Tech • AI • Robotics • Game

ENFR

Elon Spills the Beans In New Interview (Supercut)

9.4/10
SpaceXSolving The Money ProblemNovember 19, 2026 at 02:15 AM24:14
Audio player

⏱ ≈ 4 min read · 24-min video — 20 min saved

TL;DR

SpaceX, Tesla and affiliated AI and robotics efforts are pushing toward full reusability, large-scale autonomous systems, advanced chip production and massive compute expansion, with the common goal of sharply lowering costs and accelerating deployment.

KEY POINTS

First Starship upper-stage recovery

The catch of a Starship upper stage marked the first time the full vehicle architecture has been flown to space and recovered in reflyable condition. SpaceX had already completed three booster catches, including a reflown booster, but the ship recovery closes the last major technical loop for a fully reusable system. The recovered ship reportedly needs only minor repairs, though it is expected to be preserved because of its historic status.

Why reusability matters

Full and rapid reusability is expected to cut the cost of access to space by about 100-fold, with the possibility of eventually reaching 1,000-fold reductions. That change would enable payload delivery on the order of millions of tons per year, a scale viewed as necessary for permanent industrial settlements on the Moon or Mars. The broader argument is that off-world self-sufficiency would require rebuilding the entire industrial chain, from mining and refining to manufacturing finished goods.

A break from the Space Shuttle model

The reusable model being pursued differs sharply from the Space Shuttle, which reused some components but discarded the external tank on every mission and required lengthy refurbishment. Shuttle orbiters often needed around nine months between flights. Starship is intended to operate more like an aircraft, with a booster able to land, receive another upper stage and launch again rapidly.

AI training infrastructure

A major focus is the software stack used to coordinate very large clusters of GPUs, where the challenge is increasingly about networking and power management rather than just raw chips. Expertise developed in Starlink networking is being applied to AI training, on the premise that synchronizing hundreds of thousands of GPUs is fundamentally a communications problem. The current model line is in reinforcement-learning work, with iterative public-facing updates used to gather performance data and refine future versions.

Compute build-out at unusual scale

AI compute capacity is being added at roughly 0.5 gigawatts per month, described as equivalent to about 220,000 GB300s plus CPUs, networking and cooling. The stated goal is to exceed 1 gigawatt per month by the end of next year. Power supply, turbine stability and grid smoothing are central constraints, with Tesla Megapacks playing a role in buffering fluctuating training loads.

Capex efficiency and new fab ambitions

A new semiconductor manufacturing effort aims to build one of the world’s largest and most advanced chip plants, ultimately projected at 100 million square feet. A research fab is targeted to begin operating around June next year, with a first test chip around that time and phase-one volume production by late 2028. The strategy emphasizes faster iteration and greater willingness to alter production equipment than traditional chipmakers, whose large capital exposure can make them more risk-averse.

Optimus and authentic robotics demos

Optimus is being positioned as a general-purpose humanoid robot, but its developers say useful real-world autonomy remains a hard problem, especially in the hands and forearms where fine motor control is concentrated. A public demonstration is expected only when the system can perform unscripted tasks rather than staged routines. Handling everyday chores such as cleaning a dirty plate, washing it and putting it away is still beyond current humanoid robots, though that capability is projected within 12 to 18 months.

Industrial first, consumer later

The earliest strong economic use case for Optimus is expected to be heavy industry, particularly mining, ore refining and other 24/7 repetitive work. That focus is tied both to productivity and to U.S. concerns about dependence on foreign mineral processing. Home use remains a longer-term objective, with envisioned roles ranging from household chores to tutoring and child supervision.

Autonomous transport rollout

The rollout of Cybercab and broader robotaxi services is proceeding cautiously because vehicles without a steering wheel or pedals offer no human fallback. Operations are already active in Austin, with expansion planned across other large Texas cities and then Florida. California approval is being targeted for early next year, with the central priority remaining safety and regulatory acceptance.

Orbital AI compute

A proposed orbital compute platform, referred to as Star Mind, is framed as simpler than the current Starlink V3 communications satellites. The concept relies on large solar arrays generating about 250 watts per square meter, radiators shedding roughly 1,200 watts per square meter, and laser links routing data through the Starlink network. With more than 11,000 satellites already in orbit and about 30,000 space lasers in service, the argument is that the hardest satellite engineering problems are already being solved in existing systems.

CONCLUSION

The projects span rockets, AI, robotics, autonomy and chips, but the common thread is industrial scale: lower unit costs, faster iteration and tighter control of infrastructure. If even part of those targets are met, the result would reshape both access to space and the economics of advanced computing.

Ask a question

More from SpaceX