What happened

SpaceX listed on the US stock market on June 12. The stock rose about 19% on its first day, and the company was valued at roughly US$2 trillion — putting it among the most valuable companies in the world.

Ten days later, the mood flipped. After SpaceX announced a large bond issuance, the stock fell sharply in a single day, and hundreds of billions of dollars came off its market cap. There was no fatal news about the business. The market's question was something else entirely.

SpaceX was already sitting on more than US$100 billion in cash. Yet it moved to raise even more, refinancing expensive legacy debt with cheaper bonds. Investors started asking: why does a company this rich need this much more cash?

The answer traces back to a plan unveiled a few months earlier.

Tesla and SpaceX are pushing TeraFab, a massive manufacturing project for next-generation AI chips — and Musk has floated the idea that, over the long run, much of that capacity could serve space-based data centers.

How far this plan actually goes is not yet confirmed. But investors have already started pricing the direction itself.

Why it matters — the landlord of the bottleneck

AI has plenty of bottlenecks right now.

GPUs, HBM, advanced packaging, cooling, the power grid, data-center land — all of it is scarce.

The two the market is hitting hardest are chip manufacturing capacity and power supply.

A bottleneck is not a difficult idea.

It's the point on a highway where the lanes suddenly merge into one. It doesn't matter how good the cars are — every one of them has to pass through that narrow stretch. And the toll isn't set by the people driving the road. It's set by whoever owns it.

AI works the same way.

If the narrowest lanes today are chip fabs and power, Musk has chosen a strategy of owning the lanes themselves.

TeraFab targets the chip-production bottleneck.

Space data centers are a way around the power bottleneck. Musk's logic is simple: in orbit, solar power can be harvested more steadily than on the ground, and the problem of securing land for massive data centers changes fundamentally.

The core is always the same.

What others buy at a premium, he tries to secure for less.

Electricity. Data-center land. Chip manufacturing capacity.

All of them are core assets of the coming AI era.

Musk always uses the same tools

His businesses change; his toolkit barely does.

• Reuse (don't throw the rocket away)
• Automation (factories work instead of people)
• AI
• M&A
• Vertical integration

The dream changes. The playbook almost never has.

There's precedent, too.

After SpaceX proved reusable rockets and found itself short of customers to launch, it built a new customer of its own: Starlink.

This time, the direction is reversed.

Customers who want AI chips are everywhere. The factories are not.

So he's going to build the factory himself.

The real target is physical AI

TeraFab's real meaning isn't AI servers.

It's AI with a body.

Robots and self-driving cars.

Musk expects that once humanoid robots like Tesla's Optimus reach mass production, they will need far more AI chips than the car business does today.

Morgan Stanley has likewise suggested that if Optimus enters mass production, AI chip demand could far exceed automotive demand.

Musk recently thanked Samsung Electronics, TSMC, and Micron — and then added one thing.

Their pace of fab expansion, he said, would struggle to keep up with his demand.

If the physical-AI era arrives the way he expects, the shortage won't be chip performance. It will be the factories that make the chips.

It's the scenario we've been tracking all along.

Bottlenecks are confirmed by sellers

But there's one rule that has to hold.

A buyer's words are not yet evidence.

A bottleneck is only confirmed when the seller says it.

The memory market is already showing that signal.

In its latest earnings call, Micron guided that supply would trail demand for some time, and said long-term contracts locking in future output had risen sharply.

The memory bottleneck is now a fact acknowledged by the supplier itself.

Robot-grade AI chip capacity isn't at that stage yet.

What we're hearing today is a buyer's forecast.

For that forecast to become reality, the suppliers — TSMC, Samsung Electronics, Intel — need to start telling the same story.

So how should we read Musk?

Musk is often late.

The Falcon rocket failed three times in a row.

Reusable rockets were called impossible.

Then they worked.

AI5, Tesla's next-generation AI chip, has also slipped about two years from its original schedule.

But the design is done, and production preparation is underway.

I've handled broadcast equipment for more than twenty years.

The performance printed in the catalogue and the performance delivered in the field were never the same.

So I don't take Musk's "space data centers get cheaper within two to three years" timeline at face value.

My own estimate is around 2030.

But it's also hard to deny that he is someone who arrives — late, but eventually.

The most honest sentence about Musk is this:

Delay is near-certain. Success is a probability.

And I still stand by this one:

Why this reading could be wrong

The most dangerous thing in investing is believing only your own story.

This analysis could certainly fail.

First, the economics of SpaceX's space data centers depend heavily on Starship. If Starship development slips badly or costs run higher than expected, the entire calculation changes.

Second, if chip supply expands faster than expected, the bottleneck itself weakens.

Third, if AI models get dramatically more efficient — the same performance on far fewer GPUs and far less power — today's bottleneck structure could change completely.

The opposite of getting there first isn't failure.

It's glut.

What to watch

Disclosure: The author holds positions in Tesla, SpaceX, and Samsung Electronics mentioned in this post.

For information only. Not investment advice.