Ideas
AI speed-control fears are noise.
Global semiconductor capex is record and multi-year from TSMC and the memory three, so equipment suppliers should prioritize those customers over Chinese memory makers CXMT/YMTC; China's capacity expansion plans face US blacklists and an equipment bottleneck. Equipment makers therefore have a strong five-year order backdrop.
Samsung memory upcycle supports 440,000 won target.
Memory spot pricing is no longer a good barometer because spot is only 1-2% of the DRAM market; product crossover to HBM4, LPDDR6, DDR6 and GDDR7 should lift blended ASP, and leading suppliers' planned ASP should start outpacing market spot from Q3. Won strength may trim won-based estimates, but premium mix should keep operating profit rising. The cycle is longer because memory is increasingly sold in racks/systems to sovereign AI, neoclouds and enterprise customers, so valuation should blend PBR and PER rather than rely on PBR alone. His target is 440,000 won, with next-year book value around 180,000 won.
SK hynix target is 3.3M won.
SK hynix benefits from the same memory ASP/mix improvement and HBM margins: HBM long-term agreement price increases have lagged spot, which helps margins, and the cycle is longer as memory is sold in racks/systems and to sovereign AI, neocloud and enterprise customers. His target is 3.3M won; next-year book value above 1M won makes the 1.5M won area look less intimidating on a PBR 1.5x basis.
Cisco orders up 4.5x on networking.
Cisco's orders are up 4.5x because AI data centers need networking to connect with each other, not just chips; the buildout from 40GW to 120GW of US data center power capacity directly increases networking demand.
Data center power demand triples.
The speaker sees AI, power and energy as the main growth areas; US data center power use is about 40GW and internal analyses suggest it must reach 120GW by 2030, while Microsoft alone is tripling its 2GW AI data center footprint. That implies sustained demand for power/energy and data center infrastructure.
Hyperscaler free cash flow inflects next year.
Hyperscaler free cash flow should inflect as EBITDA grows: data center payback is 2-3 years, GPUs are pre-booked, operating margins are in the mid-to-high 30s, and financing activities bridge the capex gap. He expects the free-cash-flow transition to arrive sooner than feared, possibly next year.
Google JVs drive EBITDA inflection.
Google's two joint ventures, one with Blackstone and one with Blackstone, Apollo and Broadcom, could generate $50-60B of sales by 2027 and have a 2028 target of $250B at 60% operating margin, potentially turning Google EBITDA positive by Q2/Q3 next year.
Nvidia and custom ASICs both grow.
Nvidia's hyperscaler revenue growth may slow as Google TPU, Amazon Annapurna and other custom accelerators/CPUs gain share, but non-hyperscale data center demand from sovereign AI, neoclouds and enterprises is growing rapidly and accepts Nvidia pricing. As US data center power capacity triples, both Nvidia and the custom-ASIC camp can grow.
This 3PRO TV (삼프로TV) video, published September 14, 2026,
features Noh Geun-chang
discussing SMH, 005930.KS, 000660.KS, CSCO, AI-SECTOR, SKYY, GOOGL, NVDA, Custom AI ASICs.
8 trade ideas extracted by AI with direction and confidence scoring.
Speakers:
Noh Geun-chang
· Tickers:
SMH,
005930.KS,
000660.KS,
CSCO,
AI-SECTOR,
SKYY,
GOOGL,
NVDA,
Custom AI ASICs