반도체 대장주 '이 2개 주식' 지금이라도 사 모으세요 (안석현 기자 2부)

Tech / Market Themes Covered

• Shift in AI demand: learning → inference (output)

  • The discussion argues that AI semiconductor needs are increasingly moving from training (GPU-heavy) toward inference/output.
  • This shift changes which hardware becomes dominant.

• TCO (Total Cost of Ownership) driving custom silicon

  • Training on GPUs is expensive due to:
    • Infrastructure investment (TCO)
    • Power/electricity costs
  • This motivates growth in custom “A-grade” / in-house ASIC-like chips and other specialized solutions.
  • Broadcom is cited as a key leader in this shift.

• NVIDIA strategy timing “pushed back”

  • Even if NVIDIA remains #1, the expected market penetration window for ASICs (estimate mentioned: 30–50%) keeps getting delayed.
  • NVIDIA is described as aggressively and cleverly maneuvering, with fast platform/solution changes as the market transitions.

---

Key Product / Platform Details Mentioned

• NVIDIA memory + inference stack evolution

  • Earlier: NVIDIA dominance through pairing GPUs with HBM.
  • Later transitions for inference:
    • CPX (introduced at GTC last year) as an early performance/inference starting point
    • Then: acquisition of Grok
    • Followed by: scrapping CPX and switching immediately to LPX
    • LPX is framed around ultra-low-latency inference
  • The discussion references VeraRubin and claims LPX shows strong performance for both learning and inference (even if not shown “exactly” as described in the subtitles).

• Interconnect / system scaling: MBLINK and ICMS

  • MBLINK: designed to let massive infrastructure communicate in real time like “one body.”
  • ICMS: aimed at handling KV cache more easily, with KV cache expected to keep growing—driving dedicated storage proposals/usage.

---

Investment “Ideas” Emphasized (Not Formal Investment Advice)

• “Three ugly brothers” framing → memory and CPUs regain attention

  • The “three ugly brothers” are stated as: general-purpose DRAM, NAND, and CPUs.
  • Despite belief that they were past their prime, speakers say shortages are now appearing—starting around the 3rd quarter of last year—because inference changes demand patterns.

• Advanced packaging expected to become more prominent

  • As AI hardware transitions, advanced packaging is expected to matter more.
  • Intel/AMD are said to face revenue pressure due to CPU shortages.
  • Longer-term attention may extend to Intel’s advanced packaging and foundry services.

• Glass substrate / CXL angle returns

  • The speakers revisit glass substrates as an important next packaging-substrate innovation cycle.
  • Claim: semiconductor package substrates refresh roughly on a ~15-year cycle, and the next cycle is expected before 2030.
  • Intel is described as having pioneered early glass-substrate work (announced earlier) but delayed due to difficulty—now potentially restarting as conditions improve.

• Memory portrayed as the “persistent protagonist”

  • While the “main leading semiconductor player” may change over time (e.g., NVIDIA → Broadcom/Marvell → Intel/AMD), the argument is that memory remains consistently central.
  • Memory leaders named:
    • Samsung Electronics
    • SK hynix
    • Micron
  • Conclusion: “AI memory” is indispensable across learning, inference, and future physical/agentic AI scenarios.

---

“Next Wave” from the Software/AI Side: Physical AI and On-Device Requirements

• Physical AI / agentic AI may revive device-level AI

  • Subtitles claim “on-device AI” hype previously faded due to unclear real justification (example: limited practical use like travel translation).
  • For physical AI/agentic AI, utility is argued to differ—suggesting new device/service demand for chips.

• Edge constraints → low-power, low-latency, high-bandwidth memory/interconnect

  • Edge devices are battery-limited, so designs must move beyond “GPU + lots of power” toward low-power approaches.
  • Mentioned concepts/approaches:
    • “Many terminals” memory-like technologies described as low-latency high bandwidth
      • Subtitles mention “llwdrem”
      • Also likened to “mobile HBM”
    • PIM (Processing-In-Memory)
      • Described as moving certain fast tasks closer to/inside memory to bypass the main CPU
      • Uses a “spinal reflex” analogy

---

Review / Guide / Tutorial Signals

• This video is not a hardware “how-to” tutorial.
• It is mostly market/technology analysis plus investment commentary.
• Includes “advice” to investors not to react too sensitively to news (behavioral guidance rather than product instruction).

---

Advice-Style Points (Market Behavior)

• Don’t overreact to headlines/news
  • Example mentioned: during war/major news periods, stock performance and sentiment may diverge from expected EPS optimism.
  • Advice: emphasize more refined information and fundamentals.

---

Main Speakers / Sources

• Reporter Ahn Seok-hyeok (안석현 기자) (referenced as “Easy Keypost” reporter)
• Unnamed co-speaker / analyst
  • Referred to throughout as “you/they,” but not clearly identified by name in the subtitles.