The Anthropic Signal — Samsung Foundry's 2nm Anchor Book Filled Quietly While the Market Watched Burry's Short

Two headlines dominated Korea semi coverage this week. June total exports crossed $100B for the first time, and Michael Burry expanded his Korean semi short. Both stories were about memory — one riding May's $29.4B semi exports (YoY +154%) and the DDR5 16Gb spot at $46.633, the other framing the combined ~240 trillion won of SK Hynix and Samsung memory capex as cycle-peak overbuild.

But the week's real re-rating catalyst arrived quietly behind those memory headlines. At SAFE Forum 2026, Samsung Foundry named its 2nm anchor customers publicly for the first time. Tesla is in mass production on 2nm, multiple customers are ramping on 4nm — the beginning of a shift toward the named-customer disclosure discipline that TSMC has long used to justify its valuation premium.

A second signal arrived on a separate track. Samsung Foundry and Anthropic are in early-stage discussions for 2nm foundry work and advanced packaging, per thelec. That Anthropic is planning custom AI silicon is itself notable; the more important fact is that Anthropic is talking to Samsung, not to TSMC. Read against the same week's report that Google is shifting its AI-chip partnership away from the US (likely Samsung's Texas fab) toward Taiwan (likely TSMC), the high-end AI foundry market is bifurcating in a specific way. If TSMC is one axis, Samsung is building the other with Tesla + Anthropic talks + Intel legacy volume.

A third signal came from the supply chain. Intel is discontinuing legacy CPU lines and shifting related orders to Samsung Foundry. That means it is not just marquee anchor names — mid-to-low volume customers are accumulating too. Samsung Foundry's problem was never 'yield versus TSMC' — it was 'who actually mass-produces here?' Tesla + Anthropic talks + Intel legacy stacking in a single week is the beginning of the answer.

A fourth signal was roadmap reaffirmation. Samsung reconfirmed SF1.4 (1.4nm) mass production for 2029 and SF1.4+ for 2030. Not walking back the roadmap despite the Google defection tells you management believes SF2/SF2P anchor wins are sufficient to justify SF1.4 development spend.

So what is Burry shorting? His thesis is about memory — SK Hynix's 100 trillion won M17 NAND fab in Cheongju, 20 trillion won for the P&T7 packaging facility, plus Samsung's 140 trillion won Chungcheong HBM hub. Around 260 trillion won of announced capex that he reads as cycle-peak overbuild. That thesis deserves its own test, but the point for today's report is this: the foundry anchor story and the memory capex bubble story sit on different tracks. Burry is not shorting Samsung Foundry. The mispricing appears where the market conflates 'Korea semi = memory' and treats both stories as one.

The export numbers support the dual-engine framing. March semi exports $24.9B (YoY +138%), April $25.2B (+158%), May $29.4B (+154%). Three consecutive months this steep is hard to explain with memory ASP alone; a real AI foundry contribution is showing up. And with total June exports crossing $100B for the first time and semiconductors as the single largest driver category, June semi is very likely above the May $29.4B print.