報告_Daiwa_台灣資料中心硬體_20260713

PDF 原檔：報告_Daiwa_台灣資料中心硬體_20260713_original.pdf

圖片清單（已驗證 2026-07-15）

原始內容

Taiwan Data Centre Hardware

Key takeaways from the SG/KL marketing trip

• Investors remain constructive on AI theme; are blaming the recent pullback on the positioning issue
• Component shortage turns fiercer in the short term; we expect price hike stories to draw most investors' attention
• We pick substrate/NYPCB as our top ideas on highest earnings upward revisions; Delta remains the name attracting highest queries

What's new: We share our latest marketing feedback from our Singapore and Malaysia trips during the week of 6-10 July.

What's the impact: Investors remains constructive on AI. Since our previous US marketing trip, mainstream AI names have corrected by c.30% over the past 2-3 weeks. However, based on discussions with investors, most remain positive on the AI cycle and related hardware names. The pullback was largely viewed as positioning-driven, given how crowded the names had become. Fundamentally, we believe the supply-demand imbalance is becoming more evident across the supply chain. While we see substrates offering the greatest near-term pricing upside among shortage-driven components, we also note that power semiconductors have recently joined the shortage bucket.

Investment thesis: price hike > spec upgrade. As agentic AI and inference demand gain traction, AI spending is likely to shift gradually towards ASICs and CPU servers from 2026, away from Nvidia GPU-led capex in 2023-2025. While total AI spending will likely continue to rise, second-order content growth driven by spec upgrades may moderate as spending mix shifts from training to inference. We remain constructive on the ongoing server spec-upgrade cycle, but we think earnings upside from price hikes may be more meaningful from 2H26 to 2027. In our view, much of the savings from the mix shift towards ASIC/CPU servers will be absorbed by shortage components such as memory and substrates. Accordingly, we believe investors should pay close attention to the substrate industry, where pricing momentum has recently accelerated. We expect a wave of earnings upgrades, similar to what we have seen for memory vendors from 2H25. Separately, while CPU server shipments may face some risk in 2026, given component shortages, long-only investors we met remained constructive on CPU-related names over the long term.

What we recommend: We remain positive on Taiwan data centre names, given the recent read-through from ongoing component shortages. That said, we are more positive on price-hike beneficiaries, which could see stronger earnings revisions than names benefiting mainly from spec upgrades, especially when shipment assumptions might face downside risk and AI spending shifts towards relatively low-end servers. We also believe cost savings from the mix shift will be offset by higher component prices. Against this backdrop, we turn more positive on the substrate sector and continue to prefer NYPCB (8046 TT, Buy [1]) as our top idea.

How we differ: We set out our current view on how ASEAN-based investors are positioned towards the Taiwan data centre hardware sector.

13 July 2026

Daiwa

Neutral

Positive

Key stock calls

Source: Daiwa forecasts

Sector stocks: key indicators

Source: Bloomberg, Daiwa forecasts

Taiwan Data Centre Hardware: 13 July 2026

Daiwa

We believe most investors maintain a constructive view on AI; short-term pullback should be more from positioning issue

Taiwan Data Centre Hardware: 13 July 2026

ASEAN marketing key takeaways

Key feedback from investors

Constructive view remains

Despite the c.30% pullback in share prices from the recent peak during our US marketing trip in the week of 22 June, we believe sentiment remains constructive. Of the investors we met, 70-80% remained positive on the current AI cycle and on stock performance. Most viewed the recent correction as primarily driven by positioning concerns rather than a deterioration in fundamentals. Importantly, we did not encounter meaningful bearish views pointing to a capex bubble or a sharp slowdown from here. Many valuation-sensitive investors said they would look to add back on a meaningful pullback.

We believe the positive stance remains anchored in recent developments across the AI supply chain.

Short-term updates

Component shortages appear to be worsening, as evidenced by more visible price hikes in ABF/BT substrates over the past 2-3 weeks. We also note that the global ABF leader has indicated an intention to exit the BT substrate market amid tight manufacturing capacity, driven by rising CPU demand and resilient AI server demand. In addition, TSMC foundry capacity and power semis, including MOSFETs and PMICs, remain in shortage. Based on our supply chain feedback, power semis have been massively mentioned over the past 1 month (vs. little from 3 months ago).

Investors continue to focus on which subsector offers the best risk-reward profile within the price-hike theme. We currently prefer substrates , given the recent pricing momentum and the clear supply-demand imbalance. We expect substrates to follow a similar price hike pattern as we saw in memory from 2H25, despite a lower magnitude from substrates. We expect the substrate sector to deliver the strongest price hikes among the key shortage stories, with the price hike trend potentially extending through 2027.

That said, long-term concerns remain.

Long-term concerns

As highlighted in our previous US marketing note on 30 June (Key takeaways from the US marketing trip), we believe capex from the top-4 hyperscalers is already nearing a constraint, with incremental increases unlikely to be funded solely by operating cash flow. Any further upside may need to come from equity/debt financing or VC funding. ASEAN investors are also concerned about the pace of AI monetisation, though the issue is more about construction capacity failing to keep up with token demand growth rather than demand weakness.

Our view remains unchanged. While investors are aware of these concerns, AI remains a highly disruptive technology, making it difficult to draw a clear line on capex sustainability. From our marketing feedback, Singapore investors appear less focused on the capex outlook, while Malaysia-based investors want more clarity on the broader picture and timing. We think this reflects a desire to trade around the theme tactically, even as they remain constructive on AI.

Kyber delay remains a key discussion topic

Most investors do not expect Kyber to launch alongside Rubin in 2027, and ASEAN investors were largely neutral on whether Kyber arrives in 2028. We think it is still too early to conclude, as there are real technical bottlenecks, including the orthogonal mid-plane. That said, we remain confident Nvidia will work through these issues, given the importance of sustaining its moat in AI.

Daiwa

We are relatively more positive than investors on Kyber and HVDC adoption in 2028

Taiwan Data Centre Hardware: 13 July 2026

Daiwa

As the market moves towards agentic AI and inference, mid-to-low-end compute, including CPU servers, should gain importance. Still, Nvidia will likely remain dominant in AI training, thanks to its stronger compute performance. With Rubin Ultra's die count already revised down from four to two per chip, rack-level power density would otherwise remain broadly flat from GB200 to VR300 if Nvidia does not transition to Kyber at the end of the cycle. In our view, Nvidia will do what it can to improve power density, and any current bottlenecks should be addressed over the next 12-18 months.

HVDC adoption

Delta remained the most frequently mentioned name during our meetings. Investor attention is centred on the pace of 800V HVDC adoption by Nvidia, especially after Semianalysis suggested adoption could be slower than expected in 2027.

Our view is that HVDC is aimed at IT racks with power consumption above 500kW, and potentially up to 1.5MW under 800V configurations. Current Vera Rubin racks consume about 200kW, or about 400kW, including 100% redundancy, still below the 500kW threshold. If die architecture does not further scale and the total chipset count per rack remains unchanged, HVDC adoption in Rubin Ultra could remain limited. Based on Delta's current assumptions, adoption in Rubin Ultra could still be above 20-30% in 2028. We believe it also echoes our view that we are seeing some adoption of Kyber from 2028.

CPU theme

Several investors also asked for our view on Aspeed and Lotes, reflecting ongoing interest in the CPU server theme over the long term. As mentioned in our previous marketing note, CPU servers can serve not only orchestration functions but also inference workloads. This view is increasingly well accepted among long-only investors, who appear willing to stay invested in CPU beneficiaries through the current component shortage cycle.

Key investment theses

Price hike > spec upgrade

From 2023 to 2025, demand for training drove meaningful compute spec upgrades as hyperscalers prioritised performance. Looking ahead, we think interest is shifting towards inference, which prioritises the cost performance ratio, especially with the emergence of agentic AI. As a result, content growth from further spec upgrades should slow versus the past two years, even as AI capex continues to expand.

We believe the near-term focus will shift towards price hikes in shortage components, as savings from the transition from training to inference are increasingly absorbed by constrained supply. In this context, we introduce substrates as a key beneficiary, with the most meaningful price hike opportunity likely playing out from 2H26 to 2027. Within substrates, NYPCB remains our preferred name. We also like the CCL sector on both price hike and spec upgrade trends, although the price hike magnitude should be slower than how we look at substrate now.

Key risk: lower-thanexpected end-demand from AI companies

Key risks

Key downside risks for the sector

Lower-than-expected end-demand from AI companies. We believe the key downside risk to our Positive stance on the Taiwan Data Centre Hardware sector is a slowdown in the end demand from leading AI companies, such as OpenAI, Anthropic, and xAI. These companies' ability to generate sustainable ARR through monetising foundation models and applications ultimately determines the pace of infrastructure investment. If ARR growth stalls, hyperscalers may temper their capex plans, leading to weaker-than-expected AI server shipments. In our view, this is the primary downside risk to our sector call.

Taiwan Data Centre Hardware: 13 July 2026

Daiwa

Daiwa's Asia Pacific Research Directory

Taiwan Data Centre Hardware: 13 July 2026

Daiwa

Daiwa's Offices

DAIWA INSTITUTE OF RESEARCH LTD

Taiwan Data Centre Hardware: 13 July 2026

Daiwa