Fear Is Not a Fact — Testing the Meta/SpaceX 'Surplus Compute' Narrative and Korea's Sidecar Crash Against the Data

What moved the market today was fear, not fact

On July 2, 2026, the KOSPI fell more than 6% intraday, tripping the sell-side sidecar, and closed -7.9% at 7,648, below 7,800. Samsung closed at KRW 286,000 (-9%), losing the 300,000 handle; SK Hynix fell -14.6% and Hanmi Semiconductor by double digits (see the three charts above — a broad chip selloff with Korea worst and Taiwan resilient, and a pullback after a huge run). Foreign investors dumped over KRW 2 trillion in a day, and leveraged semiconductor ETNs fell ~19%.

The trigger was two overnight US narratives. First, Meta signaled a cloud business ('Meta Compute') to sell its 'surplus' AI compute externally (TechCrunch), and its stock jumped ~9%. Second, days earlier, SpaceX (merged with xAI this year) had leased Colossus data-center capacity to Anthropic, Google and Reflection (Motley Fool). Retail read it as: capacity is 90% idle, so the shortage was a lie.

Those news facts are true. The problem is the next sentence — the conclusion that "therefore demand has peaked." That conclusion is not in the data.

This is capacity monetization, not a fire sale of surplus

Neither Meta nor SpaceX is abandoning infrastructure. They are monetizing idle cycles via a merchant-compute (neocloud) model. Morgan Stanley pegs Meta's rental revenue at, at most, ~$3 of 2028 EPS — trivial against Q1 2026 EPS of $10.44 — i.e., capacity optimization, not a capex retreat. SpaceX vacated Colossus 1 not because it was 'spare' but because mixed-generation chips (Hopper/Blackwell) bottlenecked Grok training, which moved to the larger ~1GW Colossus 2. Even that older infra is leased to Anthropic and Google for over $2 billion a month (CNBC) — a price impossible if compute were truly in surplus. The post's "90%" is unverified in any primary source; Anthropic's deal alone is ~300MW of SpaceX's ~1GW (~30%).

The fundamentals point the opposite way

The same-day facts in our database run against the fear. Samsung's HBM3E/HBM4 is sold out to Nvidia and the clouds through Q1 2027, and . And capital is expanding, not shrinking (see the capex chart above): the Big-4 hyperscalers — the chips' biggest buyers — spent ~$130B in Q1'26 alone, with Alphabet and Meta raising full-year guidance to $185B and $135B. On top sit Korea's ~$919B sovereign AI data-center program to 2035 and Samsung/SK's ~KRW 800T in new fabs. Nomura is warning of an "epic" supply crunch in 2H26 — not the picture of collapsing demand.