Elon Musk Predicts Space Will Become the Cheapest AI Training Ground Within 36 Months

Interview Overview

Three days after SpaceX announced the $1.25 trillion acquisition of xAI, Elon Musk sat down with Dwarkesh Patel and Stripe co‑founder John Collison for a nearly three‑hour podcast. The conversation covered AI compute power, space‑based data centers, supply‑chain constraints, chip‑fab plans, robotics, and broader geopolitical implications.

1. Power as the Hard Bottleneck for AI

Dwarkesh asked why power, not chips, limits AI. Musk explained that worldwide electricity generation outside China is essentially flat while AI‑chip output is growing exponentially, leading to a future where chips cannot be powered. He noted that space solar panels are about five times more efficient than ground‑based panels and require no batteries, making the total cost of space power roughly one‑tenth of terrestrial power.

“You need electrical transformers to drive AI transformers.”

When pressed about the engineering challenges of deploying AI in orbit, Musk acknowledged unanswered questions about launch cost, thermal management, bandwidth, and accelerated GPU depreciation, but dismissed them as non‑issues for the moment.

2. Turbine‑Blade Supply Chain Constraint

Collison asked why SpaceX does not simply build its own power plant. Musk described the Colossus 2 data‑center project in Memphis, which acquired a retired Duke Energy gas‑turbine plant in Southaven, Mississippi, and deployed 200 MW within six months. However, the critical bottleneck is the manufacturing of turbine blades and vanes, a process performed by only three companies worldwide and already booked through 2030.

“The ultimate bottleneck for power expansion is not money, it’s physical supply‑chain limits.”

He quantified that 330,000 GB300 chips would require a full 1 GW of power, with cooling accounting for 40 % and redundancy 20‑25 %.

3. Space as the Cheapest AI Deployment Site

Musk predicted that within 30‑36 months, space will become the most economical location for AI workloads, citing the 5× solar efficiency and the absence of battery costs. He admitted the forecast omits launch and thermal challenges, but argued the ground‑based power wall makes space the only viable path.

4. Chip‑Fab Ambitions (TeraFab)

Musk outlined a “TeraFab” plan to produce a million chips per month, requiring 100 GW of power. He admitted he does not yet know how to build a fab but will figure it out, likening the approach to the Boring Company’s incremental tunnel‑digging strategy. He also disclosed pre‑orders with TSMC and Samsung, noting that even with full capacity, fab ramp‑up will take five years.

He warned that by the end of 2026 chip production could outpace available power, making whoever can supply electricity the AI leader.

5. Optimus Hand Challenge

When asked about the hardest part of building the Optimus robot, Musk said the hand—requiring custom motors, gears, power electronics, control systems, and sensors—poses a greater difficulty than all other electromechanical components combined. To address data‑scarcity for robot training, he proposed an “Optimus Academy” with 10‑30 k robots performing self‑play and millions of simulated runs.

He described three generations: Optimus 3 (million‑unit annual production) and Optimus 4 (ten‑million‑unit annual production), noting the lack of an existing supply chain will lengthen the S‑curve.

6. AI Alignment and Governance

Musk reiterated his belief that humanity cannot control AI far beyond human intelligence, but that AI should find humans “interesting” enough to preserve us. He mentioned xAI’s work on an “AI thought debugger” that traces decisions to neuron‑level causes, and praised Anthropic’s explainability work.

“I don’t think humans will ever control super‑intelligent AI.”

7. US‑China Manufacturing Competition

Musk argued that without breakthrough innovations, China will dominate global manufacturing because its ore‑refining capacity is roughly twice the rest of the world, it controls 98 % of gallium (key for solar cells), and its electricity output is projected to be three times that of the United States.

He highlighted that the U.S. has only one positive‑electrode material refinery, while China’s low‑cost solar cells (≈ $0.25‑0.30 /W) undercut U.S. production, especially given high tariffs on imported solar panels.

8. Management Practices

Musk described his weekly engineering reviews, “skip‑level” meetings where sub‑subordinates report directly, and a deadline‑setting philosophy that treats a five‑year task as if it will expand to fill five years, aiming for a 50 % probability of on‑time delivery.

His hiring principles emphasize “extraordinary highlights” in resumes, trusting conversation over CVs, and valuing talent, drive, trustworthiness, and kindness.

9. Future Signals to Watch

Will chip capacity exceed power‑deployment capability by the end of 2026?

Will Tesla and xAI merge?

When will TeraFab’s first fab become operational?

The interview concludes with Musk’s pragmatic optimism: “Being optimistic but wrong is better than being pessimistic and right; at least you’ll be happier.”