Musk: Currently, the most energy-consuming task is building "Optimus." In two years, we'll also launch a Starlink phone.

On September 15th, news broke that at the recently held 2025 ALL - IN Summit, "Silicon Valley's Iron Man," Elon Musk, made a remote appearance. He talked about the latest progress of Optimus, Tesla's self - developed chips, the Starlink phone project, and the Grok model, and shared his views on artificial intelligence, the population crisis, and the future of humanity.

While Musk painted a picture of a future where artificial intelligence will surpass the combined wisdom of all humans before 2030, he also warned that Western society is on the verge of a crisis of population decline and "self - destruction."

The following is a selected version of part of the dialogue:

01. Musk Devotes the Most Energy to Optimus

Question: Optimus is regarded as the greatest product in human history. What's the current progress? How much energy have you invested in this product? Is it the third or fourth version now?

Musk: We're finalizing the design of Optimus Version 3. This robot will be truly outstanding. It will have hand dexterity close to that of a human, which means it will have a highly complex robotic hand. It will also be equipped with an artificial intelligence "brain" that can navigate autonomously and understand the real - world environment, and it will be mass - produced on a large scale. These three points are the key elements that other robot companies don't have.

Currently, among all projects, Optimus probably takes up most of my energy, more than any other single project. This requires solving real - world artificial intelligence problems, Optimus' mechatronics problems, as well as challenges in the supply chain and production. We have to build everything from scratch because the supply chain for humanoid robots simply doesn't exist yet. This requires a large amount of vertical integration, and all the actuators used in Optimus can't be obtained from the existing supply chain. If successful, Optimus will become the largest - scale product ever. By 2030, after achieving large - scale production, the cost of each robot is expected to be about $40,000.

Question: Then what will be the cost of the first batch of robots? When can we buy one for ranch work? You previously mentioned that once the annual production reaches one million units, the marginal production cost will be around $20,000.

Musk: This will be affected by the cost of the artificial intelligence chips in the robot. The efficiency of the actuators also needs to be significantly improved. Each arm has 26 actuators, including motors, gearboxes, and power electronics. The cost of artificial intelligence chips is not low. It may account for $5,000 to $6,000, or even more, of the bill - of - materials cost.

So, when the annual production of Optimus reaches one million units, the production cost is expected to be between $20,000 and $25,000, and the final selling price will be determined according to market demand.

Question: Over the past few decades, you've accumulated rich experience in factory operations. Compared with developing the Cybertruck, Model Y, Model X, or building the Gigafactory, which is more difficult: building Optimus?

Musk: Building Optimus is much more difficult than these, but a bit easier than the Starship. The difficulty lies between that of the Model X and the Starship.

Question: Which is more challenging, hardware or software?

Musk: At present, the main difficulty lies in the final design of the hardware, especially the hand design. Of course, the other parts of the robot are also important, but the hand design, including the forearm, accounts for most of the engineering difficulty of the entire robot.

Question: Assuming we overcome the hardware challenges, with the current development of large - language models, what direct assistance can we get? Can consumers directly interact with the robot, give instructions, and can the robot understand and execute them?

Musk: Of course, there's no problem with that.

Question: I noticed that you often interact with artificial intelligence online. Can you create a robot similar to "Any" (referring to a certain type of virtual intelligent agent)? Why choose to design a humanoid robot?

Musk: If you want the robot to be able to do all the things that humans can do, a humanoid form is essential. If it's just for a specific type of task, the design will indeed be much simpler. The human form and abilities have evolved for a reason. For example, having five fingers is very practical, and even the little finger plays an important role. Moreover, humans have designed the world according to their own needs, so humanoid robots can directly adapt to the existing environment and achieve "backward compatibility."

02. Tesla Adopts a "Dual - Chip" Strategy, and the AI5 Chip Will See a 40 - Fold Leap Compared to the AI4

Question: Besides large - language models and hand actuators, robots also need chips to drive them. You previously mentioned AI5 and AI6. It seems that you're very optimistic about the development direction at the chip level. Can you introduce the relevant situation?

Musk: At Tesla, we have two chip projects: one is Dojo for model training, and the other is the AI series of chips for inference. Currently, all models are equipped with the AI4 chip, and now we're finalizing the design of the AI5 chip.

Compared with the AI4, its performance will have a huge leap. In some indicators, the performance of the AI5 will be 40 times that of the AI4. The reason for such a large improvement is that we've carried out very detailed co - design in the research and development of artificial intelligence software and hardware, and we clearly know where the bottlenecks of the existing technology are.

Question: For consumers, in which aspects will this performance improvement be reflected? For example, will the safety of Full Self - Driving (FSD) felt by Tesla owners be improved by an order of magnitude? And how will the performance of the robot change?

Musk: To be precise, the 40 - fold improvement is aimed at the biggest shortcoming of the AI4 chip, which is the softmax operation (a commonly used activation function in neural networks). Currently, the AI4 chip needs to run about 40 steps in the simulation mode to complete the softmax operation, while the AI5 chip can complete it in just a few steps through its native architecture.

In addition, the AI5 can easily handle mixed - precision models, achieve dynamic adjustment of precision, and there are many optimizations in technical details, making its performance even better. In terms of basic computing power, the computing power of the AI5 is about 8 times that of the AI4, the memory capacity is about 9 times, and the memory bandwidth is about 5 times. Because we've solved the core bottleneck of the AI4, on the basis of an 8 - fold improvement in computing power, we can further stack a 5 - fold underlying optimization effect, ultimately achieving a 40 - fold performance leap.

Moreover, the AI4 chip currently installed in Tesla electric vehicles can already make the safety of Full Self - Driving reach 2 to 3 times that of human driving, and it may even reach 10 times. The software that matches this performance will be pushed in the next few months. This will be the largest - scale software upgrade for Tesla since Version 12. We'll increase the number of model parameters by an order of magnitude and apply a large amount of reinforcement learning technology.

03. The Starlink Phone Project: The First Batch of Spectrum - Compatible Phones Will Be Launched in Two Years

Question: You spent about $17 billion on purchasing spectrum to enable the Starlink satellite network to directly connect with mobile phones. What will this technology look like in one or two years? Will we abandon the mobile operator Verizon and switch to Starlink's service? How many people want a Starlink phone? Is it technically feasible?

Musk: This is a long - term project. Through this project, SpaceX can directly provide high - bandwidth network connections to mobile phones via satellites. However, to achieve this, the mobile phone hardware needs to be modified because current mobile phones don't support these new spectrums, and the chipset must be upgraded. The whole process will take about two years.

So, the first batch of mobile phones that support these spectrums is expected to be launched in two years. Meanwhile, we're also developing satellites that can be compatible with these spectrums and cooperating with mobile phone manufacturers to enable them to add relevant spectrum support functions to new mobile phones. Eventually, satellites and mobile phones can achieve efficient connection, and users can smoothly watch videos on their mobile phones anywhere.

Question: That sounds amazing. Can these spectrums be used indoors like current mobile phone signals?

Musk: If a building has a thick metal roof, it may be affected, but there won't be any problem using it in ordinary residential houses.

Question: Do you hope to achieve the vision that users no longer need to rely on regional operators like AT&T, don't have to worry about roaming when traveling to the UK or India, and can just sign up for a Starlink account to use the network? Eventually, will Starlink achieve global coverage and replace regional operators to become a global operator?

Musk: This is a possible direction, but we won't let other operators go out of business. They'll still exist because they own a large amount of spectrum resources. However, users can indeed have a Starlink account: they can use it with the Starlink antenna at home to enjoy free Wi - Fi and also use high - bandwidth networks on their mobile phones, forming a comprehensive high - bandwidth solution covering both home and mobile scenarios.

Question: Then will you consider acquiring an operator to obtain more spectrum? For example, acquiring Verizon?

Musk: We don't rule out this possibility.

Question: Let's talk about the Starship. The recent launch was very successful. How far is it from achieving stable commercial application?

Musk: I think we'll be able to recover the spacecraft next year. Currently, there's only one more mission left in the Starship launch plan for the second - generation Starlink satellites, with only one set of the second - version booster and spacecraft remaining. After that, we'll launch the third - generation Starship, which will be equipped with Raptor 3 engines, and almost all parts of the rocket will be upgraded.

Due to the significant design changes, the third - generation Starship may encounter some problems in the initial stage, but it can be fully reusable and can send more than 100 tons of payload into orbit. As long as there are no major accidents, SpaceX will be able to demonstrate the full reusability of the Starship next year, including recovering the booster and the spacecraft, and can send more than 100 tons of payload into a practical orbit. Currently, the best - performing commercial rocket is the Falcon Heavy. When its side boosters are reused, its payload capacity is about 40 tons, while the payload capacity of the Starship is 2.5 times that of the Falcon Heavy, and it can be fully reused.

Question: The Starship exploded during a previous launch, and there was a lot of discussion about the environmental impact and the review by the Federal Aviation Administration (FAA). But you quickly resumed the launch preparation. This is not only an achievement in terms of technology and execution. How did you obtain regulatory approval so quickly? Many people thought there would be a lot of reviews and doubts at that time.

Musk: There were indeed many review processes, but we passed them all smoothly. This is thanks to the SpaceX team. They made great efforts, quickly completed the testing, assembly of the next set of booster and spacecraft, and transported them to the launch pad for launch.

Building a fully reusable orbital rocket is one of the most difficult engineering problems ever, at least in the top three. This has been the goal of SpaceX since its establishment in 2002. Now, 23 years have passed, and we have the world's best rocket engineer team. I believe we'll definitely be able to achieve full reusability of the rocket next year.

Question: From now until achieving full reusability, what technical obstacles are you most concerned about? Is there a key problem that requires your full attention, or is it more about integrating previous lessons into the next launch?

Musk: The main challenge in achieving full reusability of the spacecraft is the heat shield (a device or structure used to protect spacecraft, rockets, or other high - speed flying objects from extreme high temperatures).

So far, no one has been able to build a fully reusable orbital - class heat shield. For example, the heat shield of the space shuttle needed 9 months of maintenance after each flight. This involves both materials science and engineering technology problems.

We're starting from the basic physical principles to develop a heat shield that can withstand extreme high temperatures, is lightweight, doesn't transfer heat to the main structure, can resist rain, and won't dissolve in the rain. Moreover, all tens of thousands of heat - resistant tiles on the heat shield must be reliable, and there's no need to check or refurbish them one by one after each flight like the space shuttle.

04. Grok Development Progress: Reconstructing Knowledge with Synthetic Data, AI Capabilities May Surpass the Sum of All Humans by 2030

Question: Can you talk about the progress of the Grok model and the xAI company? You previously mentioned that the training of the next - generation model won't rely on conventional web data crawling but will use a large amount of synthetic data. What's the development direction of the Grok model? Why is this technological innovation important?

Musk: We're investing a large amount of inference computing resources to analyze the original data in the human knowledge corpus. In this process, Grok will sort through each piece of information one by one, supplement missing content, correct incorrect information, and eliminate false content. Taking Wikipedia as an example, this processing method can be applied not only to Wikipedia but also to all information carriers such as books, PDF documents, and websites. Specifically, Grok will analyze each page through a large number of inference calculations, judge the authenticity of the content, then rewrite the page, delete false content, correct half - true and half - false information, and supplement missing context.

Question: Can you make these corrected contents public? For example, create a "Grokipedia" (a coined word combining Grok and Wikipedia, which can be understood as "Gro Encyclopedia")? Especially the biography pages on Wikipedia have many problems, are full of biases, and often have editing disputes.

Musk: I'll discuss this idea with the team. Building a platform like "Grokipedia" should be very interesting.

Question: Regarding the training of the Grok 5 model, you're expanding the Colossus super - cluster in Memphis. Can you introduce the latest situation? Also, from the perspective of Scaling Laws, does expanding the scale of the super - cluster mean creating a more powerful artificial intelligence model? Will there be a situation of diminishing marginal returns?

Musk: There's a natural logarithmic relationship between the investment in computing resources and the improvement of the intelligence level of the artificial intelligence model. Roughly estimated, a 10 - fold increase in computing resources can double the intelligence level of the model. For example, increasing from an IQ equivalent to that of a human of 100 to 200, which is still a very significant improvement.

Moreover, the intelligence level will continue to increase until we can use the energy of the sun or even the entire galaxy to provide computing support, reaching the computing scale of a Type II or Type III civilization on the Kardashev Scale (Type II can use the energy of its entire stellar system, such as the solar system; Type III can use the entire energy of its galaxy, such as the Milky Way).

Artificial intelligence is not an end but an important part of the process of improving the overall intelligence of humanity. The intelligence level of humanity used to increase with population growth, but now, due to population decline and low fertility rates, human intelligence is stagnant or even declining. I expect that by next year, artificial intelligence may surpass any single human in some fields; by 2030, the overall capabilities of artificial intelligence may even exceed the sum of all humans.

05. The Future of Humanity Faces a Crisis: Western Countries Are at Risk of Population Self - Decline

Question: Do you think humanity will decline due to the development of artificial intelligence? Is the Earth's ecosystem undergoing structural changes that we currently can't fully understand?

Musk: I hope the fertility rate will rise. I've always supported increasing the