Ideas
Inference demand keeps AI chip demand intact.
The frontier-model slowdown debate applies mainly to training, not inference. As agentic/reasoning AI grows and recursive self-improvement raises compute intensity, inference demand for GPUs, MPUs, network chips, cables, memory, and server racks should keep growing independently. Even if frontier training growth moderates from a 45-degree to a 35-degree path, inference compute and memory demand remain supply-constrained, so AI semiconductor demand is not peaking.
Samsung gains from memory tightness.
Apple's reported approach to Samsung for DRAM at up to 30% higher prices indicates server-driven memory tightness even for the largest smartphone maker. Samsung's HBM progress is a key positive premise, giving it leverage to memory pricing and AI memory demand.
AI rules favor Alphabet and Microsoft.
AI regulation could become an entry barrier that entrenches incumbents. If leading model companies pre-design rules, capital-rich, vertically integrated platform companies like Alphabet and Microsoft are better positioned to build infrastructure, pursue M&A/partnerships, and comply, so the regulatory shift is not bad for them.
Samsung HBM outsourcing drives SFA turnaround.
Samsung's HBM success should force it to outsource general DRAM post-processing. SFA benefits as an OSAT: Samsung's HBM capacity increase can expand outsourced DDR5 testing/packaging, especially in the Philippines, and SFA's volumes are rising. SFA is loss-making now but should turn profitable next year; if Samsung HBM does well, the stock can target around 9,000 won and 12,000 won by the year after. Risks are Samsung's outsourcing policy and potential in-house Vietnam expansion, but near-term competitive damage looks unlikely.
Doosan Fuel Cell wins on data-center power.
Data-center onsite power demand makes fuel cells attractive: fast supply, decent output, low carbon, easy construction, and flexible modular scale. PAFC is a proven technology, and Doosan Fuel Cell is the representative PAFC player. If it wins orders properly, target prices point to 58,000-70,000 won; for a conservative entry, the speaker suggests using a deep discount to target. Near-term fuel-cell momentum is attractive.
Data-center power infrastructure demand grows.
Data-center power demand is not only about electricity; it requires transformers, transmission, and distribution equipment. Grid interconnection and equipment lead times are bottlenecks, making power infrastructure a growing investment need as AI data centers expand.
Nuclear is long-term data-center power solution.
Data centers require large-scale, reliable power. While fuel cells and LNG can bridge intermediate needs, the long-term (3-5 year) solution for large-scale power is nuclear, complemented by renewables and ESS. Nuclear has long build times, but it is necessary for high power capacity.
LNG bridges data-center power demand.
For data-center power, LNG gas turbines are a realistic intermediate complement to nuclear because they offer larger scale and flexibility than fuel cells and can be deployed near data centers. In a mixed power buildout, LNG gas turbines remain practically advantageous even with higher CO2.
SOFC is long-term fuel-cell winner.
Solid oxide fuel cells have better electrical efficiency and are the long-term technology direction, even though they are less mature than PAFC. As data-center power demand and technology advance, SK Eternix and MiCo's subsidiary are better positioned over time.
This 3PRO TV (삼프로TV) video, published September 15, 2026,
features Kim Jang-yeol
discussing SMH, 005930.KS, GOOG, MSFT, 036540.KQ, 336260.KS, Data center power infrastructure, URA, LNG, 475150.KS, 059090.KQ.
9 trade ideas extracted by AI with direction and confidence scoring.
Speakers:
Kim Jang-yeol
· Tickers:
SMH,
005930.KS,
GOOG,
MSFT,
036540.KQ,
336260.KS,
Data center power infrastructure,
URA,
LNG,
475150.KS,
059090.KQ
