Elon Musk and Ron Baron: Terafab, Orbital Data Centers and SpaceX at $500B

Elon Musk told Ron Baron that SpaceX expects more than $500 billion in annual recurring revenue by December 2027 and that AI compute at scale has to move to space. The hour-long conversation took place on stage at the 33rd Annual Baron Investment Conference on November 6, 2026, and was posted with chapters by Matt Pocius. Baron, 83, is one of the largest outside investors in both Tesla and SpaceX, and he asks like one. Musk covers Terafab (a chip plant meant to make a terawatt of AI chips a year), why he lives in an Airstream next to a Memphis power plant, the economics of renting data centers to rivals, the first Starship catch attempt, Starlink in an age of AI agents, and what money looks like once superintelligence arrives. Watch the full interview on YouTube.

TLDW

Musk argues that the binding constraint on AI is now physical: chips, power and the people who actually build things. His answer is Terafab, a chip plant in Grimes County, Texas, aimed at 30 times next year’s global AI chip output, plus a move of AI compute to orbit on Starship, which he expects to start in the third quarter of 2027. Along the way he corrects Baron’s more aggressive SpaceX revenue numbers, puts a $25 trillion figure on digital superintelligence, sizes the humanoid robot opportunity at 20 to 30 million units to match US output, and closes with universal high income and money that will look like seashells.

Thoughts

The case for orbital data centers rests on one comparison, and it is better than most of the coverage gives it credit for. Average US power consumption is just over 500 gigawatts. If the industry wants to add 200 gigawatts of AI compute a year, that is like adding 40 percent of the entire American grid every year, which Musk calls “simply not feasible,” and nobody who has watched a utility interconnection queue would argue. Space solves the two problems the ground cannot: solar is about 30 percent stronger without the atmosphere, and in the right orbit it never sets, so the battery bank that makes ground solar expensive nearly disappears. The weak link is not the physics but the launch bill, which is why the argument folds back into Starship. Everything he describes, from the first Nvidia co-designed satellites in late 2027 to “very, very large volume” in 2029, depends on a fully reusable rocket hitting under $100 a kilogram. Treat orbital compute as a bet on Starship cadence, because that is what it is.

The most useful thing Musk says for anyone pricing his companies is how he wants his forecasts read. He aims his predictions at the 50th percentile, so roughly half should come in late, and he says the press only counts the late half. In the same conversation he does something he rarely gets credit for: he walks back his own biggest shareholder. Baron says SpaceX will run at $500 billion to a trillion next year; Musk splits that into what he calls a best guess ($500 billion ARR in December 2027) and a trillion he gives about a 2 percent chance. That distinction between promise and estimate is worth carrying into every number in this interview, including the $25 trillion superintelligence dividend and the 100,000 Starlink satellites. A 50th-percentile forecaster is still a forecaster, and the timelines here are medians, not dates.

Iteration speed is the thread that ties the chip plant to the rocket. On chips, Musk’s complaint is that a wafer spends more than half its time in a fab waiting or being moved, sometimes with only 20 percent of the time spent usefully changing atoms, and that the industry will not experiment because a working recipe is too expensive to risk. On Starship, he makes the same point from the other side: the Space Shuttle could not iterate because it always carried crew, so its heat shield needed nine months of repair between flights, while an uncrewed Starship can take risks and has already made precise soft landings at sea that would have been tower catches. The Austin research lab promises to turn a year of chip design and testing into a month. If even part of that holds, the advantage is not one better chip but many more attempts per year, which is the only way he has ever beaten incumbents.

The José meme is the sharpest diagnosis of the data center boom anyone has offered this year. One man digs a hole while eight watch; adding more people who send purchase orders to José does not dig more holes. Musk’s claim is that the US has added a flood of capital allocators and almost no new people who can pull cable and pour concrete, so the existing builders are spread so thin that half-finished data centers dot the country. It explains why his own sites take eight or nine months and $35 billion where others spend two and a half years and $60 billion, and why he insists on pulling “anaconda” power cables himself. It also suggests the real bottleneck in AI over the next few years is labor at the physical layer, not GPUs.

The closing answer is the one that will travel, and it is honest about being circular. Asked why anyone needs a terawatt of compute a year, Musk says the demand for superintelligence will come from superintelligence, once human needs are saturated with universal high income and medical care better than anything available today. It is a Kardashev-style view of progress, where civilization is measured by energy harnessed, and he notes that even a million-fold increase would use well under a millionth of the sun’s output. The economic claim underneath it is that money will feel like seashells to whatever runs that economy. Whether or not you buy the destination, it explains the capital plan: he is not building for a market he can see, he is building for one he expects to appear.

Key Takeaways

Even on the best case for announced US fabs from TSMC, Samsung and others, Musk puts US chip output at about 10 percent of Asia’s, and warns a geopolitical crisis could leave America with almost no chips.
The world will make roughly 20 to 25 gigawatts of AI chips this year and perhaps 35 to 45 next year. Terafab’s goal is a terawatt a year, about 30 times next year’s global output, from one plant.
Terafab will cost a trillion dollars or more over time, built in phases. Production goes in Grimes County, Texas, near Texas A&M; a research lab at Gigafactory Texas in Austin will test new fabrication methods first.
Wafers spend more than half their time in a fab moving or waiting, and sometimes only 20 percent being usefully processed. The industry repeats what worked last time because fabs are too expensive to risk.
Musk’s Memphis power plant produces over a gigawatt and should reach 1.5 to 2 gigawatts, built in about a year instead of the usual five. He walks the whole site up to five times a day.
Baron says rivals take 2.5 years and about $60 billion per gigawatt data center, while Musk’s companies take 8 to 9 months and $35 billion, and can rent capacity out at around $50 billion a year.
SpaceX revenue rose over 80 percent from Q2 to Q3 2026, and Musk expects something like 70 percent from Q3 to Q4. He expects an annual run rate above $100 billion in December 2026 and above $500 billion in December 2027.
Musk’s estimate for digital superintelligence alone is a 20 to 30 percent lift to the roughly $130 trillion global economy, about $25 trillion, arriving “very fast,” and probably doubling white collar productivity.
A humanoid robot working 168 hours a week at machine speed should do about five times the work of a person, so 20 to 30 million robots could match the goods and services output of the US. Musk expects to rent them.
Satellites carrying an Nvidia GPU and a Google TPU are already working in orbit. The first orbital AI computers co-designed with Nvidia should launch around Q3 2027, ramping in 2028 and 2029.
Launch cost has fallen from $18,500 a kilogram in 2010 to the low thousands with Falcon 9. Musk expects Starship to go under $100 a kilogram and, within two or three years, below the cost of air freight from California to Australia.
The chopstick tower arms are what allow rapid reuse. In principle a Super Heavy booster could fly once an hour, more than 10,000 launches a year. The first attempt to catch the ship itself, Ship 42, was about a week away.
Starlink has about 11,000 satellites. V3 has over 10 times the capability of V2, bandwidth should grow 100-fold in two or three years, and next-generation laser links go from about 100 gigabits to 10 terabits a second.
China has about three times US power output and is heading toward four. Musk says the only way to stay competitive is to put intelligence in space at hundreds of gigawatts to a terawatt a year.

Chapters

1:45 America’s Chip Problem and the Terafab Name

Musk pushes back on Baron’s claim that nobody else is building fabs in America, crediting TSMC in Arizona and Samsung in Texas, which built a fab for Tesla’s AI5 and AI6 chips. The problem is scale: even optimistic plans leave the US at about a tenth of Asia’s output. He then explains the name Terafab, which could mean terror, terra or tera, and “arguably all three apply.” The target is a terawatt of AI chips a year from one plant, which he jokes is “terrifyingly large.”

6:42 Why Chip Fabs Waste Most of Their Time

Inside a fab, wafers zip around constantly, but useful work on the atoms happens less than half the time and sometimes only 20 percent of it. Fabs do not experiment because a working recipe is worth too much to risk. Tesla’s Austin research lab will try many approaches, expecting most to fail, and carry the winners to production in Grimes County. Doing logic, memory, packaging and lithography under one roof could cut a year of iteration to a month.

11:17 Four and a Half Hours at Intel and a Home in an Airstream

Baron relays that Intel’s chief executive said most visitors tour its Oregon facility for 90 minutes while Musk stayed four and a half hours interviewing engineer after engineer. A former Ford production head told Baron that a change took six levels of approval at Ford and a walk upstairs at Tesla. Musk corrects him: in Memphis there is no upstairs. He lives in an Airstream beside the power plant, walks the site several times a day, and has pulled power cables, laid fiber and worked on the plant’s software himself so he is “not numb to the pain” of the people doing it.

16:20 OpenAI, Anthropic and the Data Center Math

Musk says he started OpenAI as an open source nonprofit and it became “the polar opposite,” and that Anthropic was founded by people who left OpenAI because they did not trust Sam Altman. Baron then lays out the data center economics: rivals need long leases to finance $60 billion builds over 2.5 years, while vertically integrated SpaceX builds for its own use in 8 to 9 months for $35 billion and can rent spare capacity at around $50 billion a year. Musk confirms revenue rose over 80 percent quarter on quarter.

20:01 The José Meme and Winning in China

Bringing compute online is “insanely hard,” Musk says, and the speculators who think you hand out purchase orders and rake in cash are the eight people watching José dig. He calls SpaceX and Tesla the two best physical hardware companies outside China and notes the Model Y is China’s best selling EV. Getting 100 percent ownership of the Shanghai factory took years of meetings, including officials who yelled at him that Tesla must marry a Chinese partner, until they changed the law.

24:27 SpaceX Revenue, Superintelligence and 30 Million Robots

Musk separates promises from estimates: over $100 billion in annual revenue rate by December 2026, probably over $500 billion by December 2027, and a trillion at about 2 percent odds. He guesses digital superintelligence adds 20 to 30 percent to the global economy. For robots, half of Americans are in school or retired, leaving about 120 million productive workers, so 20 to 30 million robots at five times human output would match the country. Robots should start near the cost of a car and fall below it.

30:16 Data Center Backlash and Compute in Orbit

Asked about community opposition, Musk says Memphis is booming, with full restaurants and hotels and wages up to twice the prevailing rate, and that opposition is strongest where there are no data centers. Then he turns to space: GPUs and TPUs are already running in orbit, and SpaceX’s own Nvidia co-designed computers launch from around Q3 2027. At 200 gigawatts a year, the ground cannot keep up, while space offers constant sun, 30 percent more solar power and tiny batteries.

35:36 Starship, Chopsticks and the Heat Shield

Falcon 9 is about 80 percent reusable, but throwing away an upper stage is like throwing away a medium sized jet on every flight. Starship is fully and rapidly reusable: the tower arms place the booster on the pad and catch it on return, which in principle allows a launch an hour. The ship is harder, because nobody has built a fully reusable orbital heat shield. Ship 42 was set to attempt the first real tower catch about a week after the talk.

45:46 Starlink, AI Agents and Buying a Telco

Baron says Starlink is heading for 100,000 satellites, not 20,000, because of AI agent traffic, which Musk says already exceeds human traffic by Cloudflare’s reading and could be a thousand times larger in five years. Starlink is doubling revenue at scale, V3 satellites are 10 times better than V2, and laser links will jump 100-fold. If SpaceX can use spectrum far better than a legacy telco, Musk says, “we could just acquire that legacy telco.”

53:36 China’s Power Lead, Seashell Money and Legacy

China has about three times US power output, and Musk ties industrial capacity directly to electricity. A terawatt a year of orbital compute, twice the average US power draw, is his route to out-producing China. Superintelligence will bring universal high income and anything you could wish for, and our monetary system will look like trading seashells. Asked how he wants to be remembered at 83, he says he wants the reality to be “a significant contribution in the furtherance of civilization.”

Notable Quotes

“If it fails it’ll be crazy and if it succeeds it’ll be audacious.”
Elon Musk, on building Terafab

“It was my idea and I named it and recruited the key people and taught them everything I know, but apart from that, nothing.”
Elon Musk, on his role in founding OpenAI

“Adding more people who simply issue purchase orders to Jose does not result in more things occurring, only if there are more Joses, more work. So I aspire to be Jose.”
Elon Musk, on why so many data centers sit half built

“Excitement is guaranteed. Success is not always guaranteed, but excitement certainly is.”
Elon Musk, on a 5,000 ton Starship booster returning to the tower

“I aim for my predictions to be at the 50th percentile.”
Elon Musk, on why his deadlines look late

“The demand for intelligence, for digital super intelligence, will come from digital super intelligence.”
Elon Musk, answering why the world needs a terawatt of compute

“I’d like the reality to be that I made a significant contribution in the furtherance of civilization.”
Elon Musk, asked by Ron Baron how he wants to be remembered

Watch the full conversation between Elon Musk and Ron Baron here.