab already ordered; ASML accelerating deliveries | Reported remark; no vendor confirmation found |
| SRAM test chip around May or June 2027 at Austin research fab | Reported remark |
| Volume production around Q3 or Q4 2028 at Gibbons Creek | Reported remark |
| SpaceX casting turbine parts to bring power online 12 to 18 months faster | Musk claim; unverified |
| First phase about $16.8B, 100M sq ft finished, 3,000+ jobs, Grimes County | Confirmed by Tesla and SpaceX in August, per Electrek |
What is confirmed
On August 6, 2026, Tesla and SpaceX confirmed where Terafab will be built: Grimes County, Texas, about an hour northwest of Houston, on the Gibbons Creek Reservoir that formerly cooled a coal plant shut in 2018. They put the initial phase at roughly $16.8 billion, the finished site at more than 100 million square feet, and staffing at least 3,000, per Electrek. SpaceX said Terafab would combine logic, memory, packaging and testing and produce chips for Tesla's Optimus robots and Cybercabs and for planned space-based data centers, with "future expansion phases bringing total investment much higher." In April the companies broke ground on a research fab at Giga Texas, the precursor.
That is the firm ground: a site, a phase-one number, a scale target and a research fab.
How the cost number has moved
| Date | Figure | Source type |
|---|---|---|
| March 2026 | $25 billion headline at the first announcement; other reports cited $20 to 25 billion | Announcement coverage |
| Morgan Stanley estimate | $35 to 45 billion total | Analyst estimate, via search coverage |
| May 2026 | $55 billion for initial phases, up to $119 billion total | County filing, as reported |
| August 2026 | $16.8 billion initial phase | Tesla and SpaceX confirmation |
| October 2026 | "North of $100 billion" long term | Reported Musk remark |
The October statement is consistent with the reported up-to-$119 billion total, but it is not a new commitment. It is also a long-term estimate, while the equipment figure of "tens of billions" through early 2028 is a near-term capital need. Neither is binding.
Reasons for caution
Electrek's August report lists the reasons, and they remain relevant.
- SpaceX's S-1 language. In May, SpaceX's IPO filing described Terafab as only a "general framework" with no binding commitments, no finalized IP split and no obligation for either side to keep participating.
- Who builds it. A month earlier Electrek had reported Tesla would not build the fab itself, with Intel lined up for manufacturing.
- Scale is unprecedented. Electrek notes 100 million square feet is more than ten times the floor space of TSMC's biggest gigafabs, and no fab is built to output a terawatt of compute a year; that target comes from Musk's space ambitions.
- Track record. Tesla's near-term chips are made by Samsung (AI5 in Taylor, Texas) and TSMC, and Tesla signed a $16.5 billion deal with Samsung for AI6.
The timeline claims deserve the same skepticism. A functional SRAM test chip in mid-2027 would be a first-silicon milestone, not a proof of manufacturing yield. Volume production roughly 15 to 18 months later at a new site would be extremely fast by industry standards, where new fabs typically take years from groundbreaking to volume. Musk's reported explanation, that ASML and others have expedited equipment shipments, is not confirmed by the vendors in anything we found.
Milestones as claimed, and what each would prove
| Reported milestone | Date | What it would prove | What it would not prove |
|---|---|---|---|
| Most first-fab equipment ordered | Already, per remark | Procurement has started | That tools are delivered, installed or qualified |
| Functional SRAM test chip at Austin research fab | May or June 2027 | The team can make working silicon on a research line | Yield, cost per wafer or anything about advanced logic |
| Tens of billions in equipment spent | Through early 2028 | Capital is committed at scale | That the production site is ready |
| Volume production at Gibbons Creek | Q3 or Q4 2028 | A real fab at commercial output | Competitive process node, quality or price versus TSMC and Samsung |
| Output at the scale of a "top 20 country GDP" in one building | Unspecified | Not testable | Any concrete capacity number |
SRAM is the usual first test vehicle for a new process because its regular structure reveals defects quickly, so a working SRAM chip is a standard first milestone and a sensible one to announce. It says little about whether the process can build a leading-edge GPU or accelerator. Treat it as a checkpoint for the fab team, not for Tesla's or xAI's chip roadmap.
Terafab versus buying capacity
The case for Terafab, as Musk's remarks frame it, is that existing suppliers cannot or will not expand fast enough. The case against is that leading-edge chipmaking depends on process know-how, a mature supplier ecosystem and trained engineers that take years to build, which is why established foundries invest heavily in each node. Tesla currently buys chips from Samsung and TSMC, and Nvidia is separately funding suppliers' expansions in packaging and test. A new entrant must either license or acquire process technology or develop it. Tom's Hardware reports that Intel officially joined the venture in April 2026 to help co-develop the fab process, with access to its 14A node, and Electrek had reported Intel lined up to handle actual manufacturing, though SpaceX's S-1 language leaves roles open. Until the participants publish binding terms, the more accurate description of Terafab is an ambitious program with a confirmed site and a first phase, not a committed supplier.
What would change our assessment
- A binding agreement between Tesla and SpaceX that fixes funding shares and IP.
- Published terms for the Intel partnership and a confirmed target node.
- Equipment vendor confirmation of delivery and installation dates.
- A first-silicon demonstration on schedule in mid-2027.
- Environmental and utility approvals for the Gibbons Creek site, including the reservoir water plan.
The turbine and power claim
Reported remarks say SpaceX is using its rocket-engine casting shop to make turbine blades and vanes so it can bring power online 12 to 18 months faster. That is a claim about on-site generation for a very large fab. It is plausible as a way to bypass grid interconnection queues, which is a real bottleneck for data centers and fabs. It is not verified, and power equipment qualification is a separate hurdle from casting capability. Our coverage of Trump's push for data centers to supply their own electricity shows how central power has become.
Why it matters for people who build with AI
- No near-term supply. Terafab output is targeted for 2028 at the earliest. It does not change GPU availability or pricing in 2027.
- A signal of demand. SpaceX (which absorbed xAI), Tesla and others say they need more chips than the industry can make. Reported targets of 13 GW of AI compute by end of 2027 and a 440,000-GPU cluster by Q2 2027 appear in the relayed posts but are unverified.
- Competition for the packaging stack. Terafab plans to include packaging and test. Meanwhile Nvidia is paying suppliers for capacity; see Nvidia's packaging deals.
- Treat timelines as targets. Musk's schedule record is mixed; our Tesla Optimus production claims check applied the same lens.
What the excluded-supplier remark means
One relayed post says Terafab is excluding Chinese suppliers entirely and wants Korean suppliers because Japanese products are expensive. That is a single secondhand claim, unconfirmed by the companies in anything we read. If accurate, it would shape the equipment and materials supply chain and align with US export-control politics, but it would also narrow the pool of vendors for a project that already depends on a handful of tool makers.
How to verify
- Read SpaceX's and Tesla's SEC filings, including the S-1 risk factors and the Q3 2026 earnings materials, for binding commitments.
- Look for county and state documents on Grimes County permits and incentives.
- Watch for equipment vendor statements (ASML and others) confirming delivery schedules.
- Check for a named construction contractor and a published first-silicon date.
Related reading
- What is Terafab?
- Nvidia's packaging and substrate deals
- Tesla Optimus production claims: what is verified
- Trump on data centers supplying their own power
- Nvidia DSX pod: claimed vs verified
- Meta Prometheus supercluster: what is confirmed
- Sources: Electrek: Tesla, SpaceX confirm Terafab site · Tom's Hardware on the Q3 2026 call · Tom's Hardware on Terafab's first phase
Earnings-call remarks are relayed secondhand; check primary filings and transcripts before relying on any figure.