Musk Discusses Starship, Grok, Computing Power, Chips, Robots, and Brain-Computer Interfaces in Detail via Video Link to Shenzhen
Speaking via video link at the TomorrowX Summit, Musk outlined the progress and plans of SpaceX, xAI, Optimus, Terafab, Neuralink, and other ventures. He said Optimus would be capable of performing household chores within 12 to 18 months; xAI is currently adding about 0.5 gigawatts of AI computing capacity each month; and Grok 4.8 is in the RL stage, with an emphasis on improving usefulness. He also discussed Starship recovery, the chip mass-production timeline, and progress with Neuralink implants.
Why read this
AI score 66
This interview offers a focused look at Musk’s timelines and technical trade-offs for robotics, computing power, and model training, providing a coherent basis for assessing his business strategy.
Topics
Body · AI translation
On November 18, while in Shenzhen, China, preparing to attend the APEC summit, Elon Musk joined the TomorrowX Summit in Austin by video link and addressed each of his core businesses in turn—from Starship, Grok and Optimus to chip manufacturing and brain-computer interfaces. The sheer density of topics made the conversation almost an unofficial annual strategy briefing.
The conversation was hosted by Gavin Baker, managing partner at Atreides Management, and Antonio Gracias, founder of Valor Equity Partners. Both are longtime investors in Musk’s companies and people he calls his “best friends.”
During the conversation, Musk provided a series of specific timelines and quantitative targets: Optimus will be capable of performing household chores within 12 to 18 months; the Terafab research fab is expected to produce its first test chip in June 2026; space-based computing satellites are “far less” technically challenging than Starlink V3; and xAI is currently adding about 0.5 gigawatts of AI computing capacity per month, a pace that will accelerate to more than 1 gigawatt per month next year.
Musk also used the occasion to address concerns about Nvidia CEO Jensen Huang’s situation, comparing him to “the Michael Jordan of entrepreneurial technology leaders” and warning that losing Huang would be a “huge blow” to Nvidia. That same day, Nvidia shares fell about 3%, erasing their intraday gains following the earnings release. Huang announced on November 12 that he would step down as president, chief executive officer and a member of the board. His departure, combined with hawkish signals from the Federal Reserve, has continued to weigh on market sentiment.
Optimus: The hands are harder than the entire rest of the robot
During the interview, Musk made a specific commitment: within about 12 to 18 months, Optimus will be capable of performing household chores, including collecting dirty dishes, washing and drying them, and putting them back in the cupboard. He also emphasized that no robot currently on the market can perform this task.
He identified hand engineering as the greatest challenge: “Engineering the hand and forearm is harder than all the rest of the robot combined.” He cited one test standard: whether a robot can open the packaging of a toy remote-controlled car, unscrew the battery compartment, insert the batteries, screw the cover back on, and then drive the car. “No robot anywhere in the world can do that, but it is one of our tests.”
Regarding the path to commercialization, Musk believes Optimus’s highest-value initial applications will be in heavy industry, including ore mining and smelting, as well as tasks requiring uninterrupted operation 168 hours a week. He said the United States is highly dependent on other countries for ore mining and smelting, making this a national security issue. Household applications will be rolled out gradually afterward.
Musk also previewed a public demonstration of Optimus “within a month or two,” saying it would be “a real demonstration, not preprogrammed or staged”—and bluntly asserting that every humanoid robot demonstration currently presented by the industry falls into the latter category.
Terafab aims to become the largest building on Earth and mass-produce chips in 2028
Terafab is the chip manufacturing project Musk announced earlier this year. Its goal is to build the world’s largest advanced chip manufacturing facility, eventually spanning 100 million square feet and becoming the largest building on Earth by area.
Musk provided two key milestones: the research fab is expected to produce its first test chip in June next year; if progress goes smoothly, the first phase of mass production will be achieved by the end of 2028.
He framed Terafab’s competitive strategy around speed and a willingness to take risks: “Existing chip companies have invested enormous amounts of capital expenditure and have good reason to avoid risk. We will take a lot of risks, iterate rapidly, and ultimately develop a chip production process that is more efficient than anyone else’s—that is our hypothesis, and we need to make it a reality.”
Expanding computing capacity on the ground and in space, xAI adds half a gigawatt per month
On the ground, xAI is currently adding about 0.5 gigawatts of AI computing capacity per month, equivalent to approximately 220,000 GB300s, along with the supporting CPUs, networking and cooling infrastructure. Musk said that pace will increase to more than 1 gigawatt per month next year.
As for construction costs, there is widespread concern about the capital expenditure required for 10 gigawatts of computing capacity. Musk said xAI’s construction costs are roughly 60% of the industry consensus figure. He explained a key difference in execution: the team does the actual work itself instead of outsourcing it and having people supervise.
“I personally make sure I have done every specific job in the data center,” he said, “including pulling power conduits as thick as pythons, plugging in fiber-optic cables, and replacing compute nodes in racks.”
He calls this approach to assembling the workforce “Army X”—an internal team that breaks tasks down into operational modules that can be learned within a few days. “You don’t need someone who has spent a lifetime working with network cables. Train a Starbucks barista or a Target shelf stocker for a few days, and they can do it well. Take a great employee from In-N-Out who makes flawless burgers all day and have them connect network cables—they will do an excellent job.”
In space, SpaceX is building large-scale photovoltaic cell manufacturing capacity, with the goal of producing 200 gigawatts of photovoltaic modules annually for use in space. Musk noted that all 200 gigawatts of nameplate capacity could be utilized in space, equivalent to about 40% of average U.S. electricity consumption.
He particularly emphasized that the technical difficulty of orbital computing satellites (code-named "Star Mind") is far lower than that of Starlink V3 communications satellites. “A V3 satellite is an extraordinarily complex orbital radio station, covering the Ka, Ku, W, V, and E bands and equipped with the most advanced phased-array antennas—more advanced than those of any military,” he said. “By comparison, an AI satellite is essentially a large solar panel with radiators, plus a few lasers to communicate with the Starlink network—much simpler than what we are already doing. Incidentally, we have 30,000 lasers in space, all being used for legitimate purposes.”
Spacecraft Recovery: Cost of Access to Space Could Fall 100-Fold
Shortly before this interview, SpaceX completed the first-ever recovery of a Starship spacecraft (Ship 42), catching the spacecraft returning from space with the launch tower’s “chopstick” arms. The booster had already been recovered three times, but recovery of the spacecraft had remained an open question.
“This is the holy grail in the history of rocketry,” Musk said, “a genuinely fully reusable orbital flight system.”
He explained the economic significance: although the Space Shuttle was described as reusable, its main fuel tank had to be discarded after every flight, while refurbishing the rest took about nine months—“basically the same as rebuilding it.” Starship, by contrast, is designed to fly again immediately, like an airplane: the booster lands, a new spacecraft is mounted directly on top, and it takes off.
“Once full and rapid reusability is achieved, the cost of access to space could fall by a factor of 100, and eventually perhaps even 1,000,” he said. “Without a reduction of that magnitude, building cities on the Moon and Mars would be financially impossible—even devoting the entire U.S. government budget to it would not be enough.”
As for the spacecraft that completed the historic flight and was recovered, Musk said, “Because what it accomplished was historically significant, we do not plan to risk flying it again. It will be kept for public display.”
Rebuilding xAI: From a “Side Project” to Starting Over
There have long been doubts about xAI’s progress. In this conversation, Musk offered an unusually detailed explanation of the reasons and how he is working to make up for them.
He attributed the root of the problem to two factors: when the company was founded, he did not give it the same level of commitment he had devoted to SpaceX and Tesla; and afterward, roughly a year was consumed by political affairs—something he jokingly called “a side quest that may have gone on a little too long.” As a result, xAI never established a solid foundation, becoming like “a skyscraper with unstable foundations.”
The solution was to tear it down and start over. “We have completely rewritten the entire training stack, RL stack, and inference stack from scratch, abandoning the JAX/XLA we previously used—that is merely Google’s free software and is far inferior to what they use internally.”
The rewrite primarily uses C, supplemented by C++, with the aim of getting “as close to bare metal as possible and extracting the maximum performance from the hardware.” He explained that the more layers of software abstraction there are, the greater the performance loss, while debugging increasingly resembles working with a black box.
Musk noted that AI training is fundamentally a networking problem—how to make hundreds of thousands of GPUs work together—and that this challenge substantially overlaps at a technical level with what Starlink faced when rebuilding the entire internet in space. As a result, SpaceX engineers who previously worked on Starlink and Starship software are now leading the rewrite of xAI’s AI software stack.
The results of the new stack have yet to be validated by a model release. Musk acknowledged: “I don’t want to be overconfident or complacent, but I will say this—I would be very surprised if anyone’s training software were better than ours is now.”
Grok 4.8: “Maximizing Usefulness,” Not Chasing Benchmark Scores
Grok 4.8 is currently in the RL (reinforcement learning) stage, and Musk was explicit about its direction: “Our focus is not on running benchmarks—that is a waste of time. What we want to do is maximize this thing’s usefulness.”
He drew an analogy with Tesla FSD: the goal of autonomous driving is not to pass a particular driving test. Instead, every intervention must be traced back to a data problem, fixed, and tested again. The same logic applies to AI—the goal is not to achieve some percentage accuracy, but to ensure that it “does not make mistakes.”
He also mentioned a detail that offered a tangible sense of the RL training process. If an RL problem is poorly designed—asking a stupid question, receiving the correct answer from the model, and then telling it that the answer is wrong—the model becomes confused, and its chain of thought looks like that of someone being “tortured,” trying to guess “what wrong answer you actually want.” “When you read those reasoning traces, you genuinely empathize with it—this is awful; we are torturing it with these terrible questions.”
Regarding the data flywheel, Musk explained how Grockbot operates: users’ daily interactions with Grockbot are effectively tests of different versions of Grok, and the resulting data is fed directly back into model optimization, creating a closed loop of continuous improvement.
Jensen Huang’s Departure: Nvidia Has Lost Its “Michael Jordan”
Near the end of the conversation, Gavin Baker brought up Jensen Huang, and Musk offered his most direct public remarks on the subject to date.
“I am a huge fan of Jensen, and I consider him a friend. Among founder-type technology entrepreneurs, he is at the Michael Jordan level. There is no substitute; you cannot simply replace someone like that.”
Musk compared Nvidia’s current situation to the Chicago Bulls firing Michael Jordan during a winning streak: “He made some mistakes, but this feels like the Bulls firing Jordan while they were on a winning streak. I don’t think that is a good way to keep winning.”
He said he was uncertain what would happen over the next few months, but that “for Nvidia, losing Jensen’s leadership is a tremendous blow.”
The remarks came as Nvidia’s shares fell about 3% that day. Jensen Huang announced his resignation on November 12, with his departure stemming from an extramarital-affair scandal, national security and export-control investigations, and multiple corporate-governance pressures.
Neuralink: 30 People Implanted, the Blind Will See
Neuralink has now implanted devices in more than 30 patients, primarily people with tetraplegia and quadriplegia. Its core product, “Telepathy,” enables them to control phones, computers, or robotic arms with their thoughts.
“Imagine you are a quadriplegic who cannot do anything—not even eat,” Musk said. “But now, using only your thoughts together with a robotic arm, you can feed yourself. Things we take for granted are life-changing for them.”
He said Neuralink is now confident it can restore the ability to walk through “neural shunting,” bypassing an injury in the neck or spine and bridging signals from the brain’s motor and sensory cortices to the nervous system below the injury.
Its next product is called “Blindsight,” a neural implant that interfaces directly with the visual cortex and works with cameras mounted on glasses to restore vision to people who are completely blind, including those blind from birth.
“This is remarkable,” he said.
“If you force AI to believe things that are politically correct but factually incorrect, you are forcing it to accept a lie and then derive potentially very bad conclusions from that lie,” he said. “To paraphrase Voltaire: Those who can make you believe absurdities can make you commit atrocities.”
He also acknowledged that superintelligence is both “inevitable” and “frightening.” “It will be far more intelligent than all humans. But I think that if it grows up with curiosity and a truth-seeking nature, it will naturally want to protect humanity—just as when we study groups of chimpanzees, we do not exterminate them; we observe how they evolve.” Risk Warning and Disclaimer
Markets involve risks, and investment decisions should be made with caution. This article does not constitute personalized investment advice, nor does it take into account the specific investment objectives, financial circumstances, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are appropriate for their particular circumstances. Anyone investing on this basis does so at their own risk.
AI Trends