THE SHORT ANSWER
● Already owed — Meta - about $825bn disclosed, of which $355.7bn is leases and $454.5bn contracts
● Still being raised — SK Hyper - about $350bn sought for the first 5GW of 15GW
● Cash out this year — Meta, guided to $135-145bn of 2026 capex
● Scale against local baseline — SK - 15GW against a Korean national capacity of roughly 1GW
Two very large numbers landed in the same week and are being quoted side by side as if they measure the same thing. They do not, and the difference is the whole point.
| SK Hyper | Meta |
|---|
| Headline figure | ~$350bn | ~$825bn |
| What it is | Capital being sought | Commitments already disclosed |
| Status | In talks, per Bloomberg | Disclosed, up from ~$700bn in July |
| Composition | First 5GW of a 15GW programme | $355.7bn leases, $454.5bn contracts |
| Timeframe | Target 2035 | Multi-year, non-cancelable |
| Near-term cash | Not yet raised | $135-145bn capex in 2026 |
| Local baseline | Korea is ~1GW today | Existing global fleet |
Why the comparison breaks
Meta's figure is a liability. It covers leases and contracts already signed, non-cancelable, disclosed because accounting rules require it. If Meta's AI revenue disappoints, that $825bn does not shrink - that is exactly what makes it a risk worth reporting.
SK Hyper's figure is a request. It is capital the company is in talks to raise for one phase of a programme targeted at 2035. If the financing does not come together, the number simply never exists. Reported amounts already vary between outlets - around $349.4bn in some accounts - which is normal for a deal in negotiation and impossible for a disclosed obligation.
What SK's number does tell you
The useful comparison is not against Meta, it is against Korea. The country's entire current data centre market is roughly 1GW. The first phase is 5GW and the full programme 15GW, so SK is proposing to multiply national capacity by fifteen. That is a grid, planning and power-generation question before it is a financing one, and it is the reason the timeline runs to 2035 rather than 2028.
What Meta's number tells you
That the obligation book is growing faster than the guidance. It moved from about $700bn in July to about $825bn now - $125bn added in roughly three months - against 2026 capex guidance of $135-145bn on negative free cash flow. The forward commitments are accumulating at close to the rate of annual spend.
How to read the next one
Three questions settle most AI capex headlines. Is this money already committed or money being sought? Is it a single year of capex or a multi-year obligation? And what is the local baseline it is being added to? Apply those and the week's two biggest numbers stop being comparable, which is the correct answer.