Original Core Argument

AI data centers crowding out HBM capacity causes consumer memory prices to surge (DDR5 +300%, DDR4 +172%). Geeker’s old devices are not on current supply chain, becoming “anti-inflation assets.” Three recommended items: M1 headless MacBook (<1000 RMB), 2080Ti modded 22G (1800 RMB), Angle Cloud/Wanke Cloud (40-200 RMB). Judgment: LLM has no moat, edge inference is the endgame, bubble will burst.

Fact-Checking

ClaimVerificationSource
DDR4 spot price surged 172%✅ Confirmed: Jan 2026 vs Dec 2025Cailian Press
DDR5 price up 300%+✅ Confirmed: Since Sep 202521 Jingji
Micron stock up 600% in one year⚠️ Underestimated: ~$93 a year ago, ~$864 now, ~830% gainYahoo Finance
Memory shortage until 2027✅ Confirmed: Morgan Stanley says 2-3 yearsInvestopedia
HBM crowding out consumer memory✅ Confirmed: HBM4 unit price $560, DDR4/LPDDR4 wafer production decliningEE Times China
256G DDR5 server memory >40K RMB✅ ConfirmedJW View, Zhihu

Logic Analysis

What the Article Gets Right

  1. Memory price chain is correct

    • HBM demand → capacity shift → DDR5 up → DDR4 follows up
    • Transmission logic is clear and data-supported

  2. “Geeker anti-inflation” angle is interesting and real

    • Old devices not on current supply chain, unaffected by price surge
    • DDR4 from 155 RMB to 897 RMB, ~6x increase
    • This is a real “alternative asset”

  3. Edge inference as endgame judgment makes sense

    • Open-source models getting stronger (Qwen, Llama, etc.)
    • Quantization, distillation, MoE lower local deployment barriers
    • Privacy, latency, cost triple drivers

Where the Article Falls Short

  1. “LLM has no moat” is too absolute

    • Model capabilities can catch up, but data flywheel, ecosystem, user stickiness can’t
    • OpenAI’s moat isn’t GPT-4 itself, but user habits and API ecosystem

  2. “Bubble will burst” is too optimistic

    • No timeframe given
    • Doesn’t consider AI may create new demand (not just replace old)
    • Historical “bubble burst” predictions are often too early

  3. Geeker item values are overstated

    • 2080Ti modded 22G can run 7B models, but not larger ones
    • M1 MacBook 8GB struggles with local models
    • These are “usable,” not “good”

Conclusion

This is an interesting “non-mainstream” perspective article with solid factual foundation (memory price data accurate) and coherent logic. Its greatest value is explaining AI supply chain cost transmission clearly — from HBM to consumer memory, from data centers to geeker’s second-hand market.

The “edge inference is endgame” judgment deserves serious consideration, but “LLM has no moat” and “bubble will burst” are too absolute.

One-line summary: Memory price chain logic is clear; edge inference judgment has value; but “bubble burst” is too optimistic.