Evergreen Capital
Technology Investor, Semiconductors & AI Infrastructure
Terafab's distant start and difficulty making memory at competitive yields are incremental positives for DRAM makers. Elon Musk said its first SRAM chip is slated for 2H28 volume production; that suggests logic first, not DRAM. DRAM names remain among the most disconnected AI bets: valuations imply a sharp 2027-28 cyclical decline despite growth and margins that should stay above prior peaks if AI demand holds. Ramping capacity before demand arrives risks oversupply. x.com
Agentic AI is driving a sustainable inflection in memory demand, so memory stocks still look extremely cheap and should re-rate, with near-term peak-pricing fears overblown. Smarter reasoning models finish agentic tasks more reliably in fewer steps and tend to be larger. Agents also run best with long-lived context held in memory as KV cache, avoiding costly GPU re-prefill; GPT-5.6 Sol prices cached input at $0.50 per million tokens versus $5 for regular input. x.com x.com x.com
China is not doing as well as many think, and its memory-production ramp is only potential. Micron, SK hynix and the DRAM ETF look insanely cheap. A quoted StockSavvyShay post reports Elon Musk's view that memory shortages will worsen in 2027, while Chinese LPDDR production might ramp in 2028 and HBM remains harder to scale. x.com
3D-stacked DRAM could constrain bit output even more than today's HBM, preventing memory oversupply beyond 2030. HBM demand already diverts capacity from conventional DRAM; logic-integrated stacks could add pressure if they scale around 2029-30. That comparison depends on the architecture being understood correctly, and memory makers have not settled the DRAM process nodes. x.com
Networking, power and compute look more durable than cyclical memory in the AI buildout. Lumentum's revenue compounded year after year across the 2010s and 2020s, unlike Micron's; AWS revenue also compounded. Lumentum, Nebius, Bloom Energy and Credo should heavily outperform SanDisk and Micron from here, though AI itself remains heavily doubted. x.com x.com
Memory remains structurally scarce because SK hynix says conventional DRAM and HBM are practically sold out through 2029, while customers request roughly twice the HBM volume it can produce. Long-term agreements cover more than half of 2027 and 2028 capacity, and the company expects the supply-demand imbalance to persist through 2030. x.com x.com
Prepaid contracts and demand-paced fabs weaken a broad 2028 memory-glut case. SK hynix says over half its 2027-28 capacity is signed with deposits; 'sold out through 2029' includes negotiations. Samsung targets 60-70% contracted capacity. Micron's three projects begin output in 2H28; tooling follows demand. Reuters says CXMT's 20-30%-yield HBM3E may ship this year to export-restricted Chinese buyers, but performs like older HBM3. Commodity DRAM risks 2028-29 oversupply; NAND could ease in 2H27. … Show more
The memory sector is finished because China's CXMT has advanced a 4F² DRAM architecture that could greatly increase density. Micron, SanDisk and other memory-stock holders will bleed for months; China will lead AI moving forward, and those stocks are already paying the price. x.com
Rising memory contract prices strengthen the case for memory stocks still far below their highs. JPMorgan forecasts Q4 DRAM prices up 13% and NAND up 18%; Micron is 28% below its 52-week high, Seagate 44%, Sandisk 47% and Western Digital 55%. x.com
Micron and SK hynix pricing in a memory shortage past 2030 extends the memory-cycle case beyond the Street's 2028 expectation. SK hynix's record quarter adds evidence: per Wall St Engine, revenue rose 328% to $78.9B, operating profit rose 632% to $62.8B, and operating margin hit 80%. Agentic and physical AI could add demand as they roll out. x.com x.com
Deployment limits will slow AI hardware spending and expose cyclical suppliers such as SK hynix. Amazon has hardware waiting for power, and Broadcom's CEO said some clients delayed shipments because they could not deploy. Cheap forward multiples can mark a cyclical peak as analysts extrapolate growth. The slowdown could have begun or be a year away; timing is uncertain, so reducing cyclical risk in favor of healthier, genuinely cheap names looks better. x.com x.com x.com
Micron's bet on 3D DRAM over rivals' 4F2 could give it an edge as AI memory demand rises. Micron expects shortages for at least two more years: HBM per GPU rose 3.6x from H100 to B300, inference memory per server doubled in three years, and agent context grows about 30x annually. It plans further 6F2 nodes, 1-delta DRAM in 2H27 and 400-plus-layer G11 NAND in 2027; processor-stacked vHBM is targeted around 2028. If Micron is right, rivals are spending on the wrong path. Overweight Micron. x.com
