Musk's Terafab: The Energy Bet Behind AI's Next Frontier

Elon Musk has a unique ability of translating science fiction into science fact, for a profit. Terafab, his latest plan to set up computing capacity, a large part of it in space, that needs all the electricity the US generates, draws on his previous visions: driverless cars, robotic assistants and reusable rockets. Yet, at its core, Terafab is not a semiconductor strategy, it is an energy bet.

Musk's argument is the world cannot feed AI's appetite for power, and the solution must be extraterrestrial. Not everyone agrees. Other AI infrastructure companies are seeking to go beyond the energy constraint through more conventional means. Nuclear plants are reliable, but they take a long time to build. Terrestrial RE involves low initial investment but tends to be less efficient. Capturing solar radiation in space poses engineering challenges, which when overcome can provide a limitless energy source.

Musk reckons he can cross the technical hurdle because no new physics is involved. He needs to think bigger than the competition to establish his underlying assumptions about Terafab. Robotaxis need many more chips than existing Teslas'. The company's Optimus robot is being pitched as a solution to labour shortage. SpaceX is transiting from the interplanetary travel business into operating data centres in space. All of this involves quantum jumps in computing within Musk's business empire, leading inexorably to Terafab's chip manufacturing ambitions.

Nothing like this has been attempted before in semiconductor fabrication. Not only is Terafab going to pull the entire chip ecosystem—logic, memory, packaging and testing—under one roof, it will by its scale upend supply chains in the global semiconductor industry. Chipmakers have baulked at Musk's demands due to capacity constraints in their vendor trains. Those constraints will apply equally to Terafab. Machines to make chips are not keeping pace with existing AI demand. Musk will stress the industry even more.

Then there is the question of going to space. Deploying solar panels in space to generate the energy Musk seeks is sound theoretically but has not yet been tried out. Additional engineering advances are needed for in-space manufacturing, robotic assembly and in-orbit servicing. Critically, though, the cost of reusable rockets must fall appreciably for the idea to make economic sense. SpaceX's Starship has brought down the cost of placing payloads in orbit into the realm of possibility, not feasibility.

There is enough evidence, including Musk's record, to suggest Terafab may not deliver to timelines drawn for investors. Terafab does not exist in isolation; however, the AI infrastructure build-up is surging. Microsoft, Alphabet and Amazon are committing investments of similar magnitude as Terafab. Only, they are exploring more conventional approaches to scaling up. The underlying thread in AI infrastructure remains the energy bottleneck. Where do you source the power?

Even on a non-Musk scale, the industry needs solutions soon. Data centres are consuming just a shade under 2% of the world's electricity but their share is growing exponentially. AI models improve with computing, data and parameters. Training large AI models burns prodigious amounts of energy even when operating at 1/10th of the capacity Musk is looking at. This demand will keep rising even if the world is not swamped by Tesla's self-driving cars or helpful robots.

The choice is about solving the energy issue on earth or seeking it somewhere else. Neither is an easy choice. Say, Musk gets everything right. One man would then control the chips, rockets, orbital networks and robots that complete the AI loop. And this would be without any governance framework.

If Musk is being too ambitious, the world will have to reapply itself to the issue of climate control. All the additional energy will not be from clean sources. If we are partially successful with both approaches, new structures will evolve for AI to shed its terrestrial compute ceiling.

Musk's reimagining of the semiconductor industry will diversify supply chains and lower geopolitical risks. The US, China, the EU and even India are prioritising stability of their semiconductor supply chains. The AI boom has filled out order books across the industry, although advanced computing chips by Nvidia, AMD and Intel are leading the surge.

Intel, which is a partner in Terafab, expects structural changes in the industry from the way Musk is reimagining semiconductor fabrication. He is demolishing silos that will change how fabs operate in the future. The chips Terafab intends to make are customised for specific purposes—terrestrial and space—in a vertically integrated ecosystem and that sets them apart from other AI infrastructure companies who have to depend on the market.

Most speciality chips like those used in Tesla cars have a single optimised design, a specialisation Terafab plans to convert into industrial strategy. This, in itself, is a qualitative leap. Securing assured supply of chips whose prices should drop as they are produced in-house will speed up business transition plans for Tesla and SpaceX. Having a clear path to critical hardware should come in handy for Musk to bring investors on board his version of Starship Enterprise.