Don't Be Fooled by the AI Speed Control Theory: AI Data Centers to Triple Explosively; Why Money Is Flowing into Optical Communications, Substrates, and Power | SemiCon Research Lab CEO Noh Geun-chang
[#Global Insight] Don't Be Fooled, AI Speed Control Theory… AI Data Centers Triple in Explosion / Why Money Is Flowing into Optical Communications, Substrates, and Power | SemiCon Research Lab CEO Noh Geun-chang
Speakers
Noh Geun-chang
— Center Head, Hyundai Motor Securities Research Center
Summary
Noh Geun-chang, CEO of SemiCon Research Lab, reviews macro noise, OpenAI Astra, and the AI data center buildout. He argues that AI demand is shifting the bottleneck toward power, data center interconnection, and semiconductor substrates. He highlights Marvell, Cisco, Bloom Energy, and several substrate suppliers, while warning that some second-tier fuel-cell names still need earnings proof.
- Macro concerns eased despite a likely 25bp rate hike and lower oil prices.
- OpenAI Astra and agentic AI are expected to lift token demand and semiconductor demand.
- AI data center power demand may triple by 2030, creating a power bottleneck.
- Scale-across data center interconnection is emerging as a new AI infrastructure bottleneck.
- Marvell and Cisco are positioned as key networking beneficiaries.
- Substrate suppliers such as Ibiden, Samsung Electro-Mechanics, and Isu Petasys are supply-chain focus points.
- Bloom Energy is profitable and growing, while Doosan Fuel Cell needs earnings verification.
- The speaker favors tracking AI infrastructure leaders and buying on macro-driven dips.
Ideas
Agentic AI lifts semiconductor demand
Weakness in memory semiconductor stocks tied to crypto-market trading and stablecoin collateral is mostly retail/sentiment-driven and not a real institutional signal. The Anthropic CEO's comments should be read as a push to tighten China-related export controls and compliance, not as an AI slowdown, so any crypto-driven selloff in memory semiconductors is likely temporary and the uncertainty should fade.
AI data center power demand triples
AI data center power demand is becoming a major bottleneck: 1GW of data center capacity is comparable to a large nuclear reactor, current US data center power demand is about 40GW, and Microsoft plans to triple its capacity by 2030. With OpenAI Astra and agentic AI potentially lifting token output by around 50%, US data center demand could reach 120GW by 2030, requiring large power infrastructure investment and creating waiting lines for available power.
Marvell leads AI scale-across networking
Marvell is the best-positioned company in scale-across/data center interconnect, with about 60% DSP market share and a broad portfolio spanning DCI, switches, storage controllers, and custom silicon. Google is adding Marvell and MediaTek for TPU customization as it diversifies away from Broadcom, and Marvell has joined Nvidia's NVLink Fusion ecosystem. Data center revenue rose 46% YoY, leaving continued earnings momentum likely as AI data centers interconnect.
MediaTek gains Google TPU role
Google added MediaTek as a TPU custom-silicon supplier starting with TPU v8, and later added Marvell, because demand is so large and Google wants to reduce dependence on Broadcom's pricing and control. MediaTek therefore has a company-specific role as a beneficiary of Google's multi-sourcing in AI ASICs.
Broadcom loses Google TPU share
Broadcom is losing relative appeal in Google's TPU custom-silicon supply chain because Google is diversifying to MediaTek and Marvell, partly due to Broadcom's dominant pricing/control and slow responsiveness. The stock has been sideways while Marvell has outperformed, making Broadcom less attractive on this specific AI ASIC theme.
Cisco is AI infrastructure revival
Cisco is being revived as an AI infrastructure company: AI infrastructure orders are up 4.5x, data traffic tied to ChatGPT/agentic AI is up 450%, and it is a scale-across beneficiary with its own switch chips designed internally and manufactured by TSMC. Cisco is a top-two global networking vendor, and while the speaker does not recommend buying immediately, he sees it as a long-term tracking candidate and a buy-on-macro-weakness name.
Ibiden supplies Cisco networking substrates
Ibiden is the substrate supplier for Cisco's internally designed networking switch chips, so Cisco's cumulative one-million-unit chip shipments and AI networking growth should support Ibiden's substrate demand. The speaker also notes Japanese government subsidies for such Japanese materials and parts makers, aiding capacity buildout.
Samsung Electro-Mechanics gains substrate traction
Samsung Electro-Mechanics is following Ibiden in networking-chip substrates, making various efforts with customers and actually gaining traction, so it can benefit from the same AI networking substrate demand that is lifting Ibiden.
Isu Petasys is key Cisco PCB supplier
Isu Petasys remains focused on networking-equipment MLB, unlike Samsung Electro-Mechanics and LG Innotek which exited that business, and has the best technology with Cisco as its most recognized customer. As Cisco's AI networking and switch-chip volumes grow, Isu Petasys should benefit as a key networking PCB supplier.
Bloom Energy powers AI data centers profitably
Bloom Energy offers on-site solid-oxide fuel cells for data centers with the fastest installation time, and it is showing real profits: revenue is expected to double and operating profit nearly quadruple. Customers include Oracle and Equinix, with plans to add more hyperscalers and data center developers, making it a leveraged play on AI data center power shortages.
Doosan Fuel Cell needs profit proof
Doosan Fuel Cell is another hydrogen fuel-cell play on data center power and the direction is correct, but it is still loss-making and analysts' forecast of a breakeven around next year needs verification. The speaker warns that second-tier companies often rally and then fall, so he prefers to wait for earnings confirmation before treating it as a clean long.
Up Next
This 815 Money Talk (815머니톡) video, published September 14, 2026, features Noh Geun-chang discussing SMH, AI-SECTOR, MRVL, 2454.TW, AVGO, CSCO, 4062.T, 009150.KS, 007660.KS, BE, 336260.KS. 11 trade ideas extracted by AI with direction and confidence scoring.
Speakers: Noh Geun-chang · Tickers: SMH, AI-SECTOR, MRVL, 2454.TW, AVGO, CSCO, 4062.T, 009150.KS, 007660.KS, BE, 336260.KS