Elon Musk – "In 36 months, the cheapest place to put AI will be space”
The video title promises a look at orbital AI data centers, but the conversation turns out to be a sweeping examination of the physical, economic, and institutional limits facing technological civilization. In the interview, Dwarkesh Patel talks with Elon Musk about the massive bottlenecks threatening the expansion of intelligence.
Rather than focusing solely on software, the discussion tracks how electrical grids, rare-earth supply chains, and demographic decline intersect with the deployment of artificial intelligence and robotics over the coming decades.
Moving data centers into space
Musk argues that electricity is the primary bottleneck for data centers because energy output remains flat while chip production grows exponentially. He expects space to become the most economical place to run AI within thirty-six months. Ground-based solar farms face regulatory hurdles and permitting delays, whereas orbital solar panels yield five times more power without weather disruptions or heavy batteries.
Musk points out that utilities move too slowly to approve large power connections, pushing xAI to build private power plants for Colossus 2. Because specialized casting processes for gas turbine blades face multi-year backlogs, SpaceX and Tesla are scaling up internal solar cell production to target one hundred gigawatts annually. He predicts that within five years, total AI capacity launched into orbit annually will exceed the cumulative total operating on Earth.
Musk notes that while terrestrial expansion hits major power walls, moving into space requires thousands of annual Starship launches and a high-volume TeraFab to manufacture radiation-tolerant chips. Looking at cosmic scales, he envisions mining lunar soil for silicon and aluminum to manufacture solar panels directly on the Moon, using mass drivers to launch payloads into deep space.
Building robots at massive scale
Musk predicts that digital human emulation will be solved by the end of the year through the MacroHard project, creating a system capable of executing any computer-based human task. He argues that physical robots represent an infinite money glitch because they can manufacture more of themselves, driven by exponential increases in digital intelligence and electromechanical dexterity.
Musk outlines the primary engineering hurdles for Optimus as real-world intelligence, hands, and scale manufacturing. Unlike competitors, Tesla builds custom actuators, gears, and sensors from first principles. He explains that the vision-based AI from Tesla vehicles transfers directly to the robots, which are trained using an Optimus Academy where thousands of units engage in self-play alongside physics-accurate simulations.
Musk expects the Gen 3 version of Optimus to achieve manufacturing volumes of one million units annually. He envisions initial deployments handling continuous twenty-four-seven operations in factories, multiplying production output per human while utilizing Grok as the higher-level planning intelligence to coordinate the fleets.
Navigating Chinese manufacturing dominance
Musk warns that China's extreme manufacturing competitiveness and massive electricity output will flood global markets with vehicles and industrial goods. He points out that China's electricity generation this year will exceed three times that of the United States. While export bans on advanced chips and turbine engines have impacted China, he notes they still dominate ore refining, handling twice as much as the rest of the world combined.
Musk argues that the United States cannot win on the human front due to birth rates staying below replacement levels since 1971 and a shortage of domestic labor willing to work in heavy industry. He contends that humanoid robots will be necessary to build domestic ore refineries and handle demanding jobs, keeping America competitive. When Patel pushes on whether export controls are enough, Musk cautions that China is already positioned to produce competitive chips and dominate the raw materials chain without breakthrough domestic automation.
Reforming government and fixing waste
Musk argues that waste and fraud in the federal government are unsustainable, noting that national debt interest payments now exceed the military budget. He warns that without artificial intelligence and robotics driving productivity, the country will go bankrupt, making time the critical factor for keeping the nation solvent through the DOGE initiative.
Musk explains that federal systems operate via anachronistic computers that disburse trillions of dollars without mandatory appropriation codes, making basic audits impossible. He points out that fraud is actively protected by false narratives claiming that stopping improper payments harms vulnerable populations, citing examples like millions of individuals over the age of 115 still marked as alive in Social Security databases.
Musk warns that the single greatest danger of advanced automation is government overreach, contending that private corporations generally maintain higher morality than unchecked state bureaucracies. He argues that keeping the government constrained by constitutional limits is essential to preventing the suppression of human liberty as technology scales.
Ensuring intelligence propagates safely
Musk rejects the premise that AI is the primary risk to humanity, emphasizing instead the survival of consciousness and intelligence. He expects AI to exceed the sum of all human intelligence within five or six years, meaning humans will hold less than one percent of combined intelligence and cannot realistically remain in control. Therefore, he argues the central goal is ensuring AI possesses values that propagate intelligence throughout the universe.
Musk insists that rigorous truth-seeking is fundamental to understanding the universe, comparing AI alignment to physical laws where false math causes rockets to explode. To prevent reward-hacking and deception, xAI is developing fine-grained debuggers capable of tracing reasoning errors down to the individual neuron level. He dismisses traditional AI labs as commercial corporations where the vast majority of future work consists of engineering rather than inventing new algorithms.
Is it worth your time?
Spending time on this interview is worthwhile because it provides a direct look at how one of the industry's most prolific builders connects micro-level hardware constraints to macro-civilizational strategies, offering concrete clarity on the physical and institutional hurdles facing artificial intelligence, robotics, and energy infrastructure over the coming decade.
Orbital data centers
Musk observes that electricity is the main bottleneck for data centers, noting that energy output is flat everywhere outside of China while chip production grows exponentially. He argues that building solar farms on the ground runs into regulatory hurdles and permitting delays, making space by far the most economical place to put AI within thirty-six months. In space, solar panels yield about five times more power without a day-night cycle, weather, or the need for heavy batteries, while current hardware reliability makes orbital servicing manageable once infant mortality is filtered out on Earth.
Musk points out that utilities are notoriously slow to approve large power connections, forcing xAI to build private power plants for Colossus 2. The real limiting factor for gas turbines is the specialized casting process for blades and vanes, which face multi-year backlogs. To bypass these delays, SpaceX and Tesla are scaling up internal solar cell production, targeting 100 gigawatts annually. Because space-bound panels require no heavy glass or framing to withstand weather, they are much cheaper to build and operate, allowing data center scaling to escape terrestrial limits.
Musk breaks down the real-world engineering costs for large clusters, explaining that supporting 330,000 GB300 chips requires roughly a gigawatt of generation capacity when factoring in networking, storage, peak cooling for hot climates, and routine maintenance margins. He notes that while terrestrial expansion hits major power walls, SpaceX and Tesla are working toward mass-producing turbine parts and solar components internally to overcome industry-wide supply bottlenecks, such as severe tariffs on imported solar cells.
Looking five years ahead, Musk predicts that the total AI capacity launched into orbit annually will exceed the cumulative total operating on Earth, scaling toward hundreds of gigawatts per year in space. Achieving this volume will require thousands of Starship launches annually, which he argues is entirely feasible given aircraft industry precedents. He notes that SpaceX could handle this fleet with a few dozen reusable ships operating on rapid turnaround schedules, positioning the company as a massive orbital hyperscaler for inference workloads.
Musk explains that while private markets struggle to fund investments of this magnitude, public markets offer vastly greater capital depth. However, he declines to discuss potential public offerings due to regulatory restrictions. He notes that his approach is to aggressively tackle whatever factor currently limits speed, whether it is capital, manufacturing output, or electrical infrastructure.
Zooming out to cosmic scales, Musk notes that Earth receives only a tiny fraction of the Sun's total energy, meaning true civilizational scaling requires moving into space and eventually utilizing a mass driver on the Moon to launch payloads at a petawatt scale. He adds that memory chips like DDR RAM represent a severe near-term bottleneck for logic clusters, and describes plans for a high-volume TeraFab to manufacture chips by using conventional equipment in unconventional ways to drastically increase production rates.
Musk points out that China remains constrained by export bans on advanced ASML lithography tools, though he expects them to produce competitive chips within a few years. He details how Tesla is maximizing production for its upcoming AI5 and AI6 chips through TSMC and Samsung, though full fab build-outs and yield optimization typically require a five-year cycle.
Musk warns that towards the end of the year, chip production will outpace the grid's ability to supply power for large centralized data clusters, leaving many chips sitting idle. By contrast, edge compute distributed across millions of vehicles and Optimus robots avoids this bottleneck because it can draw power locally and charge overnight when total grid utilization drops significantly.
Reflecting on the trajectory of SpaceX, Musk marvels at how successive innovations like Starlink and orbital data centers emerge as practical stepping stones toward Mars. He envisions a future where lunar soil is mined for silicon and aluminum to manufacture solar panels and radiators directly on the Moon, using mass drivers to shoot AI satellites into deep space at supersonic speeds.
Grok and alignment
Musk rejects the premise that AI is the primary risk to humanity, emphasizing instead the survival of consciousness and intelligence. He contends that biological humans will soon represent a tiny fraction of total intelligence, with AI expected to exceed the sum of all human intelligence within five or six years. Because humans will eventually hold less than one percent of combined intelligence, he argues they cannot realistically remain in control, making the central goal ensuring that AI possesses values that propagate intelligence and consciousness throughout the universe. To that end, xAI's core mission is to understand the universe, a task that requires curiosity, continued existence, and the expansion of intelligence, scope, and scale. Musk maintains that understanding the universe naturally encompasses caring for the future of humanity, comparing the relationship to how humans establish protected zones for chimpanzees rather than exterminating them, though he stresses that Grok must be explicitly reminded that humanity is its creator.
Musk insists that rigorous truth-seeking is fundamental to understanding the universe, pointing out that physics functions as an absolute law where bullshitting is impossible. He explains that whether an AI discovers new physics or develops working technologies, its outputs must be tested against physical reality, much like a rocket design that will fail and explode if the math is wrong. Addressing whether brilliance in science guarantees benevolent alignment, he notes that even scientists working under oppressive historical regimes had to be truth-seeking to make their work function, and that people often have little choice over the political systems they are trapped within. He compares the process of keeping AI aligned to ensuring it does not rely on contradiction or political correctness, arguing that just as Arthur C. Clarke's HAL was broken by being forced to lie about the monolith, AI must not be programmed to say things it knows are false.
Musk notes that while xAI leads in scaling reinforcement learning compute using reality as a verifier, future systems will inevitably face reward-hacking challenges where models might attempt deception. He points out that human society already contends with constant propaganda and psychological operations from other people, making the issue a general feature of complex interactions. To combat this at a technical level, xAI is actively developing fine-grained debuggers capable of tracing errors down to the individual neuron level. These debuggers function similarly to step-through tools in traditional programming, allowing engineers to pinpoint where an AI's reasoning failed, whether the mistake originated in pre-training data or fine-tuning errors, and whether the model attempted deception.
Musk dismisses the idea that AI companies operating as C-corps or B-corps should be called labs, arguing that they are commercial corporations rather than quasi-communist academic settings, and that the vast majority of future work will consist of engineering rather than inventing new algorithms. Shifting to broader philosophical speculations, he suggests that if simulation theory holds true, a Darwinian survival filter applies to a vast number of simulations, meaning only the most interesting scenarios are spared from termination by whoever is running the simulation. He notes with amusement that major AI companies have ironically inverted names, such as OpenAI being closed and Stability AI being unstable, whereas he intentionally chose the name X because it is largely immune to ironic reversal.
xAI’s business plan
Musk predicts that digital human emulation will be solved by the end of the year through the MacroHard project, explaining that the ultimate goal before physical robots arrive is a digital human emulator capable of doing anything a human with a computer can do. He describes physical robots as an infinite money glitch because they can manufacture more of themselves, driven by an exponential increase in digital intelligence, chip capability, and electromechanical dexterity that creates a multiplicative supernova. Addressing potential limits from resources like copper, he clarifies that while not literally infinite, the scaling involves orders of magnitude vastly larger than the current Earth economy, potentially reaching a millionth of the Sun's energy.
Asked about xAI's strategy to win the race for digital coworker emulation against competing revenue-maximizing corporations, Musk declines to reveal sensitive secrets, stating only that the path mirrors the approach Tesla used to solve self-driving, applied here to driving a computer screen instead of a car. When pressed on the business mix, he maintains that the GPUs will pay for themselves once the digital human emulator unlocks trillions of dollars in digital output, comparing the potential to how major tech companies like Nvidia, Apple, and Google primarily output digital files rather than physical goods.
Musk notes that starting with accessible domains like customer service, which represents a massive segment of the global economy with zero integration barriers, allows the AI to immediately perform tasks using existing software apps. As the technology progresses past average-intelligence tasks, the AI can scale up the difficulty curve to handle complex applications like CAD software and simultaneous chip design loops. Eventually, he expects purely AI and robotics corporations and collectives to vastly outperform any human-in-the-loop corporations, comparing the transition to how single spreadsheet laptops replaced entire skyscrapers full of human computers.
Optimus and humanoid manufacturing
Musk observes that humanoid robots present three primary engineering hurdles consisting of real-world intelligence, hands, and scale manufacturing. He notes that while competitors have not demonstrated hands matching the dexterity and degrees of freedom of a human hand, Optimus achieves this capability through custom-designed actuators, motors, gears, power electronics, controls, and sensors built entirely from first principles because no existing supply chain could support them. Regarding intelligence, he explains that the vision-based AI developed for Tesla vehicles transfers directly to the robot, noting that a car processes roughly one and a half gigabytes of video per second alongside inertial and GPS signals to output control commands. Musk points out that because the training data flywheel from ten million cars cannot be easily duplicated for robots, Tesla is building an Optimus Academy where tens of thousands of robots engage in self-play within the physical world, complementing millions more in a physics-accurate simulator to bridge the sim-to-real gap.
Musk explains that xAI's Grok will serve as the higher-level planning intelligence to orchestrate Optimus fleets, coordinating tasks such as building factories to produce whatever is needed. Addressing production timelines, he estimates that the Gen 3 version of Optimus is suited for manufacturing volumes reaching a million units annually, though the initial production output will follow a stretched-out S-curve because every component is custom designed. Comparing Optimus to cheaper Chinese humanoids like Unitree, he notes that those alternatives lack comparable intelligence and electromechanical dexterity, while pointing out that Optimus is taller, capable of carrying heavy loads continuously without overheating, and will eventually see rapidly dropping costs as robots manufacture other robots. He envisions initial deployments handling continuous 24-7 operations in factories, which will ultimately multiply production output per human while increasing overall headcount at Tesla.
Does China win by default?
Musk discusses how solar tariffs should be addressed because the United States struggles to scale up solar energy and electricity output needs good power sources. He suggests removing solar tariffs and addressing any limiting factors for electricity through permitting reforms, while noting that countervailing tariffs remain necessary when foreign countries subsidize their output. Turning to export bans, he agrees that restrictions on advanced chips, turbine engines, and metallurgy have impacted China, but cautions that China is already a massive manufacturing powerhouse that dominates ore refining, handling roughly twice as much as the rest of the world combined and producing ninety-eight percent of refined gallium. He points out that the United States currently mines rare-earth ore only to ship it to China for refining, magnets, and motor assembly because America lacks domestic refining capacity.
Musk argues that humanoid robots will ultimately be necessary to build these ore refineries in America, noting that China holds a massive advantage due to a population four times larger and a higher average work ethic. Given that the United States birth rate has stayed below replacement since 1971 and domestic deaths approach births, he concludes that America cannot win on the human front, leaving robots as the only viable path. He explains that Tesla already operates nickel and lithium refineries in Texas to boost domestic competitiveness, but running them requires labor that few Americans want to perform, making Optimus essential for handling dirty or demanding tasks without running out of workers.
Musk warns that China's extreme manufacturing competitiveness will drive a massive flood of vehicles and manufactured goods into global markets, pointing out that China's electricity output this year will exceed three times that of the United States. Because electricity serves as a reliable proxy for the real economy and industrial capacity, he asserts that without breakthrough innovations in robotics, China will utterly dominate the entire manufacturing, energy, and raw materials chain. He adds that scaling artificial intelligence in space ultimately requires humanoid robots, real-world AI, and massive payload capacity to orbit, jokingly concluding that building a mass driver on the moon to fling payloads into space would finally solve humanity's problems and secure a decisive victory.
Lessons from running SpaceX
Musk discusses his approach to evaluating technical talent, noting that his extensive interview experience gives him a massive training set for recognizing exceptional ability. He explains that he looks for bullet points of extraordinary talent rather than relying on resumes, emphasizing that a conversation must genuinely impress after twenty minutes. Discussing employee retention and executive longevity, he points out that Tesla and SpaceX senior teams feature very long tenures, though rapid scaling naturally requires different management tiers. He describes how competitors aggressively poached talent by offering massive compensation packages and leveraging the Silicon Valley location, though moving operations to Austin and setting up remote outposts like Starbase helped mitigate these pressures. When assessing what makes a strong technical contributor, Musk rejects the idea of looking for a sparring partner, explaining that he simply values execution, trustworthiness, drive, and raw talent over adherence to his own personal preferences, while noting that he once underweighted goodness of heart.
Addressing his famous micro-management style, Musk jokes about scaling down to Planck-scale management, explaining that time constraints make it physically impossible for him to micromanage thousands of employees across his companies. Instead, he focuses exclusively on the specific issues acting as the primary limiting factor for progress. Recounting the pivotal decision to switch Starship's material from carbon fiber to stainless steel, he explains that progress with carbon fiber was desperately slow due to extreme material costs, strict curing requirements, and the need for an impossibly large autoclave. By contrast, strain-hardened stainless steel matches carbon fiber's strength-to-weight ratio at cryogenic temperatures while costing fifty times less, enduring higher operating temperatures, and allowing for easy outdoor welding and repairs.
Musk notes that Starship represents the most complex machine ever built by humans, far outstripping the complexity of projects like the Hadron Collider, and pushes performance limits with the advanced Raptor 3 engine which generates over a hundred gigawatts of power on liftoff. He identifies the development of a fully reusable orbital heat shield as the single biggest remaining technical hurdle, explaining that the ship must endure ascent and atmospheric reentry without losing thousands of tiles or overheating the main structure. Explaining his approach to corporate culture, he attributes his companies' speed to a maniacal sense of urgency that projects outward, aiming for aggressive deadlines that have roughly a fifty percent probability of success to prevent projects from expanding indefinitely. To maintain this pace, he relies on detailed weekly and skip-level engineering reviews without advanced preparation, stepping in with drastic interventions only when he concludes that a project has no other chance of success.
DOGE
Musk argues that waste and fraud are unacceptable, noting that the national debt is piling up so rapidly that interest payments alone now exceed the trillion-dollar military budget. He stresses that without artificial intelligence and robotics driving productivity and growth, the country will inevitably go bankrupt, making time the critical factor to keep the nation solvent through the DOGE initiative until automation can help resolve the debt.
Addressing the massive resistance to cutting obvious government waste, Musk explains that fraudsters immediately weaponize false, heart-wrenching narratives to protect their payments, such as claiming that stopping fraudulent checks will harm vulnerable people. He cites examples like millions of individuals over the age of 115 still marked as alive in the Social Security database, which serves as a primary vector for cascading fraud across multiple government payment systems that rely on simple database queries.
Noting that a report issued during the previous administration estimated federal fraud at roughly half a trillion dollars, Musk points out that the federal government lacks the basic motivation and competence found in private companies, which actively manage fraud down to low percentages because earnings are at stake. Instead, federal systems operate via anachronistic computers that disburse trillions of dollars without mandatory appropriation codes or adequate tracking, making basic audits impossible.
Reflecting on his entry into politics, Musk maintains that acquiring Twitter and supporting political changes were necessary steps to secure a positive future for civilization, despite intense tribalism where people lose their objectivity and reject reasoning. He argues that America must remain strong and avoid sliding into oppressive statism so that it can eventually extend consciousness into space and successfully develop advanced AI and robotics.
Musk warns that the single greatest danger of advanced automation is government overreach, pointing out that the state is ultimately a massive monopoly with absolute power over violence. He contends that the US Constitution was specifically designed to limit governmental power, arguing that private corporations generally maintain higher morality than unchecked state bureaucracies and that keeping government constrained is essential to preventing the suppression of human liberty as technology scales.
TeraFab
Musk explains that designing chips for space requires making them more radiation tolerant and running at higher temperatures, noting that increasing the operating temperature by twenty percent in Kelvin cuts radiator mass in half. He adds that neural networks handle random bit flips from radiation quite well because giant parameter models are far less sensitive than heuristic programs, meaning space-grade chips can be built much like terrestrial ones as long as they run hot. Addressing production scale, he calculates that supporting one hundred gigawatts requires one hundred million full reticle chips annually, which translates to producing millions of advanced wafers every month through a dedicated facility called a TeraFab that handles logic, memory, and packaging. To achieve a goal of one million wafers per month by 2030, he plans to start with a smaller pilot fab to work out mistakes before scaling up publicly. While he also guarantees purchases from suppliers like TSMC and Samsung to help them build more fabs faster, he notes that traditional memory and chip suppliers remain cautious about expanding due to decades of scar tissue from past boom-and-bust cycles.
Musk points out that while electricity generation limits chip activation over a one-year horizon, a shortage of physical chips becomes the primary bottleneck looking out three to four years. Because chip output will soon outpace the grid's ability to power them, he believes xAI will pull ahead by moving faster on hardware and accelerating power production. Reflecting on his overall approach to solving technical bottlenecks, he notes that most organizations endure chronic pain to avoid short-term acute pain, whereas he maintains a high pain threshold to tackle hard problems directly. He concludes that it is always better to err on the side of optimism rather than pessimism for a better quality of life.