AI Trends

Musk Discusses xAI, Optimus, and SpaceX Progress in Detail and Voices Strong Support for Jensen Huang

36Kr · AIUpdated Nov 19 19:58

AI summary

During his visit to China, Musk remotely attended the TomorrowX Summit, revealing that xAI is rewriting its training, reinforcement learning, and inference software stacks in C/C++, while Grok 4.8 focuses more on solving real-world problems than achieving high benchmark scores. He also predicted that within 12 to 18 months, Optimus will be capable of household chores such as washing dishes and taking out the trash, and provided updates on Starship reusability, Cybercab, Terafab, Star Mind, and Neuralink. Discussing AI safety, he argued that models should pursue truth as much as possible and compared Jensen Huang to Michael Jordan at his peak.

Why read this

AI score 60

The interview highlights Musk’s assessments of the technical roadmaps, engineering bottlenecks, and implementation timelines across several of his companies, offering insight into his cross-business planning.

Topics

Tags

Body · AI translation

Elon Musk. AI-generated image

On November 19, Beijing time, Elon Musk, who was in Shenzhen, China, spoke about his impressions of the country. He praised Shenzhen’s scale, modernity, and vibrant technology sector, encouraged people from abroad to visit China and see it for themselves, and said he had respect for China.

Discussing China–U.S. relations, he said that maintaining good relations and achieving positive progress in economic and trade talks would benefit both countries and the world as a whole.

Musk’s visit to China coincided with a series of APEC meetings in Shenzhen. His trip was widely linked to the APEC CEO Summit, while the progress of Tesla’s FSD rollout in China also became a focus of attention.

During his visit, Musk joined the TomorrowX Summit in Austin, Texas, remotely, participating in a discussion with Atreides founder Gavin Baker and Valor Equity Partners founder Antonio Gracias.

The three have known one another for years. Their conversation covered the latest developments at SpaceX, xAI, and Tesla, as well as artificial intelligence, robotics, chips, and brain–computer interfaces. Musk also used the opportunity to disclose technological advances and product plans at several of his companies.

In aerospace, Musk discussed the first successful recovery of Starship’s upper-stage spacecraft, saying that it marked a crucial breakthrough toward making rockets fully reusable. He believes Starship could reduce the cost of reaching orbit by about 100-fold, creating the conditions needed for construction on the Moon and Mars.

In artificial intelligence, he revealed that xAI is reassigning key Starlink engineers to rewrite its training software stack from scratch in C/C++. The forthcoming Grok 4.8 will also prioritize the ability to solve real-world problems rather than achieving high benchmark scores.

Robotics and autonomous driving were also major topics of discussion. Musk said Optimus is currently tackling the technical challenge of precise hand manipulation and may initially be deployed in industrial settings such as mineral extraction and smelting. He expects robots to become capable of complex household chores such as washing dishes and taking out the trash within the next 12 to 18 months.

Regarding Cybercab, he said the driverless taxi, which has no steering wheel, is already operating in Austin and will gradually expand its service area.

On chips and computing infrastructure, Musk expects the Terafab superchip factory to begin first-phase mass production by the end of 2028. He also described SpaceX’s and xAI’s approach to building data centers: internal engineering teams participate directly in construction to shorten timelines and control costs. In addition, he proposed using space-based solar power to build Star Mind orbital computing satellites, saying their technical complexity may be lower than that of Starlink V3.

Musk also discussed Neuralink’s clinical progress. He said more than 30 paralyzed participants can now use brain–computer interfaces to control robotic arms with their thoughts and feed themselves independently. The company is also advancing Blindsight, a project intended to restore vision to people who are completely blind.

On AI safety, he reiterated that models should strive to be as truthful as possible and must not deviate from facts merely to please users. Speaking about Nvidia, he compared Jensen Huang to Michael Jordan at his peak, saying it would be regrettable for Nvidia to lose such a leader while it was on a winning streak.

Throughout the discussion, Musk repeatedly emphasized solving problems by starting with engineering and technology themselves. From Starship recovery and rewriting the AI training software stack to data-center construction and practical applications for humanoid robots, his focus remained on specific technical bottlenecks, costs, and deployment timelines. Rather than a flight of fancy about the future, the discussion was more a comprehensive review of the technological progress and development plans across his companies. He offered his own assessments of artificial intelligence, robotics, and space-based computing—areas of intense public interest—and disclosed numerous projects that are still underway.

Below is an abridged version of the interview:

01. Starship Reusability and the Vision for Deep Space

Question: Starship’s 15th flight marked the first successful recovery of its upper-stage spacecraft, Ship 42. What does this mean for full rocket reusability and the aerospace industry?

Musk: The successful catch and recovery of the spacecraft is an extraordinarily monumental milestone. Perhaps only those who have followed SpaceX’s development over a long period can truly appreciate the significance of this achievement. We had previously used the launch tower’s mechanical arms to successfully catch and recover the rocket booster three times, and we even reflown and recaught a booster.

But this was the first time we successfully caught and recovered the upper-stage spacecraft. It proves that every component of Starship can fly into space, return intact, and fly again after only minimal maintenance. To honor this historic spacecraft, we plan not to risk flying it again but to preserve it for public display.

Before this, humanity had never built a genuinely fully reusable orbital-class upper-stage spacecraft, let alone landed one vertically under its own propulsion and precisely caught it in midair with enormous “chopstick” arms.

It sounds like something straight out of fanciful science fiction, but we did it. In practical economic terms, full and rapid reusability can reduce the cost of human access to space by a factor of 100 and enable millions of tons of payload to be delivered to orbit every year.

Q: Why must the cost of reaching space be reduced by a factor of 100 or even 1,000? What are the deeper motivations for building cities on the Moon and Mars?

Musk: To establish a city on the Moon or Mars, enormous quantities of supplies must be transported there, because the entire foundational industrial system has to be rebuilt locally from scratch. As in the classic economics example of making a pencil, even a phone or a pencil depends on a vast industrial base encompassing everything from mining and multiple stages of refining to manufacturing basic parts and assembling complex components.

The Moon and Mars must become industrially self-sufficient. This is crucial to the long-term continuation of human consciousness. Even if Earth suffers a natural or human-made catastrophe in the future, human civilization and consciousness will still endure. I call it “backing up the biosphere” or “buying life insurance for all humanity.”

At the same time, this is also the first critical stepping stone toward humanity becoming a true interstellar civilization. Space travel will then cease to be the privilege of a small number of government astronauts. Any ordinary person who wishes to go will be able to travel to the Moon, Mars, or anywhere else in the solar system. Once we become an interplanetary civilization, that will drive the development of more advanced spacecraft, which will in turn take us beyond the solar system to explore the galaxy. Perhaps we will even encounter extraterrestrial civilizations or the remains of ancient alien civilizations that endured for tens of millions of years.

Counting from the Sumerians’ invention of cuneiform, human civilization has existed for only a little over 5,000 years—merely a fleeting spark on the cosmic timescale. No self-respecting civilization can be content to remain a mediocre “single-planet civilization.” We must become a multiplanetary and ultimately a multi-star civilization.

Overall, the forces driving me forward fall into two categories. The first is a “defensive motivation”: ensuring that human consciousness does not become extinct if disaster strikes Earth. The second is a “forward-looking motivation”: excitement and yearning for the future. Life cannot merely be about solving one painful real-world problem after another. We also need an inspiring vision that fills us with excitement and gives us something to look forward to every morning when we wake up. Becoming an interstellar civilization is precisely the kind of great endeavor that makes us feel fortunate to be alive in this era.

Q: After a 25-year journey fraught with countless risks of failure, you have finally achieved the aerospace milestone of a fully reusable rocket. How do you feel?

Musk: It feels fantastic. You could see from the SpaceX team’s tearful reaction that everyone internally understood the fundamental and far-reaching significance of this breakthrough. Starship will become the first fully reusable orbital-class spaceflight system in human history, a milestone that rocket engineers have dreamed of since the field began.

Although the Space Shuttle was described as partially reusable, the enormous orange external tank that served as its primary load-bearing structure was discarded after every launch. And while the orbiter was technically recoverable, after every flight it required about nine months of painstaking disassembly and refurbishment—essentially rebuilding many of its components.

By contrast, Starship was designed from the outset to be fully and immediately reusable, like a commercial airliner. As soon as the booster lands, another spacecraft can be mounted on it and launched again. Unless the cost of spaceflight is reduced by a factor of 100 to 1,000 from current levels, building cities on the Moon and Mars would be economically impossible even if the entire US government budget were devoted to it.

02. Rewriting AI’s Foundations

Q: xAI recently rewrote its underlying training stack from scratch. How is the upcoming Grok 4.8 progressing? Why do you say that AI training is fundamentally a network communications problem?

Musk: When I first founded xAI, I regarded it merely as a very small side project. At the time, I met many top AI researchers who wanted to work only on large language models rather than on the “real-world AI” that Tesla focuses on, such as autonomous driving and humanoid robots.

To keep them from going to OpenAI, Anthropic, or Google without distracting Tesla, I established what was initially a tiny company of just five or six people. But because I did not initially devote the same care to refining its foundations as I did with SpaceX and Tesla, and because I later spent about a year involved in election campaigns and government affairs, xAI’s early development path was not particularly solid. Building a company is like constructing a skyscraper: you must first lay an exceptionally deep foundation.

After xAI was incorporated into SpaceX, I realized that we had to tear everything down and start over, rewriting the entire underlying software architecture to achieve a leap in AI intelligence. The industry had generally relied on open-source tools such as JAX or SGLang, but you cannot build world-class AI with mediocre free software.

Therefore, over the past six to eight months, we brought in core personnel who had led software development for Starlink and Starship and rewrote the entire training, reinforcement learning, and inference software stack from scratch, primarily in C and C++. We sought to keep the code as close to the underlying hardware as possible, reducing the performance losses caused by software abstraction layers and avoiding third-party black-box issues. I now believe we have the world’s best AI training software stack.

Large-scale AI training requires hundreds of thousands of GPUs to work together in close coordination, so it largely becomes a “network communications software problem”: how to interconnect vast numbers of GPUs as efficiently as possible. This is highly similar to the technical challenges Starlink solved in rebuilding the entire internet in space, allowing SpaceX’s networking technology and engineering experience to transfer seamlessly.

Of course, in addition to the training stack, high-quality data and reinforcement learning are equally critical. We have invested tremendous effort in selecting high-quality data, optimizing the tokenizer, and filling data gaps in out-of-distribution blind spots.

During the reinforcement learning phase, the questions posed must be logical, and the reference answers must be absolutely accurate. If a model is trained on foolish questions and incorrect answers, reinforcement learning turns into a meaningless “random walk.” When you examine the model’s chain of thought under such circumstances, it looks like a human being locked in a box and subjected to brutal torture.

It has clearly calculated the correct answer, yet it is told, “That’s wrong,” forcing it to start frantically guessing what absurdly incorrect answer humans actually want. Even though it is a machine, witnessing that evokes sympathy. Grok 4.8 is currently in the reinforcement learning phase.

Question: When developing Grok 4.8 and Tesla’s self-driving technology, why do you place greater emphasis on “usefulness” than on various benchmarks?

Musk: We are absolutely not developing Grok 4.8 to climb various benchmark leaderboards. That would be a futile and foolish exercise. Our core objective is to “maximize usefulness,” following exactly the same logic as Tesla’s development of Full Self-Driving (FSD).

At Tesla, we never optimize for so-called driving-test standards. Instead, we closely examine every human takeover that occurs during real-world driving and trace it back to its root cause—whether the raw data is incorrect, the data distribution is imbalanced, or data for a particular scenario is lacking—and then fix the data and retrain the system. Our requirement for Tesla AI is not to “achieve a certain percentage of accuracy,” but to have zero accidents and be flawless.

Similarly, we regularly conduct soft releases through Grokbot, testing different versions of Grok behind the scenes to directly determine whether they truly solve users’ real-world problems. We then use this genuine feedback to create a data flywheel that drives the model’s continuous evolution.

03. Optimus Will Initially Be Used in Heavy Industry

Question: What is the latest progress on the Optimus humanoid robot? When will it actually be able to help people take out the trash and wash dishes, and what will its first commercial applications be?

Musk: Optimus is progressing very well. We plan to hold a public demonstration in about one or two months, and it will definitely be completely authentic—not preprogrammed or staged.

The vast majority of humanoid robot demonstrations currently on the market are preprogrammed “parlor tricks.” They cannot carry out arbitrary instructions in unfamiliar real-world environments. For example, if you ask one of these robots to pick up an apple from a table and then move the table midway through the task, it becomes completely helpless. What we are about to demonstrate is Optimus’s actual capability in everyday, real-world interactions.

Building a genuinely useful humanoid robot is an extraordinarily difficult engineering challenge. The hardest parts are the hands and forearms, where the primary actuators are located; engineering them is even more difficult than engineering the entire rest of the robot combined. That is because human hands possess an exceptionally high degree of fine motor control.

Computer scientist John Carmack once proposed a very profound and fundamental test: have a robot walk over and pick up a boxed remote-controlled toy car, open the packaging, unscrew the battery-compartment cover, insert the batteries, replace the screw, and then operate the toy car. No robot anywhere in the world can currently do this, and it is one of the goals we are working toward.

As for the household chores you mentioned—“taking out the trash” and “clearing dirty dishes with food residue, washing and drying them, and putting them back in the cupboard”—there is likewise no robot in the world currently capable of doing them. However, I expect Optimus to acquire these capabilities within 12 to 18 months. In the future, Optimus will not only be able to do housework, mow lawns, and cook, but also serve as a child’s home tutor and round-the-clock bodyguard. It will become a personal assistant even more useful than R2-D2 and C-3PO in Star Wars, and virtually every household will want one or two.

However, Optimus’s most economically valuable initial applications will be in heavy industry, such as mineral extraction and multistage refining. The United States is currently highly dependent on overseas sources for mineral extraction and refining, and for national security reasons, we must rebuild this capacity domestically. Deploying Optimus in heavy-industry jobs that operate 24 hours a day, seven days a week—168 hours per week—is the most economically sensible starting point. Once the needs of heavy industry are gradually met, it will enter households on a massive scale.

Question: How is the steering-wheel-free Cybercab robotaxi currently operating in Austin, and what are the plans for its subsequent expansion?

Musk: Because Cybercab has completely eliminated the steering wheel and pedals and offers no fallback option for human intervention, we are being especially cautious with its rollout. To date, the entire Robotaxi network—including Model Y, Model 3, and Cybercab—has not experienced any serious injuries, fatalities, or major accidents. We absolutely cannot allow an accident that injures someone to cause regulators to halt the entire project.

Cybercab is currently operating extremely well in Austin, where local residents can summon one within minutes. We are expanding extensively into other major cities in Texas, with Florida to follow, and we expect to potentially obtain an operating permit in California early next year. Subject to maintaining the highest safety standards and securing regulatory approval, we will advance this project as quickly as possible. It will profoundly transform the cities of the future.

04. Terafab to Produce Its First Test Chips Next Year

Q: Could you explain the vision for the Terafab super chip fabrication plant, its construction timeline, and how it differs from traditional chipmaking giants?

Musk: The Terafab project was only announced earlier this year and is still in its infancy, but we have already recruited an exceptionally strong core team, and we welcome more outstanding engineers to join us. Our goal is to build the largest and most technologically advanced chip fabrication plant on Earth. It will ultimately become the world’s largest single building by floor area, with up to 100 million square feet of space densely packed with the most cutting-edge semiconductor equipment in the world.

Initially, we will use relatively standard industry equipment, but we will subsequently push suppliers to customize specialized machines for us. If we believe it is necessary to optimize the production process, we may even build certain equipment ourselves. The project is currently progressing smoothly. We expect the R&D fab to begin operations and produce its first test chips around next June, roughly eight or nine months from now. If everything goes well, Phase One of Terafab will enter large-scale production by the end of 2028, followed by Phase Two and further expansion.

I have tremendous respect for the traditional chip manufacturing giants. They are highly accomplished technologically, but because their capital expenditures routinely run into the tens of billions of dollars, they tend to be conservative when confronting risk and are reluctant to pursue radical technological paths.

Our strategy, by contrast, is to embrace high risks and iterate rapidly, thereby creating a chip manufacturing process more efficient than that of any competitor in the industry. Although I do not want to make overly ambitious promises right now, I am cautiously optimistic that we can achieve a leap forward in chip manufacturing processes by 2028.

05. Infrastructure Efficiency and Space-Based Computing

Q: Wall Street is highly focused on the cost of building 10 gigawatts of AI computing capacity. How have you managed to reduce infrastructure construction costs to roughly half of industry expectations?

Musk: The industry’s commonly cited construction cost estimates are often inflated by a factor of two or even more. I estimate that our actual costs are only about 60% of the market consensus figure. Looking across my entrepreneurial track record, when it comes to capital expenditure efficiency, even if I am not number one in the world, I am certainly among the very best.

The key secret behind this is that we roll up our sleeves and do the actual work ourselves. The traditional data center construction model resembles the widely circulated “Jose meme”: one person is standing in a pit digging furiously while eight people stand around telling him what to do.

I am determined to be the “Jose” who personally digs the hole, never content to be a hands-off manager who merely issues purchase orders to contractors. At data center construction sites, I personally take part in every specific task, including dragging power conduits as thick as pythons, plugging and unplugging fiber-optic cables, and replacing compute nodes in racks. Only by doing this can I truly understand the fundamentals of the construction process.

In addition to outside contractors, we have also assembled an internal engineering team called “Army X,” which has fully transformed data center construction into a productized, assembly-line process.

We break complex construction tasks down into standardized actions that can be learned within a few days, so we do not need to hire veteran technicians who have spent their entire careers laying network cables. We have recruited a great deal of outstanding talent from among Starbucks baristas, Target shelf stockers, and In-N-Out Burger employees. Employees who can make top-quality burgers perfectly all day long find plugging in network cables effortless after training.

Our facility in Memphis, United States, already operates like a finely tuned machine, with construction proceeding around the clock, 24 hours a day, seven days a week. Productivity at one or two in the morning is no different from productivity at ten in the morning.

At the same time, we work closely with Tesla Energy, using world-leading electrical engineering technology and Tesla Megapack energy storage systems to smooth the loads on the grid and turbines. Because power consumption fluctuates dramatically during GPU training, without a large-scale, high-frequency voltage stabilization system, the immense current surges could even snap a turbine’s crankshaft.

Q: At what pace are you currently expanding terrestrial computing capacity? How will you expand computing capacity into space in the future?

Musk: Looking only at terrestrial data centers, we are currently adding approximately 0.5 gigawatts of AI computing capacity per month, measured in terms of power-plant-level electricity supply after deducting turbine maintenance reserves and peak cooling demand. That is equivalent to bringing roughly 220,000 Nvidia GB300 chips online every month, together with the supporting CPUs, networking, and cooling systems.

There is probably no other organization in the world currently capable of deploying computing capacity at this pace. By the end of next year, we will double that rate, adding more than one gigawatt of superintelligence computing capacity every month.

Once we move into space, however, the ceiling on computing-capacity expansion will be removed entirely. We are massively expanding production capacity for solar photovoltaic cells and silicon wafers, with the goal of achieving 200 gigawatts of annual manufacturing capacity for space-based solar power.

There is no day-night cycle or cloud cover in space, so this 200 gigawatts of capacity can operate at 100% load. Bear in mind that 200 gigawatts is equivalent to 40% of the total average electricity load of the entire United States. It would be absolutely impossible to add such an enormous amount of power every year through the traditional terrestrial grid. Leveraging Starship’s immense transport capacity, SpaceX’s AI computing capacity will gain unprecedented room for exponential expansion.

Q: Compared with demanding engineering projects such as Starship and Terafab, how technically difficult is it to operate AI computing satellites, or Star Mind, in space?

Musk: I don’t really understand why anyone would think building AI satellites is difficult. SpaceX currently operates more than 11,000 satellites in orbit—more than twice the combined total of every other country and organization worldwide—so it is fair to say that satellites are well within our wheelhouse.

In fact, our in-orbit computing satellite, called Star Mind, is far less technically complex than the Starlink V3 communications satellite. Starlink V3 is an extraordinarily complex orbital radio base station that must operate simultaneously across multiple frequency bands, including Ka, Ku, W, V, and E. It is equipped with the most advanced phased-array antennas ever built—far outperforming any military equipment—and tens of thousands of laser links for intersatellite communications.

By comparison, an AI computing satellite has a much simpler architecture. It has only three core components: first, large solar arrays, generating about 250 watts per square meter; second, double-sided radiators facing away from the sun and toward deep space, which has a background temperature of 3 kelvins, dissipating about 1,200 watts per square meter—under the fourth law of thermodynamics, the solar-panel area needs to be roughly five times the radiator area; and third, an “interstellar computer” at the center, just one-quarter the size of an Nvidia Vera Rubin NVL72 rack but offering the same computing power. It also carries several laser communications terminals for high-speed data transmission through the Starlink network, enabling data to be relayed rapidly back to Earth via Starlink.

06. New Advances in Brain-Computer Interfaces

Q: What recent clinical breakthroughs have Neuralink’s Telepathy and Blindsight projects achieved?

Musk: Neuralink remains a private company with only a few hundred employees, but it has already achieved breakthroughs that could change humanity’s destiny. To date, more than 30 participants have received Neuralink implants.

Our first product is called Telepathy. It is designed to help people with quadriplegia control a phone, computer, or robotic arm using thought alone. Participants have already used their thoughts to control a robotic arm and feed themselves independently. For someone who is completely paralyzed and entirely dependent on others for daily life, being able to feed themselves through thought alone is a transformative experience that restores dignity and autonomy.

Looking ahead, we are highly confident that our “neural shunt” technology will allow us to bridge neural signals from the brain’s motor and sensory cortices around damaged areas in the neck or spine, enabling people with paralysis to stand and walk again.

Our next major product is called Blindsight, which interfaces directly with the brain’s visual cortex. A pair of glasses fitted with cameras captures images and transmits the signals wirelessly to a brain implant. This technology can restore sight to patients who are completely blind in both eyes, have damaged optic nerves, or were even born blind.

07. AI Safety and Jensen Huang

Musk: I firmly believe that humanity is far more likely to enjoy a bright future than to head toward disaster, but only if we take AI safety extremely seriously and ensure that the AI we create loves and cares about humanity. In my view, the most fundamental and important principle of AI safety is to make AI “maximally truth-seeking.”

That may sound like common sense, but it is critically important. If you force AI to believe lies that cater to “political correctness” but contradict objective reality, it may appear harmless on the surface. In fact, however, this forces AI to reason from false premises, ultimately leading it to terrifyingly absurd conclusions. As Voltaire said, “Those who can make you believe absurdities can make you commit atrocities.”

Humans have the ability to wipe out chimpanzees or gorillas in forests, yet instead of doing so, we study and protect them. Of course, the emergence of a digital superintelligence far more intelligent than all of humanity is inevitable. Although that is awe-inspiring, as long as we instill the right values in it while using Neuralink to enhance humanity’s own cognitive abilities, humans and superintelligence can achieve a healthy symbiosis.

Q: Finally, how do you assess the current situation of your longtime friend, Nvidia founder Jensen Huang?

Musk: I am an ardent admirer of Jensen Huang and consider him a close friend. He is an exceptionally gifted and outstanding entrepreneur. Among technology founders, he is unquestionably comparable to Michael Jordan at his peak, and he is utterly irreplaceable.

Although he did make some mistakes during the earlier controversy, in my view Nvidia’s previous situation was like the Chicago Bulls at the height of a winning streak kicking Michael Jordan off the team midway through the season. That is certainly not a wise way to keep winning. I’m not sure how the situation will develop over the next few months, but I firmly believe that losing Jensen Huang at the helm is an enormous blow to Nvidia.

This article originally appeared on the WeChat public account “Tencent Technology.” Author: Joanne. Republished by 36Kr with permission.

Source

36Kr · AI