On a random Thursday in August, Musk posted a rendering video on X and called it “the largest and most valuable building on Earth by far.” Then he added, almost as an afterthought, “And it will be stunningly beautiful.” No context, no thread, just the video and that line. Within an hour it had a few hundred thousand views and my group chat had already split into “this is insane” and “he’s done this before, remember the Hyperloop.”
I’ve been following this project since March, when Musk first teased it at the old Seaholm Power Plant site in Austin. Back then it was called a chip fab for Tesla and SpaceX. By August it had a name, a location, a price tag, and a video that looked like something out of a Denis Villeneuve movie. So let’s actually go through what Terafab is, because the headlines have gotten a bit ahead of the concrete.
One thing first, and I think it matters: this is not a datacenter. People keep calling it one, and I get why, since the whole point of it is to feed compute into datacenters. But Terafab itself is a chip factory. It’s the building that makes the chips that later go into datacenters, EVs, robots, and eventually orbit. That distinction changes what you should expect from it.
What Terafab actually is
Terafab is a joint venture between Tesla and SpaceX, run by Musk since he’s CEO of both. It’s planned for Grimes County, Texas, a rural county east of College Station, about 45 miles northwest of downtown Houston. The site sits on what used to be a reservoir.
The pitch is vertical integration. Instead of designing a chip and sending it off to TSMC or Samsung to fabricate, Terafab is supposed to handle the whole chain in one place: design, mask fabrication, wafer manufacturing, testing, and packaging. Musk called it “the most epic chip-building exercise in history.” He also called it that back in March about a smaller version of the same idea, so take the superlatives with a pinch of salt, if that makes sense.
The confirmed first-phase investment is $16.8 billion. That’s actually less than SpaceX’s own earlier estimate of $55 billion for phase one, which is a pretty big number to walk back without much explanation. A regulatory filing from SpaceX suggests the total across multiple phases could eventually reach $119 billion. Texas is chipping in too, reportedly around $30 billion in incentives, though the exact structure of that deal hasn’t been fully laid out yet.
The size, and why the Central Park comparison doesn’t quite hold
Once complete, Terafab is expected to span more than 100 million square feet. I went looking for a clean way to picture that and honestly the numbers do most of the work themselves. Sorting through it took longer than I expected too, since three different outlets picked three different buildings to compare it to, and the comparison itself changes depending on which source you trust. It would be roughly five times the floor area of the New Century Global Center in Chengdu, which currently holds the record for the world’s largest building (that one has an indoor beach and a ski slope inside it, for context). Terafab is projected to be bigger than the Pentagon, Apple Park, the Mall of America, and Tesla’s own Gigafactory Texas combined. Yahoo Finance put it at about 36 times the office footprint of the Empire State Building. Fast Company said you could fit six of the whole structure inside Manhattan.
So if someone tells you it’s “four or five times the length of Central Park,” that’s not quite the right comparison, Central Park is about 843 acres of park land, and Terafab is a building footprint measured in square feet. But the instinct behind the comparison isn’t wrong. This thing really is bigger than anything humans have built for a single purpose, at least on paper.
I say on paper because right now, as of this writing, it doesn’t exist. What’s out there is a rendering video, some groundbreaking activity, and a permit process. Fast Company put it more bluntly than I would: it’s “a megacool gigafuturistic terarender.” That’s a fair jab. Musk’s brand has always been the announcement first, the delivery somewhere down the line, sometimes years down the line.
Why build it at all
The size numbers feel like marketing, but the logic behind building this actually holds up. AI compute demand right now is outpacing supply, and by a wide margin. Everyone, from OpenAI to Google to Musk’s own xAI, is fighting for the same limited pool of advanced chips, mostly made by TSMC in Taiwan on machines built by ASML in the Netherlands. If you’re Musk and you’re running Tesla, SpaceX, and xAI at the same time, you’re competing against your own peers, and sometimes your own companies, for the same fab slots.
Terafab is meant to fix that by making Musk’s companies self-sufficient. The target is more than one terawatt of compute capacity a year. For scale, a terawatt is close to the entire power generation capacity of the United States, which tells you the number is only describing an aspiration, not a near-term output figure, if that makes sense.
The facility is planned to produce two different chip lines. One is edge inference silicon, meant for Tesla’s Optimus robots and its self-driving Cybercabs, chips that need to run AI models locally without much power draw. The other is a more powerful line, hardened for space environments, which ties directly into SpaceX’s plan for a constellation of orbital data centers. SpaceX filed for permission earlier this year to eventually launch up to a million data center satellites, with Musk claiming the network could provide hundreds of gigawatts of compute once built out. I won’t pretend I fully believe that number yet, a million satellites is a lot of satellites, but the direction of the bet is clear enough.
What this actually buys Musk
If I try to strip away the theater and just look at the business logic, Terafab is a supply chain move dressed up as a moonshot. Right now every AI company on Earth is dependent on TSMC and a handful of other fabs. That dependency is a real bottleneck, and it’s one Musk doesn’t control. Owning your own fab means you’re not standing in the same line as Nvidia’s other customers, you set your own priorities.
For Tesla, that means Optimus and Cybercab chips don’t have to wait behind whatever queue Nvidia or Qualcomm customers are in. For xAI, it means Grok’s training and inference compute isn’t entirely at the mercy of external chip allocation. For SpaceX, it means the orbital datacenter plan, which sounds like science fiction today, has an actual chip source lined up instead of being a plan with no supply behind it.
There’s also a jobs and political angle that’s worth naming plainly rather than dressing up. The facility is expected to employ at least 3,000 people, with 60 to 80 percent of those jobs going to residents of Grimes and neighboring Brazos counties. That’s landed at a slightly awkward moment, with data center backlash becoming a real issue heading into the 2026 election, so a local jobs number this large is doing some political work too, not just economic work.
One detail I liked, and it’s a small one: the facility is planned to draw water from the Gibbons Creek Reservoir instead of local groundwater, with on-site wastewater treatment and recycling built in. Chip fabs use enormous amounts of water, and given how much backlash data centers have already drawn over water use, this feels like a lesson learned in advance rather than an afterthought.
The honest catch
I’ll say this plainly: Musk has no prior experience running a semiconductor fab.
None.
Building EVs and spacecraft is hard. Advanced chip manufacturing is a different kind of hard on top of that, one of the most capital-intensive, failure-prone industries that exists. TSMC has been doing this for almost forty years and it still takes them years to bring a new node online cleanly.
ASML’s CEO has apparently confirmed direct talks with Musk’s team and said he’s “very serious” about the project, which is a decent signal, ASML doesn’t hand out that kind of comment lightly since their lithography machines are the actual bottleneck of the entire chip industry, not some vague throwaway compliment. But being serious about wanting something and actually delivering it are two different things. Given how often Musk’s own stated timelines have slipped at Tesla and SpaceX, I’d bet on the completion date moving before I’d bet on the terawatt figure holding.
The funding gap is the other thing nagging at me. Going from a floated $55 billion phase-one number down to a confirmed $16.8 billion is a big cut, and it’s not clear whether that reflects a smarter, leaner plan or just a company being more careful about what it commits to on paper before ground is fully broken. No firm completion date has been given either. When Musk first floated the idea back in November of last year, there was no timeline attached, and there still isn’t one that I’ve seen reported with any real confidence.
How it stacks up against actual chip fabs
It helps to compare Terafab against the fabs that already exist and actually ship chips today, rather than just against buildings. TSMC’s most advanced fab, Fab 18 in Tainan, is nowhere near 100 million square feet. It’s a fraction of that. And it still produces the bulk of the world’s leading-edge chips. Intel’s Ohio project, the one they call “Silicon Heartland,” is budgeted at over $20 billion for two fabs and has already slipped its original timeline by a couple of years. Samsung’s Taylor, Texas fab, not too far from where Terafab is going up, took roughly three years longer than planned and cost billions more than the original estimate.
I bring this up because it’s the closest thing to a real baseline. Every advanced fab built in the last decade, by companies with decades of fab experience, has run over budget and past schedule. Terafab is being built by two companies with zero fab experience between them, at a scale nobody has attempted before. That doesn’t mean it can’t work. It means the base rate for “on time and on budget” in this specific industry is already bad before you add in a scale nobody has tried.
There’s also the question of yield. A fab this size existing is one thing. A fab this size actually producing usable chips at a defect rate low enough to be worth the investment is a separate and much harder problem. TSMC’s edge over the rest of the industry isn’t really about square footage, it’s decades of process knowledge about exactly how to keep yields high at advanced nodes. That’s not something you buy with a construction budget, it’s something you earn through years of failed batches and incremental fixes. Terafab’s website talks about consolidating logic, memory, packaging, and testing under one roof to fix supply chain problems, and that’s a real problem worth fixing, but it says nothing about whether the actual chips coming off the line will be competitive on yield with what TSMC already ships.
What happens if it slips
Say the timeline stretches, which given the industry’s track record is a reasonable thing to plan for rather than a cynical guess. What does that actually cost Musk’s companies?
Optimus probably stays a smaller-scale product for longer, since a lot of its cost curve depends on cheap, purpose-built edge chips rather than repurposed GPU silicon. xAI keeps leasing or buying Nvidia and other third-party chips at whatever price the market sets, and that market isn’t getting cheaper. The one that worries me most is SpaceX’s orbital datacenter plan, since the space-hardened chip line was the one piece of actual hardware supply behind that whole satellite constellation filing. Without it, that plan stays theoretical for even longer than its already vague timeline suggests.
None of that is catastrophic on its own. Tesla, SpaceX, and xAI all have other paths to compute even without Terafab, they’d just be more expensive and less differentiated ones. But it does mean the “vertical integration solves everything” pitch is really a bet on execution speed as much as it’s a bet on the concept itself. The concept, building your own chip supply so you’re not stuck behind everyone else in line, is sound. Whether these two companies can actually pull it off is going to show up first in supplier contracts and hiring numbers, long before it shows up in any drone flyover video.
Where it stands right now
As of this month, Terafab is a confirmed site, a confirmed first-phase budget, and a rendering video. Groundbreaking has started on a smaller precursor facility at Tesla’s Giga Texas North Campus, which the companies describe as a testbed for the real thing. The main Grimes County site itself is still mostly permits and dirt work.
So is it going to be the biggest building on Earth, and will it actually reroute the AI chip supply chain around Musk’s own companies? Maybe. The engineering logic behind wanting it is sound, chip supply really is the chokepoint everyone in AI is fighting over right now. But “sound logic” and “built on schedule for the stated budget” are two very different bars, and Musk’s own history says the second one is where these projects usually slip.
I’d watch the Giga Texas precursor facility over the next year, that’s the one that’ll actually tell you whether Terafab proper is realistic or just another tera-render.