Meta jumped roughly 9% this week on reports it’s building a cloud business to sell excess AI compute and hosted models to outside customers. Samsung and SK Hynix, meanwhile, are down sharply on the back of a $1.3 trillion, decade-long semiconductor investment blueprint. On the surface, that looks inconsistent — both are AI infrastructure spending stories, and the market is rewarding one while punishing the other.
But the more interesting read isn’t that investors are being illogical. It’s that they’re drawing a sharp distinction between two layers of the same buildout, and pricing them very differently.
Two Capex Stories, Two Very Different Receptions
Meta’s cloud pivot isn’t really a spending announcement — Meta already raised 2026 capex guidance to $125–145 billion, and that guidance had been weighing on the stock. What moved shares was the reframing: idle GPU capacity turning from a pure cost center into a potential revenue line, competing directly with AWS, Azure, and Google Cloud. The market is pricing “AI infrastructure optionality” — a credible near-term path to monetizing spend that’s already been committed.
Samsung and SK Hynix’s $1.3 trillion plan is the opposite shape. It’s capex with no attached monetization story — new fabs, HBM lines, and packaging capacity, funded largely by the companies themselves with government support, but no new customer or contract disclosed alongside it. It’s exactly the kind of pure spending announcement that would likely have hurt Meta’s stock too, had it not been paired with a revenue angle.
The Real Tension: Where the Market Thinks Margin Sits
The more interesting tension, if there is one: you could argue the market is currently pricing “spend on AI infrastructure to sell compute to others” as good, but “spend on AI infrastructure to make more chips for others to sell compute” as neutral-to-bad — even though Samsung and SK Hynix sit one layer further upstream in the same buildout Meta is monetizing.
That’s less “illogical” than it is a reflection of where investors currently think margin and pricing power sit in the AI stack. Hyperscalers monetizing GPU-hours look like they’re capturing value. Chipmakers pouring capital into a decade-long HBM and fab buildout — amid softening memory pricing signals and crowded positioning — look like they’re taking on the risk without a matching visible payoff yet.
In other words: it’s not that the market doubts the AI buildout. It’s that it currently rewards proximity to the customer relationship and monetization mechanism, and discounts capacity that sits further back in the supply chain, exposed to pricing and demand risk it doesn’t fully control.