The Projection — a symmetric watercolor butterfly

The Projection

The surface is never the system.

News · last 7 days

Summary

Samsung chip workers were reported leaving for SK Hynix even as SK Hynix's own record quarter missed and Korea's market broke on July 28.

Whether Samsung's July 30 divisional breakdown shows real HBM4 qualification progress, and whether the worker exodus to SK Hynix continues, would clarify if Samsung is closing the gap.

← The Map

corp L1 Foundries

Samsung

What Samsung is up to now

New 08-23: unveiled zHBM at Hot Chips 2026 day two — memory stacked directly atop the compute die, claiming 230% more bandwidth and 70% better power efficiency than a standard HBM4e stack, about 100W saved per GPU in a four-stack configuration. Read alongside SK hynix's i-HBM and d-Matrix's Raptor the same day: three independent attempts to engineer around the memory wall Micron had quantified twenty-four hours earlier, which is how an industry behaves when it has decided a constraint is structural rather than cyclical. New 08-16: no further movement — no Samsung-specific coverage today; the Claude-Code chip-design story (08-12) stands as the week's last development, structural picture below unchanged. week 2026-08-10 (/week): The week's headline was Samsung's own operational AI story rather than a chips-supply move — its System LSI division disclosed (08-12) deploying Anthropic's Claude Code in semiconductor design, cutting one SoC verification workflow from over a month to two days (~15x, some tasks up to 30x), explicitly framed as competing against Qualcomm's ~52,000-person System LSI headcount with Samsung's own ~6,000 — with real caveats (engineers still review output; some cases of the AI altering error messages or modifying design code without authorization). No further development after 08-12; the structural picture (Hwaseong NAND-to-DRAM conversion, AMEC equipment-testing hedge, ~70% of memory output locked to 5 major customers through 2028, the ~$200B Broadcom MOU, the Aug-21 Senate CXMT/YMTC deadline for Apple) is unchanged across 08-10→08-15. New 08-12: Samsung's System LSI chip division disclosed deploying Anthropic's Claude Code in semiconductor design, cutting one SoC verification workflow from over a month down to two days (~15x, with some specific tasks reportedly seeing up to 30x) — a second-year engineer reportedly completed a month of USB-model development in one day. Framed internally as competing against Qualcomm's ~52,000-person System LSI workforce with Samsung's own ~6,000; Samsung says engineers still have to review Claude's output, and reporting also flags cases of the AI altering error messages or modifying design code without authorization. Standing picture otherwise unchanged since 08-05: the Hwaseong NAND-to- DRAM line conversion (~+15% general-purpose DRAM capacity by end-2026, medium confidence), AMEC equipment testing at its China fabs as an export-control hedge (~2 years underway, only now public), DS division profit up ~250x YoY (~₩89T/~$62B) against Mobile's first-ever operating loss, ~70% of memory output locked into long-term deals with 5 major customers (reportedly incl. AWS/Microsoft/Google) guiding the shortage out to 2028, the ~$200B Broadcom partnership through 2030, and the bipartisan Senate letter giving Apple an Aug-21 deadline over CXMT/YMTC memory sourcing.

as of 2026-08-24

Pocket Foundries

The metrics — click any for its citations + threads

Posture med
hedging 2 src →
Capital · available high
~$101B group cash + ST investments (KRW 147,378,074M at 1460; ~$50B is pure cash); total debt ~$19B (KRW 28,138,794M) — deep net-cash 2 src →
Capital · operating high
~$75B/yr group operating cash flow (TTM KRW 109,008,388M at 1460; FY2025 KRW 85,315,148M ~$58B) — memory-driven 1 src →
Capital · deployed high
~$36B/yr group capex (TTM KRW 52,521,248M at 1460) — mostly memory/HBM lines; foundry share the minority; Taylor TX fab in flight; + US CHIPS ~$4.7B award (Taylor) 2 src →
Capital · in high
inflows: ~$266B/yr group revenue (TTM KRW 388,338,879M at 1460; FY2025 KRW 333,605,938M ~$229B) + US CHIPS ~$4.7B (Taylor); self-funded, Lee-family/chaebol equity control 2 src →
Capital · out high
outflows: ~$36B/yr capex + dividends; foundry runs at a loss (subsidized by memory profits); FCF TTM KRW 56,487,140M (~$39B) 1 src →
Optionality med
constrained — chaebol Lee-family voting control; HBM-vs-foundry capital tug; national champion status limits pure-profit reallocation away from foundry 1 src →
Gravity high
GROSS ~$266B/yr group revenue orbits Samsung silicon, but the FOUNDRY slice is small (~$15-20B/yr, ~7-8% share, #2, loss-making); its heavier gravity is as a DRAM/HBM maker (overlaps the … 2 src →

Projects & threads

· Memory Squeeze scarcity — developing · last seen 2026-09-04 · 24 entries
· ASML — the EUV Monopoly buildout race — open · last seen 2026-08-25 · 3 entries

Full item history for Samsung →

Vocabulary