Some thoughts on the "Meta" ghost story

Last night, Meta made a major announcement about its plan to rent out idle AI computing power, building its own AI cloud business.

Upon the news, Meta's stock price surged, while hardware stocks collectively pulled back.

I immediately wrote a simple post for members, saying if it were me, I would opportunistically and in batches bottom-fish equipment, storage etc., as a preliminary judgment that the market is overly emotional.

1.

The market's logic for selling off hardware is: if even Meta is starting to rent out computing power, it means there's an oversupply. In the future, spending on hardware like GPUs, HBM, and SSDs will decrease.

The market really proved my previous point — no matter what the asset is, excessive overbuying makes it sensitive to negative news.

Anyway, after a big rally, everything looks bad..

Actually, Meta renting out computing power is itself a positive development.

The market's concern was that Meta is investing hundreds of billions of dollars annually in building AI infrastructure but lacks a mature cloud business like Amazon AWS or Microsoft Azure, raising persistent doubts about its return on investment, which is why Meta's stock kept falling in the past.

Now, renting out idle computing power turns assets that were previously only for "own use" into assets that can generate continuous revenue, significantly improving the business model. This is the most direct reason for Meta's surge.

2.

Actually, Meta renting out computing power only changes the user of the computing power; it doesn't reduce the overall societal demand for it.

Previously, GPUs were used by Meta to train models; now they are rented to OpenAI, Anthropic, or other AI companies. Those GPUs are still working, and HBM, enterprise SSDs, and network bandwidth are still being consumed.

What determines the long-term demand for the AI industry is not the number of GPUs, but the growth of Tokens.

If AI applications continue to proliferate in the future and the daily global generation of Tokens continues to grow rapidly, then even with improved GPU utilization, the entire industry will still need to build more data centers and purchase more GPUs, HBM, and enterprise SSDs.

Remember the storage ghost story from before? It was that Google paper, the gist of which was that improved storage efficiency would lead the market to believe storage usage would also decrease, causing market panic.

The improved GPU utilization here can still be explained by Jevons paradox: technological efficiency improves, consumption per unit of output decreases, but total resource consumption actually increases.

There's another key point.

If Meta truly believed computing power was already in surplus, the most direct action would be to cut capital expenditure (Capex).

But the opposite is true. Meta still maintains its capital expenditure guidance of $125-145 billion, indicating the company itself doesn't believe AI infrastructure construction is over.

So, like the last storage incident, this is more of a "ghost story" — the rally has been too big, and excessive overbuying is sensitive to any market movement.

A pullback is definitely a good thing.

Renting out computing power changes the business model and resource utilization efficiency, not the AI demand itself.

3.

Whether this wave of AI revolution can continue still boils down to three core questions:

Will AI capital expenditure continue to grow?

Will Token computing power demand persist?

Are data centers still expanding?

If the answer is still yes, then we can still focus on the big picture and ignore minor details; the long-term logic hasn't fundamentally changed.

In the short term, the market will keep telling stories; in the long term, what truly determines the direction of the industry is always demand.

$Meta Platforms(META.US) $NVIDIA(NVDA.US) $Alphabet(GOOGL.US) $Sandisk(SNDK.US) $Micron Tech(MU.US)