Can Elon Musk Actually Deliver His $119 Billion Terafab Plan?

Kind: captions Language: en A hundred and nineteen billion dollars. That's the number in the titles of some of the biggest Terafab videos about Musk's chip factory in Texas. It's real. It's right there in the tax filings. It just isn't what the biggest of them say it is. Then there's the terawatt. Here's Musk's goal for Terafab. >> This is a terawatt of compute for your >> Put that next to the other number he's floated for the same factory and in our base case, depending on how you count it, they're up to about three times apart. And a question nobody seems to ask, as of today, has anyone seen a single chip made by Terafab? We went through the filings, the county contract, SpaceX's IPO prospectus, and Musk's own words. Quick guess before we start, is that 119 billion all for making chips? Hold on to your answer. Terafab is really two things. Part one is a research fab Tesla's building on its Giga Texas campus in Austin. Musk has put it at probably three billion dollars or so for maybe a few thousand wafers a month. Part two is the giant one, a campus in Grimes County being developed by a SpaceX subsidiary, Terafab AI. Intel joined in April with no terms disclosed. Where is it now? Tesla says construction and equipment purchasing are underway at its Austin research fab and local news reported land clearing at the Grimes County site in August. And that chip question? So far, no chip made by Terafab has been shown in public. SpaceX's IPO prospectus says it's in the very early stages with timelines and capital spending not yet determined and neither Tesla nor Intel obligated to stay. Receipt one, the money. Remember your guess? Some of the biggest Terafab videos call it a hundred and nineteen billion dollars on one single building. >> Let's open the paperwork. It comes from eight applications for a Texas school tax break called Jedi filed by Terafab AI, a SpaceX subsidiary. Together, they list $119.39 billion of estimated investment over up to four phases from 2026 to 2036. The filings give a range of 55 to 119 billion and say that after each phase, the company would decide whether to build the next. The $16.8 billion first phase announced in August, that's phase one of these same filings. So, the answer to the guess? No. About 72 billion is machinery and equipment, and about 47 billion is buildings and other improvements. The list of eligible property includes chip-making tools, yes, but also gas turbines, utility-scale batteries, water treatment, a space compute test facility, and the GPUs and AI accelerators themselves. The filings never say how much of it is the chip factory itself. So, what is SpaceX actually signed? The biggest commitment in the contracts we've read is the county tax deal signed in June. At least $5 billion invested by the end of 2030 and 1,800 jobs by 2035 with a default only below 90%. The agreement says nothing in it obliges the owner to build, though the tax break depends on building. And SpaceX can walk away on 30 days notice. 5 billion is about 4% of the headline. It's not a scandal. A tax filing estimates what you might build. A contract sets the least you've promised. The headlines pick the bigger number. Now, file this away. The most advanced chips are patterned with EUV lithography machines from ASML. And last year, ASML sold 48 of them to the whole planet. Keep that number in your head. And at the end, would owning a fab even pay off? In our model, one number swings it more than any other. >> Receipt two, the output. Musk has put two different numbers on what Terafab would make. The first came at Tesla's shareholder meeting in November 2025. >> It's it's got to be at least 100,000 wafer starts per month, size fab. And maybe that would be one of 10 in a complex. So, it'll ultimately be a million wafer starts per month. >> A wafer start is one silicon disk entering the factory. So, a million a month. Then at the launch came the terawatt, a year's output of computing hardware measured by the power it draws. Do those two numbers describe the same factory? Here's the check. Take Nvidia's Blackwell, its flagship AI chip of 2025. Each one uses two big dies of about 800 square millimeters. About 65 of those fit on a wafer. Lose some to defects, pair them up, and you get roughly 19 GPUs per wafer, around 27 kilowatts of computing power. Now, a million of those wafers a month for a year. Yes, more than a terawatt or less? Less. In our model, about a third of a terawatt, somewhere between 0.16 and 0.5, depending on yields and power per chip. To reach a full terawatt, you'd need about 3 million wafers a month, anywhere from two to six and a half million. And our model leaves out the memory wafers, which only widens the gap. To be fair, they could still line up if Musk's chips carry two to six times more power per wafer than Blackwell, or if the terawatt counts whole data centers, not just chips. And there's a clue. Here's his starting point. >> Uh uh the current output of AI compute uh is roughly 20 gigawatts per year. >> That's about right, but only if you count whole data centers. Count his terawatt the same way, and a million wafers comes much closer. He hasn't said which he means. Counted as chips at Blackwell's density, a million wafers and a terawatt are two different ambitions. >> Receive three, the machines. Let's take the million wafers at face value. That's about 71% of TSMC's entire 2025 capacity across all its process nodes. It's roughly the whole world's capacity at 7 nanometers and below. And it's about twice the world's forecast capacity for 2 nanometer class chips in 2028. Remember those 48 EUV machines? So, how many would a million wafers a month need? Take a guess. In our model, about 440. Somewhere between 250 and 860. At last year's 48, that's about 9 years of ASML's entire EUV output, or about 4 years at the more than 110 a year capacity it's reportedly exploring for 2028. Then, the bill. In our model, a million wafers a month costs 400 to 840 billion dollars. For scale, TSMC plans 60 to 64 billion of capital spending in all of 2026. And power. About 4.6 gigawatts around the clock in our base case. Anywhere from 2 to 10. That's about 1.6 times all the electricity TSMC used in 2024. And the Grimes County plans do call for gas-fired power plants and big battery arrays on site. >> Now, flip it around. If every dollar of the 119 billion went into chip-making capacity, our model says it buys about 140 to 300,000 wafers a month, roughly 200,000 in the base case. That's about two of Musk's own 100,000 wafer fabs, not 10, before you subtract the turbines, batteries, and GPUs. The binding 5 billion buys about 6 to 12 and 1/2 thousand. So, who is this really for? Mostly not Tesla's cars. Tesla's next AI chip, AI 5, isn't coming from Terafab. Musk said Samsung in Texas and TSMC in Arizona will make it. And Samsung has a separate deal for the next one, AI 6, worth about 16 and a half billion dollars through 2033. In our model, Tesla's own cars and near-term robots need roughly 1,600 to 20,000 wafers a month. In the base case, about half a percent of a million. Even Musk's hope of 10 million Optimus robots a year comes to around 10,000. So, why build it at all? Here's Musk, asked whether it's about getting better economics on chip purchases. >> The Terafab is is not some sort of mechanism for to generate leverage over our chip suppliers. Uh it's just literally we we don't see a path to having enough >> The launch, he put it more bluntly. >> So, we we either build Terafab or we don't have the chips. And uh we need the chips, so we're going to build Terafab. >> So, the scale fits two bigger dreams. Musk expects humanoid robots to reach 1 to 10 billion a year someday. In our model, a billion AI 5-size chips a year would need about a million wafers a month, roughly his number. The other dream is space. >> I think it's probably 100 to 200 gigawatts a year of terrestrial chips. Um and probably on the order of a terawatt of chips in space. Um just because of our strengths on the ground. >> Even so, SpaceX's prospectus says it expects to keep buying a significant portion of its compute hardware from outside suppliers. But would owning a fab at least be cheaper than buying wafers? The 2-nanometer wafer TSMC reportedly costs about $30,000, and that price includes TSMC's profit. Build your own, and you keep that margin. You pay for the building, the tools, and every wafer that comes out wrong. So, last guess. 100,000 wafer fab over 15 years. Does it beat buying? In our base case, the fab costs $56 billion. Counting capital and running costs, each wafer start costs about 19,000. But, a newcomer won't match TSMC's yields on day one. At 80% of TSMC's good chips, each useful wafer costs about 23 and 1/2 thousand. Cheaper than 30. But, after 3 years of building and 2 of ramping, 15 years nets out at roughly break even, about plus 1 and 1/2 billion, with payback around year 10. >> And the range is huge. In our optimistic case, with near TSMC yields, high loading, cheaper construction, and pricier wafers to buy, it's plus 151 billion. But, that's what a seasoned foundry achieves, not a Terafab forecast. In our stress case, with weak yields, high build costs, and cheaper wafers to buy, it's minus 90 billion. That one number? The build cost. In our model, over the ranges we tested, it swings the result most, ahead of the yield and the price of the wafers you'd otherwise buy. In our base case, if yields fall below about 63% of TSMC's, or the build cost tops about $780,000 per wafer of monthly capacity, owning loses. So, where would a first-timer land? Even TSMC, the most experienced foundry there is, took about 3 and 1/2 years from breaking ground in Arizona to volume production. Intel's Ohio fabs have slipped from 2025 to 2030 or 2031. Samsung's 2-nanometer yields have reportedly trailed TSMC's. And as chip analyst Dylan Patel put it before the launch, the really complex part is actually developing a process technology and building wafers. The fab only pays if a newcomer performs like the incumbents. >> So, is it impossible? Not necessarily. Here's what would have to be true. Chips with several times more power per wafer, or a terawatt that counts whole data centers. A proven process from a partner, whether Intel's 14A, which Musk called not yet totally complete, or TSMC, where Musk reportedly says there are just discussions so far. Far more EUV scanners from ASML, which is targeting at least 80 in 2027. Its own power, which the Grimes plan provides for, and yields close to the incumbents, fast. What to watch. The first wafers from the Austin research fab. Real construction in Grimes County, where SpaceX has reportedly talked about limited operations toward the end of 2027. Any EUV order. None has been reported for Terafab so far. And the two dates that matter. Whether SpaceX goes ahead with phase two, penciled in for 2029, and whether that 5 billion is in the ground by the end of 2030. So, what does a 119 billion dollars actually buy? On paper, up to four optional phases of fabs, power plants, and computers through 2036, with a 5 billion dollar floor in the county contract, and a 30-day exit. Even if every dollar went into fabs, that's about two of Musk's big ones, not 10. And a terawatt of chips at Blackwell's density would take about three times the million wafers he talked about. Terafab is a real ambition backed by real money. Right now, it's a framework with a very high ceiling. Check back in 2030.