AI scarcity moved to atoms: memory, GPU and wafer margins hit 58%-80%
sujingshen · x · 2026-10-31
Building on Orange AI's "AI bubble" essay, the author uses public financials to argue AI scarcity didn't vanish—it moved from intelligence to physical compute inputs, and money follows it.
Key points:
- Money flows to the scarcest layer: memory, GPU and wafer margins run 58%–80%; model providers are only starting to profit; application-layer winners are numerous but unevenly distributed.
- HBM consumes roughly 3x the wafer area per unit capacity vs DDR5; SK Hynix has locked over half of 2027–2028 capacity with 14 customers, effectively selling out DRAM/HBM through 2029; TSMC's Q3 revenue beat its guidance ceiling with demand "still ahead of supply."
- In China, the bottleneck is chips, not power: money flows to domestic compute, optical modules and cloud—Innolight earned RMB 22.2B in the first three quarters.
- The bubble will deflate scarcity premiums, not AI's underlying value; cheaper compute upstream benefits the downstream.
- Opportunities: enterprise customers, vertical industries, distribution first, paying markets, and selling tools to AI builders.
More from Venture
- AI's TAM Is Infinite, Our Imagination Isn't: The TAM Framework Wasn't Built for Market-Rewriting AI — YogeshMalik · 2026-10-31
- Mobile app with $141 monthly revenue sells for $2K in 68 days on TrustMRR — marclou · 2026-10-31
- CXMT validates G6 DRAM engineering samples in Q3, mass production expected within a year — zephyr_z9 · 2026-10-31
- Scratching His Own Itch: Dev Turns Unreliable Social APIs Into ReachSocial.dev — FewScarcity6957 · 2026-10-31
- Anthropic's S-1 risk factors: compute tops the list at 13.4%, edging out regulation — ayushtweetshere · 2026-10-31
- Analyst model projects Anthropic at $1.5T revenue by 2028, met with disbelief — bookwormengr · 2026-10-31