,

Elon Musk at TomorrowX: Half a Gigawatt a Month and the Starmind Bet

Elon Musk closed Day Two of the TomorrowX Summit in Austin by video from Shenzhen, covering Starship’s first ship catch, xAI’s rebuilt Grok training stack, half a gigawatt of new compute a month, and orbital AI satellites. He spoke on the Arena Stage with two of his closest friends, Gavin Baker of Atreides Management and Antonio Gracias of Valor Equity Partners, in a session the summit’s organizer iConnections posted as Elon Musk on Starship, Grok, Optimus and the Future of AI. It ran just under an hour and moved through nearly every company Musk runs, ending on AI safety and an unprompted defense of Jensen Huang.

TLDW

Musk calls the catch of Ship 42 on Flight 15 the final proof that every part of Starship can fly to space and come back reusable, which he says can cut the cost of reaching orbit a hundredfold. He says xAI had to be rebuilt after he neglected it, with the whole training, RL and inference stack rewritten in C and C++, and that Grok 4.8 is being tuned for usefulness rather than benchmarks. He puts Optimus doing household chores 12 to 18 months out, Terafab’s first test chip around June 2027, and terrestrial compute growth at half a gigawatt a month, rising past a gigawatt a month by the end of 2027. He closes on orbital compute, Neuralink’s Telepathy and Blindsight, and his view that a maximally truth-seeking AI is the core of AI safety.

Thoughts

The most revealing answer in the session is about money, and it is not really about money. Baker raised the Wall Street worry over how SpaceX will pay for 10 gigawatts of compute, and Musk said the outside estimates are off by a factor of two, putting his own cost at roughly 60% of the industry consensus. His explanation is a management philosophy: he calls today’s data center construction the “Jose meme,” one man digging a hole while eight watch, and says he personally does at least some of every task on site, from pulling anaconda-sized power conduit to swapping compute nodes. Then comes the part other builders should study. An in-house crew he calls Army X breaks the work into tasks that can be learned in a few days, so the company can hire former baristas, Target shelf stockers and In-N-Out cooks instead of lifelong cabling specialists. That is the Tesla production-line playbook applied to construction, and it is the real reason he can claim half a gigawatt a month (about 220,000 GB300s, measured at the power plant). The Megapack detail matters too: GPU training loads swing hard enough to break turbine shafts unless the power is smoothed. Whether the 60% figure holds is unverifiable from a stage interview, but the method behind it is concrete and copyable.

The orbital compute answer is the boldest technical claim, and Musk makes it by flipping the usual objection. Critics treat a space data center as an exotic engineering problem. Musk argues the Starmind satellite is simpler than the Starlink V3 satellites SpaceX already builds, because V3 is a flying radio station with phased arrays, lasers and half a dozen frequency bands, while Starmind is mostly solar panels, a double-sided radiator facing deep space, and a compute module about a quarter the size of a Vera Rubin NVL72 rack. His numbers are checkable physics: panels producing around 250 watts per square meter, a radiator shedding about 1,200 watts per square meter, so roughly five times as much panel area as radiator area. The ambition behind it is staggering, a target of 200 gigawatts a year of space solar manufacturing, which he puts at about 40% of average US power use. The honest gap is everything between the physics and a working fleet: launch cadence, failure rates, and servicing hardware you cannot visit. He concedes it is “not trivial,” and the whole plan rests on Starship’s tonnage to orbit.

The most human moment came from an unlikely place: reinforcement learning hygiene. Musk explained that if the RL questions are badly posed or the “correct” answers are wrong, the model gets punished for right answers, and its thinking traces start to read “like a human being tortured” as it tries to guess which wrong answer the grader wants. Set aside whether a model can suffer; the operational point is sharp. Much of what separates good models from mediocre ones is unglamorous data work, checking that every question is sensible and every answer correct, and Musk says plainly these tasks “are not hard, but they’re just a lot of work.” It ties straight into his line that benchmarking Grok 4.8 is “a fool’s errand.” His alternative is the Full Self-Driving method: trace every failure back to wrong, missing or skewed data, fix the data, retrain. Shipping test versions quietly under Grok Bot and learning from whether users’ problems got solved is that same flywheel pointed at a chatbot.

The safety answer is familiar Musk, but the framing is worth taking seriously. Forcing an AI to hold a politically convenient falsehood, he argues, means forcing it to reason from a lie, and conclusions built on a lie can be terrible. He paraphrased Voltaire, that those who believe absurdities can commit atrocities, and added that a curious, truth-seeking superintelligence would want humanity to flourish the way we study gorillas rather than wipe them out. That is a hopeful bet more than a proof, and he admits superintelligence being “vastly smarter than all humans” is scary as well as inevitable. Right after, asked about Jensen Huang, he compared Nvidia losing its CEO to the Chicago Bulls firing Michael Jordan mid-streak. Huang resigned as Nvidia’s CEO earlier in November after disclosing an affair, and the board has signaled it is not considering his return. Musk’s comment is a friend’s loyalty, not an analysis of the board’s reasons, but it is the most newsworthy line of the day.

On Optimus, the most credible part was the restraint. Musk said current humanoid demos are staged and pre-programmed, and that Tesla will only show Optimus when the demo is authentic, perhaps in a month or two. He borrowed John Carmack’s test of a useful robot: open a toy remote-control car, unscrew the battery cover, insert batteries, close it and drive it. No robot can do that today, he says. His forecast of 12 to 18 months for loading dishes and taking out the trash fits his history of aggressive timelines, so treat it accordingly. The more grounded claim is that the first real jobs will be heavy industry running 168 hours a week, especially mining and refining ore, which he frames as a national security gap for the United States.

Key Takeaways

  • Starship has now had three booster catches and its first ship catch, which Musk says proves every part of the system can fly to space and return in reflyable shape. Ship 42 will go on public display instead of flying again.
  • Full and rapid reusability should cut the cost of access to space about 100 times, maybe 1,000 times eventually, and reach a mass-to-orbit rate in the million-ton-per-year range.
  • Musk gives two reasons to go multiplanetary: a defensive one (backing up the biosphere) and an inspirational one (a future among the stars is worth looking forward to).
  • xAI began as a small side project for people who wanted to work on large models rather than Tesla’s real-world AI. Musk admits he did not build it with the rigor of SpaceX or Tesla, and calls his year in politics a side quest that ran too long.
  • xAI, now part of SpaceX, replaced JAX and SGLang with a training, RL and inference stack written mostly in C and some C++, over about six to eight months. Musk says he would be very surprised if anyone has better training software.
  • Large-scale training is a networking problem, so engineers who ran Starlink and Starship software now lead the AI stack rewrite.
  • Grok 4.8 is in its RL phase, and test versions ship quietly under Grok Bot so real user outcomes feed the data flywheel.
  • Optimus will get a public demo only when it is authentic. Musk says household chores like washing and putting away dishes are 12 to 18 months out, and the first jobs will be 24/7 heavy industry.
  • Cybercab has no steering wheel or pedals, and Tesla reports no serious injuries so far. Expansion is planned for other Texas cities, then Florida, with California approval hoped for early next year.
  • Terafab is planned to be the largest building on Earth at about 100 million square feet. Musk targets a first research fab test chip around June 2027 and phase one volume production toward the end of 2028.
  • Musk puts his compute build cost at about 60% of industry consensus, and says SpaceX is adding roughly half a gigawatt of AI compute a month, set to pass a gigawatt a month by the end of 2027.
  • SpaceX is aiming for 200 gigawatts a year of photovoltaic production for space, and says the Starmind satellite is simpler than a Starlink V3.
  • Neuralink has implanted more than 30 people. Blindsight, its next product, is designed to send camera signals straight to the visual cortex, potentially even for people blind from birth.
  • Musk argues a maximally truth-seeking, curious AI is the most important ingredient of AI safety, and that Neuralink’s long-term role is human symbiosis with digital superintelligence.

Chapters

0:00 Elon Joins From Shenzhen

Musk calls the session a best-friend buddy talk and opens with praise for China, which he urges people to visit. He expects geopolitical bumps over Taiwan but hopes for good long-term relations, and says he came partly to thank the Tesla Shanghai factory team.

2:12 Starship’s First Ship Catch and the Multiplanet Case

Musk explains why catching the upper stage with the tower’s “chopsticks” was the last major piece of full reusability. He compares Starship with the Space Shuttle, whose orbiter took about nine months to refurbish and whose external tank was thrown away every flight. Citing Milton Friedman’s pencil example, he argues a Moon or Mars city needs an entire industrial base shipped there, which only a hundredfold cost drop makes affordable.

11:24 Rebuilding xAI and the Grok Training Stack

Musk tells the founding story of xAI and admits it grew on weak foundations. Building a company is like building a skyscraper, he says, so the team rewrote everything close to the bare metal. Progress looked slow during the rewrite because it was, but he believes the new stack can carry an Empire State Building.

16:22 Networking, Data Quality and Usefulness Over Benchmarks

Training hundreds of thousands of GPUs coherently is a networking problem, which is why Starlink talent transfers. Musk walks through the less visible work: data filtering, tokenization, filling gaps in data distribution, and an RL harness whose questions and answers must be correct. Grok 4.8 is being optimized for usefulness the way Tesla optimizes FSD, by fixing the data behind every intervention.

21:59 Optimus, Authentic Demos and the First Jobs

Musk dismisses staged humanoid demos and says the hands and forearms are harder to engineer than the rest of the robot combined. Asked when Optimus can do the dishes, he says 12 to 18 months. He also pictures it as a tutor or babysitter and “more useful” than R2-D2 or C-3PO, but says mining and refining come first.

28:53 Cybercab, FSD and Terafab

Cybercab is rolling out carefully because there is no fallback control, and one accident could shut the program down. On Terafab, Musk says the team will start with standard equipment, push suppliers for custom machines, and sometimes build its own. He respects the chip industry but sees its risk aversion as an opening for a faster-iterating fab.

33:54 The Cost of 10 Gigawatts and Doing the Physical Work

Musk calls himself possibly the best at capex efficiency and explains the Army X approach in Memphis, where the site runs around the clock seven days a week. Tesla Megapacks smooth the wildly varying GPU loads that would otherwise break turbine shafts. He says no other organization is adding half a gigawatt of advanced AI compute a month.

41:06 Scaling Compute in Space With Starmind

Space solar can use its full nameplate power, which is why SpaceX wants 200 gigawatts a year of photovoltaic capacity. Musk notes SpaceX has more than 11,000 satellites in orbit and about 30,000 low-power lasers in space, and walks through the Starmind satellite’s panel, radiator and compute layout.

46:41 Neuralink Telepathy and Blindsight

Prompted by Gracias, Musk describes Telepathy users who control phones, computers and robot arms by thought, some feeding themselves for the first time. He expects to bridge signals past spinal damage so paralyzed people can walk, and describes Blindsight as the next product.

49:39 Truth-Seeking AI and Jensen Huang

Gracias frames the summit’s goal as “P(boom) greater than P(doom),” and Musk agrees a great future is more likely than a bad one. He makes his truth-seeking case for AI safety, then calls Jensen Huang a Michael Jordan-level founder and calls his departure from Nvidia a huge blow to the company.

Notable Quotes

“I think any self-respecting civilization has got to be a multiplanet civilization.”

Elon Musk, on why Starship matters

“You just can’t build world class AI using mediocre freeware to state the obvious.”

Elon Musk, on rewriting xAI’s training stack

“Our focus with Grok 4.8 is not trying to benchmark this thing because I think that’s a fool’s errand.”

Elon Musk, on usefulness over benchmarks

“If you get some aces out of In-N-Out who’ve been making epic burgers all day long, you know, perfectly, they’re going to do a damn good job of network cables.”

Elon Musk, on staffing data center construction

“I’m not saying it’s trivial, but it’s not as hard as what we’re already doing.”

Elon Musk, on the Starmind satellite versus Starlink V3

“You want to make sure that you have a maximally truth-seeking AI. This is the most important thing in my opinion for AI safety.”

Elon Musk, on AI safety

“They’re on such a winning streak and then they fire Michael Jordan, you know, halfway through their winning streak. I’m like, not a good way to keep winning in my opinion.”

Elon Musk, on Nvidia without Jensen Huang

Watch the full conversation here.

More From the TomorrowX Summit

Related Reading