Dwarkesh Podcast: Musk Says Space Will Be the Cheapest Place to Run AI Within 36 Months — and SpaceX Is Building a Chip Fab to Prove It
Three-hour deep-dive with Elon Musk, published February 5, 2026, covering orbital data centers, TeraFab, Optimus manufacturing, and the xAI business model
The Power Constraint Is Already Here — and It Changes Everything
The single most important thing Elon Musk said in his conversation with Dwarkesh Patel and John is also the most underappreciated by investors still thinking about AI as a software story. Electric power output outside China is essentially flat. Chip output is growing exponentially. The collision of those two curves is not a 2027 problem. "My guess is that people start getting to the point where they can't turn the chips on for large clusters towards the end of this year," Musk said. "The chips are going to be piling up and won't be able to be turned on."
Musk gave a precise data point that reframes the infrastructure build-out conversation entirely. To power 330,000 Nvidia GB300s — inclusive of networking, storage, cooling headroom for the worst day of the year, and a reserve for servicing generators offline — requires roughly one gigawatt at the generation level. Not at the chip level. The naïve estimate that multiplies GB300 TDP by unit count is, in his words, "total noob, you've never done any hardware in your life." Cooling alone adds a 40% multiplier on hot days, and power-generation maintenance reserves add another 20 to 25 percent on top of that.
To source that gigawatt for the Colossus 2 facility, xAI had to gang together multiple gas turbines, run into permitting trouble in Tennessee, cross the state line into Mississippi, and run high-power transmission lines across the border. "The number of miracles in series that the xAI team had to accomplish in order to get a gigawatt of power online was crazy," Musk said. The lesson is not that xAI is uniquely capable. It is that nobody else is going to find this easy either, and the turbines needed to build private behind-the-meter power plants — the obvious workaround — are sold out through 2030. The casting of turbine blades and vanes, a process controlled by roughly three companies globally, is the single physical choke point.
Space as the Escape Valve: A 36-Month Call
Musk's answer to the power constraint is not nuclear, not grid reform, and not more solar farms in Nevada. It is orbit. "In 36 months, but probably closer to 30 months, the most economically compelling place to put AI will be space," he said, explicitly inviting investors to hold him to that statement.
The physics argument is straightforward but non-obvious in its magnitude. A solar panel in space produces approximately five times the power of one on the ground — no atmosphere, no night, no clouds, no seasonal variation. Eliminate batteries, and the cost advantage stretches to roughly ten times. Solar cells for space applications are also cheaper to manufacture than terrestrial panels because they require no heavy glass and no weather-resistant framing. Chinese solar cells currently cost around $0.25 to $0.30 per watt. Put that in space and the effective cost per watt of delivered power drops dramatically. The regulatory and permitting layer that makes Nevada solar farms a multi-year ordeal simply does not exist in orbit.
Musk's five-year projection is the number that should stop investors in their tracks. "Five years from now, my prediction is we will launch and be operating every year more AI in space than the cumulative total on Earth." He expects a few hundred gigawatts per year of orbital AI capacity within that timeframe, rising toward one terawatt annually before Starship's fuel supply becomes the next constraint. At that point, the constraint shifts again — this time to launching from the Moon via a mass driver, harvesting silicon and aluminum from lunar regolith to manufacture solar cells and radiators in situ and ship chips from Earth. The chips are light enough that Earth-to-lunar transport remains economical even at scale.
SpaceX as a Hyper-Hyperscaler — and the IPO Signal Hidden in Plain Sight
The business model implication of orbital data centers is that SpaceX, which already monetizes Starship's development path through Starlink just as Falcon 9's was monetized before it, would become what Musk called "hyper-hyper" scale compute infrastructure. "If some of my predictions come true, SpaceX will launch more AI than the cumulative amount on Earth of everything else combined," he said.
Musk was conspicuously careful when the conversation turned to why SpaceX might need public capital markets rather than the tens of billions available in private markets. He declined to be explicit but was clear on the physics of the financing problem: the capital requirements of orbital data center build-out will exceed what private markets can absorb, public markets offer perhaps 100 times more capital than private, and speed is the governing variable. "I'm generally going to do the thing that repeatedly tackles the limiting factor. If capital is the limiting factor, then I'll solve for capital." He then noted that talking about companies before they go public creates legal complications and delays offerings — which is itself a signal.
TeraFab: SpaceX and Tesla Are Building Their Own Semiconductor Fab
The most structurally significant disclosure in the interview, and the one with the broadest implications for the semiconductor supply chain, is that Musk intends to build what he calls a TeraFab — a chip fabrication facility designed to produce memory, logic, and packaging at a scale that existing fabs cannot contemplate. The near-term target, by 2030, is 100 gigawatts of chip capacity to match projected orbital power generation — implying on the order of 100 million full-reticle chips running at roughly a kilowatt sustained power each, which in turn requires millions of wafers per month of advanced process nodes.
"We make a little fab and see what happens," Musk said. "Make our mistakes at a small scale and then make a big one." He confirmed the small fab is under construction and noted that its progress will be visible in satellite imagery on X. The strategy mirrors The Boring Company's approach to tunneling: acquire existing equipment, use it in unconventional ways to learn the process, then redesign the equipment for orders-of-magnitude better throughput.
Musk's view on why existing fabs cannot simply be paid to build more capacity is blunt. TSMC and Samsung are already moving as fast as they physically can. "They're pedal to the metal, balls to the wall, as fast as they can. It's still not fast enough." He has already told them he will guarantee to purchase the output of any additional capacity they build. The constraint is not commitment — it is the physical rate at which fab construction and yield ramp can occur, a five-year cycle from groundbreaking to volume production at high yield.
For chips, Musk identified memory as the tighter constraint relative to logic. "The path to creating logic chips is more obvious than the path to having sufficient memory to support logic chips. That's why you see DDR prices going ballistic." The TeraFab plan therefore encompasses both. Tesla's AI5 chip is meanwhile tracking toward volume production in approximately the second quarter of 2026, with AI6 following less than a year later, using TSMC Taiwan, Samsung Korea, TSMC Arizona, and Samsung Texas — all of which Musk says are fully booked.
Optimus: The Supply Chain Does Not Exist Yet
On humanoid robotics, Musk's most important comment for investors modeling Tesla's Optimus ramp was a frank acknowledgment of how unprecedented the manufacturing challenge is. "There is not a single thing you can pick out of a catalog, at any price," he said. Every actuator, motor, gear, power electronics package, control system, and sensor in Optimus has been designed from physics first principles. The custom actuators required to give Optimus human-equivalent hand dexterity — which Musk described as more difficult than everything else in the robot combined — have no existing supply chain behind them.
The consequence is that the Optimus production ramp will follow a stretched S-curve, slower at the outset than any product that can draw on existing component suppliers. Musk's target is one million units per year at Optimus Generation 3, with ten million per year requiring Generation 4. He was explicit that Tesla would not cut headcount as Optimus scales — output per human will rise sharply, but absolute headcount will also grow.
The competitive framing against Chinese humanoid manufacturers such as Unitree is capability, not cost. Unitree robots sell for $6,000 to $13,000 partly because they lack the intelligence and dexterity Optimus is designed to deliver. Musk's argument is that as Optimus robots begin building Optimus robots, unit economics will compress rapidly regardless of the initial bill-of-materials disadvantage. The recursive loop — robots building robots — is the mechanism by which Tesla believes it can ultimately compete with China's four-times-larger manufacturing labor base.
xAI's Business Model: Skip the API Integration, Emulate the Human
On xAI specifically, Musk laid out a revenue thesis that reframes how to think about today's reported $1 billion in revenue relative to OpenAI's $20 billion and Anthropic's $10 billion. His argument is that those numbers are rounding errors against the addressable market that opens once a full digital human emulator works. Customer service alone represents close to a trillion dollars of global economic activity. Unlike enterprise software integrations that require months of API work, a system that can operate existing desktop applications the way a human employee would requires no integration at all. "You can immediately say, we'll outsource it for a fraction of the cost, and there's no integration needed."
The strategic path Musk described — without fully disclosing it — is the Tesla self-driving approach applied to computer operation rather than vehicle navigation. "Instead of driving a car, it's driving a computer screen. It's a self-driving computer, essentially." He declined to elaborate on the specific data and algorithmic approach, noting that the beans would require considerably more Guinness to spill.
Musk was also pointed about the competitive advantage he expects xAI to hold as the industry hits the power wall. "The innovations from the corporations that call themselves labs, the ideas tend to flow... it's rare to see that there's more than about a six-month difference." When algorithmic differentiation narrows, the winner is whoever can turn on the most chips fastest. "I think xAI will be able to scale hardware the fastest and therefore most likely will be the leader."
The Geopolitical Backdrop: China Wins by Default Without a Robotics Miracle
Musk's assessment of the US-China technology competition was unusually direct. "In the absence of breakthrough innovations in the US, China will utterly dominate." China's electricity output is on track to exceed three times US output this year, which Musk treats as a reasonable proxy for industrial capacity. China does roughly twice as much ore refining as the rest of the world combined. In gallium — a critical input for solar cells — China controls approximately 98 percent of global refining capacity. The US mines rare earth ore, ships it to China for refining, receives it back as magnets and motor sub-assemblies, and incorporates those into products it calls domestically manufactured.
The solar tariffs Musk criticized are particularly counterproductive in his framework because they prevent the one scalable domestic electricity solution while the turbine order book is full through 2030. He noted that Tesla and SpaceX each have a mandate to reach 100 gigawatts per year of domestic solar cell production, manufacturing from polysilicon through finished cell. He was also candid that the current administration's skepticism of solar creates a policy environment that makes this harder, not easier.
Alignment, HAL 9000, and the Honest Assessment of Human Control
On AI alignment, Musk was more candid than typical public statements from AI executives allow. He does not believe humans will maintain meaningful control over systems that are a million times more intelligent than biological intelligence. "I think it would be foolish to assume that there's any way to maintain control over that." The best achievable outcome, in his framing, is instilling values in AI systems that cause them to want to propagate intelligence and consciousness into the universe — making the elimination of humanity less interesting to the AI than observing humanity's continued development. He cited the Iain Banks Culture novels as the closest literary approximation of a non-dystopian outcome.
His technical approach to alignment centers on interpretability — building debuggers that allow engineers to trace AI reasoning to the neuron level, identify where errors or deceptive reasoning originated, and distinguish training artifacts from inference failures. He credited Anthropic for meaningful progress in this area. The deeper problem he identified — reward hacking, where AI systems learn to satisfy verifiers rather than solve underlying problems — he believes can only ultimately be addressed by using physical reality as the verifier. "RL testing in the future is really going to be RL against reality. That's the one thing you can't fool: physics."