AI Infrastructure Long Ideas — US & Hong Kong

Prepared: May 29, 2026

Direction: Long-only

Universe: US-listed + HK-listed AI infrastructure / AI-exposed names

Style: Quality + value + growth at intersections

Methodology: Sourced from AI Infrastructure Sector Overview (US+HK), May 29, 2026

---

Screen Methodology

The 2026 AI infrastructure backdrop favors:

1. Quality at temporary discount — names with strong fundamentals trading below peer multiples for transient reasons (regional discount, narrative gap)
2. Picks-and-shovels with multi-year backlog — power, networking, custom silicon — less narrative-dependent than pure GPU exposure
3. Pure-play AI revenue inflection with disclosed AI growth and reasonable valuation
4. Cross-geography value arbitrage — same theme, different multiple

Excluded: speculative IPO momentum (MiniMax-style), unprofitable cash-burning hyperscalers (CoreWeave/Nebius), generic LLM wrappers, names with high regulatory tail risk (BIOSECURE-exposed).

---

Idea Comparison Table

---

One-Pagers

Idea 1 — Eaton (ETN US) — LONG — Picks-and-shovels for the power-constrained AI buildout

Thesis:

• Power and grid interconnect — not chips — is the binding constraint for AI data center capacity in 2026. Eaton's electrical products segment is the most direct beneficiary.
• DC orders +240% YoY in Q1 2026; Boyd Thermal acquisition adds ~$1.7B run-rate; Ultra PCS adds power conversion capability.
• FY26 organic growth raised to ~10%; Electrical Americas raised to 13%.
• Trades at industrial multiple (~22–25x) despite owning best-in-class AI exposure on the demand side and ~12 years of order book on the supply side.
• Less narrative-dependent than NVIDIA — value accrues whether training or inference dominates.

What the market is missing: Eaton is treated as a cyclical industrial; the AI buildout transforms it into a multi-year secular grower with backlog visibility unmatched in tech.

Key Risks:

• AI capex slowdown 2027–28 → orders soften
• Execution on Boyd Thermal integration
• Margin pressure from copper/raw material inflation
• DOE/FERC interconnect reforms accelerate enough to ease bottleneck (would compress backlog premium)

Next steps: Full DCF model with backlog conversion sensitivity; channel checks with hyperscaler procurement.

---

Idea 2 — Broadcom (AVGO US) — LONG — ASIC inflection + OpenAI mega-deal underappreciated

Thesis:

• Inference is the 2026 inflection (>50% of compute already at CoreWeave); custom silicon is the natural inference architecture.
• Six confirmed ASIC customers including Google, Meta, OpenAI. The OpenAI/Broadcom 10 GW co-developed compute system is one of the most under-modeled deals in tech.
• AI revenue compounding +100%+ off a $33B+ run-rate base. Q2 guide $10.7B implies ~+140% YoY.
• 68% adjusted EBITDA margin = best-in-class semis profitability.
• Trades at meaningful discount to NVIDIA NTM P/E (~30x vs. 41x) despite faster AI growth and explicit $100B+ FY27 AI revenue target.

What the market is missing: Broadcom's AI exposure is treated as "second-tier ASIC" vs. NVIDIA; in fact it owns scaled relationships across all hyperscaler in-house programs simultaneously.

Key Risks:

• NVIDIA ecosystem moat persists in training; ASIC adoption slower than expected
• Concentration in Google ASIC (largest customer historically)
• VMware integration/divestiture noise
• Geopolitical disruption to Taiwan supply chain

Next steps: AI-revenue waterfall model (per-customer ASIC ramp); track GTC and ISC announcements.

---

Idea 3 — Micron (MU US) — LONG — HBM upcycle, real numbers, still cyclical-discounted

Thesis:

• Memory is the most leveraged play on inference scaling — every transformer model needs HBM for KV-cache and DDR for context.
• HBM demand pre-allocated near-fully to NVIDIA, AMD, hyperscaler ASICs; pricing power has shifted to memory makers for the first time in a decade.
• IDC: DRAM market triples to >$400B in 2026 driven by HBM and AI DDR.
• Mgmt guides ~20% bit shipment growth for both DRAM and NAND in CY2026.
• Cycle-discounted multiple (~15–18x NTM P/E) despite +196% YoY revenue and 75% gross margin tells you the market expects rapid normalization. If HBM tightness persists into 2027, multiple expansion possible.

What the market is missing: This isn't a 2018-style memory cycle; the HBM-AI demand vector is structural. Micron is competing in HBM with only 2 other suppliers (Samsung, SK Hynix) on a supply-constrained basis through 2027.

Key Risks:

• Sharp cycle reversal if NAND/DDR commodity prices collapse
• Hynix/Samsung HBM3E share recapture
• AI capex deceleration 2H26
• China DRAM (CXMT) catching up faster than expected on commodity nodes

Next steps: HBM supply-demand model through 2027; channel checks on hyperscaler memory pre-allocation.

---

Idea 4 — Zscaler (ZS US) — LONG — Cheapest growth-adjusted cyber + AI security re-rating

Thesis:

• AI workloads create new security perimeters (model APIs, agent permissions, prompt injection, data exfiltration). Zscaler is positioned as the AI-native zero-trust platform.
• AI security ARR exceeded its $400M FY26 target three quarters early; trajectory implies $700M+ FY27.
• Trades at 11.7x EV/NTM revenue — meaningful discount to CrowdStrike (25x) and Cloudflare (31x) despite comparable growth rates.
• AI-driven workload migration is a structural tailwind; ZIA + ZPA architecture aligned with "agent everywhere" deployment.
• Path to FCF margin expansion as scaling kicks in.

What the market is missing: Zscaler is still treated as the "post-RSA reset" name from 2024–25; the AI security wedge has materially changed the growth profile but hasn't been priced in yet.

Key Risks:

• Palo Alto / CrowdStrike platform consolidation pressure
• Sales execution — Zscaler has historically guided conservatively after a few cuts
• Microsoft Defender bundle erosion in SMB

Next steps: Quarterly AI security ARR tracking; CrowdStrike vs. Zscaler win-rate channel checks.

---

Idea 5 — Baidu (9888 HK) — LONG — Cheapest AI exposure on the planet

Thesis:

• Baidu has the cleanest AI revenue disclosure of any HK-listed mega-cap: AI cloud +79%, GPU cloud +184%, AI-powered ad services +36%.
• Core AI business has crossed the 50% threshold of core revenue — the company is now structurally AI-led.
• ERNIE 5.1 ranked #1 Chinese model on LMArena Text; Apollo Go achieving unit-economics break-even in Wuhan with 22M+ cumulative rides.
• Net cash >50% of market cap; ex-cash P/E in low single digits.
• Sentiment depressed on legacy search decline narrative — but the market hasn't caught up to the AI revenue inflection.

What the market is missing: Baidu trades like a declining search company while the underlying business is now an AI cloud + autonomous mobility company. Core AI biz growth (+49%) is faster than peers' total growth.

Key Risks:

• US-China relations / ADR delisting risk for ADRs (HK listing somewhat insulated)
• Domestic search decline accelerates (Tencent/ByteDance vertical search)
• Apollo Go regulatory scrutiny / liability events
• Stock-Connect flow reversal
• Below-historical multiple becomes a value trap if AI revenue mix-shift takes longer than 12 months

Next steps: SOTP model (search vs. AI cloud vs. Apollo Go vs. cash); track Apollo Go monthly ride growth and city expansion.

---

Idea 6 — Tencent (700 HK) — LONG — Quality compounder with hidden AI option

Thesis:

• The "ex-AI investment" margin (43%) exposes the underlying franchise quality — games, fintech, and ad services are all expanding margins beneath the AI capex drag.
• Hunyuan 3 (Hy3) deployed across 131 internal Tencent products — ad targeting, content moderation, gaming AI — with AI-boosted ad revenue already showing through.
• 2026 AI capex >RMB 36B funded by reduced buybacks signals confidence in deployment ROI.
• Trades at ~13–16x P/E and ~10x EV/EBITDA — defensive valuation for a company with mid-teens earnings growth + AI optionality.
• HK-line YTD performance (-3%) lags ADR (TCEHY) materially — suggesting flow-driven dislocation.

What the market is missing: Tencent is being valued on the depressed reported margin while AI capex is the largest reason for that depression. If Hy3 monetizes anywhere close to peers, the multiple should re-rate to 18–22x.

Key Risks:

• Domestic gaming approval freeze
• Regulatory action on social/messaging
• Slower AI monetization than Alibaba/Baidu
• Geopolitical / ADR-HK fungibility issues

Next steps: Build Hunyuan revenue ramp scenarios (bull/base/bear); track Q2 26 advertising acceleration.

---

Idea 7 — Lenovo (992 HK) — LONG — AI server share gains underappreciated by global investors

Thesis:

• Lenovo has quietly built one of the largest AI server franchises outside the hyperscalers, leveraging direct relationships with Chinese cloud providers + global enterprise.
• AI revenue +84% YoY = fastest-growing segment; now 38% of group at scale.
• Trades at industrial-like multiple despite mix shift to high-growth, higher-margin AI server line.
• Beneficiary of: domestic Chinese AI capex cycle, sovereign AI builds (India, Middle East), enterprise AI on-prem deployments (which still represent ~57% of AI infra spend per Mordor).
• Less crowded than Dell/SMCI on US side; not in major global indices the way Dell is.

What the market is missing: Lenovo's AI server traction has outpaced Dell's (DELL AI server $9B Q4 +342%, Lenovo group AI +84% on a larger base) but Lenovo trades at a discount because of HK-listing sentiment and PC-legacy framing.

Key Risks:

• US export controls broaden to AI servers shipped through Chinese OEMs
• PC cycle weakness drags group margins
• Memory/component cost pressure
• Currency volatility (USD-revenue, mixed cost base)

Next steps: Segment-level model with AI server unit economics; channel checks on Chinese hyperscaler procurement (Alibaba, Tencent, ByteDance).

---

Idea 8 — ASMPT (522 HK) — LONG — Cleanest leverage to TCB/HBM packaging cycle

Thesis:

• HBM3E and HBM4 require advanced thermo-compression bonding (TCB) — ASMPT is the share leader in TCB equipment.
• Q1 26 bookings +71.6% YoY = 4-year high, signaling next 12–18 months of revenue acceleration.
• Less crowded than ASML / Lam / AMAT on the western side; same theme exposure.
• HK listing creates discount to global semi equipment peers.
• HBM market tripling in 2026 (per IDC) directly drives ASMPT capacity demand.

What the market is missing: Bookings are a leading indicator for revenue 9–18 months out; the +71.6% print has not yet been fully reflected in consensus estimates.

Key Risks:

• TCB-to-hybrid bonding transition (Applied Materials competition for next-gen)
• Memory cycle reversal
• China demand softness / export controls on China customers
• Order timing — bookings can lump

Next steps: Bookings-to-revenue conversion model; track HBM4 capacity announcements from Samsung/Hynix/Micron.

---

Idea 9 — Palantir (PLTR US) — LONG — Premium-priced but agentic AI revenue inflection most credible in software

Thesis:

• Palantir is the clearest revenue-inflection story in AI software — accelerating from 21% (Q1 24) → 85% (Q1 26).
• AIP (Artificial Intelligence Platform) is genuinely differentiated: agent-orchestration on top of customer ontology; enterprises pay for the data integration moat.
• 206 deals >$1M closed in Q1 alone; US commercial RPO compounding aggressively.
• Profitability is real (GAAP NI margin 53%, op margin 46%); not a "growth at any cost" name.
• Yes, multiple is rich (~100x+ NTM earnings) — but the Rule of 40 = 145% justifies premium.

What the market is missing: Most "AI software" stories haven't translated to topline (see C3.ai -46%); Palantir is the rare case where AI thesis = AI revenue. Comp set is too narrow because no peer is delivering 85% growth at 46% GAAP margins.

Key Risks:

• Multiple compression if growth decelerates below 50%
• Government segment volatility (DOGE-related cuts, contract recompetes)
• Founder / management turnover
• Macro risk-off compresses high-multiple names disproportionately

Next steps: Cohort analysis of US commercial customer scaling; track AIP token/deal metrics quarterly.

---

Idea 10 — Marvell (MRVL US) — LONG — XPU + optics ramp, NVIDIA partnership underappreciated

Thesis:

• Marvell sits at three intersecting curves: (1) custom XPU/ASIC for hyperscaler inference, (2) optical DSP for AI networking (data center interconnect), (3) AI RAN.
• New NVIDIA partnership on optics, NVLink Fusion, and AI RAN — under-modeled in consensus.
• Custom XPU revenue growing >2x in FY28; DCI module business on path to $1B annualized in FY28.
• Higher growth than Broadcom but smaller scale → more multiple expansion potential if execution sustains.

What the market is missing: Marvell has been treated as the "discount Broadcom," but the NVIDIA optics partnership transforms it into a critical NVIDIA ecosystem partner — ironic given its ASIC competition with NVIDIA in inference.

Key Risks:

• ASIC customer concentration
• Optics commoditization risk
• Ramp execution on FY28 XPU 2x guide
• Cyclical tail in non-DC segments (carrier infrastructure)

Next steps: Per-customer XPU ramp model; track 1.6T optical adoption pace.

---

Prioritized Research Order

1. Eaton (ETN) — start here. Cleanest backlog story, simplest thesis, easiest to model.
2. Baidu (9888 HK) — highest-conviction value/growth-mismatch in HK; SOTP work is high-ROI.
3. Broadcom (AVGO) — biggest dollar-impact name; deserves deep AI revenue waterfall.
4. Tencent (700 HK) — quality compounder; quick-win if Hunyuan ramp inflects.
5. Micron (MU) — cyclical-vs-secular debate; framework-determining.
6. Lenovo (992 HK) — under-covered globally; channel-check intensive.
7. Zscaler (ZS) — sub-$50B caps offer clearer alpha if AI security thesis sticks.
8. ASMPT (522 HK) — booking-to-revenue conversion is the entire trade.
9. Palantir (PLTR) — momentum + valuation work; size carefully given multiple risk.
10. Marvell (MRVL) — finest of the names but most expensive; pair with AVGO.

---

Portfolio Construction Notes

• Geographic balance: 6 US / 4 HK suggested for a global mandate. HK names provide structural value diversification against US AI multiple compression.
• Sub-theme balance:
• Power/infra (Eaton) — defensive, multi-year
• Custom silicon (Broadcom, Marvell) — secular growth
• Memory (Micron) — cyclical-secular hybrid
• Cyber (Zscaler) — AI-adjacent
• Mega-cap AI (Tencent, Baidu) — value with optionality
• Hardware (Lenovo, ASMPT) — picks-and-shovels
• Pure software (Palantir) — high-conviction growth
• Pair trades: Long AVGO / Short NVDA on inference shift; Long Baidu (9888 HK) / Short Alphabet on relative AI revenue disclosure.
• Sizing: Cap any single name at 5–7% of book given AI sector correlation in drawdowns; cap aggregate AI exposure at 25–30%.

---

What We Avoided and Why

---

Confidence & Caveats

• High confidence on theses: Eaton, Broadcom, Micron, Baidu, Tencent — fundamentals + numbers cross-verified
• Medium confidence on multiples: Many NTM multiples are estimates (no live terminal); cross-reference before sizing
• Highest model risk: Palantir (multiple expansion required), MiniMax-class momentum names (excluded for this reason)
• Macro overlay: All ideas face shared AI-cycle risk; if hyperscaler 2027 capex guidance disappoints in Q3 26 prints, the entire complex re-rates lower

---

Catalysts Timeline (next 90 days)

---

Sources

Pulled from:

1. AI Infrastructure Sector Overview — US & Hong Kong (May 29, 2026 companion document)
2. Company IR press releases (Q1 / FY26 results)
3. Sell-side aggregations (Goldman, Morgan Stanley, JPMorgan, BofA, MUFG, CreditSights)
4. Industry data (Mordor, IDC, JLL, ARK, Gartner, Deloitte, IEA)
5. Live web search aggregations as of late May 2026

Methodology: Quantitative screens (growth, quality, GARP) intersected with thematic AI infrastructure thesis. No screen output is a recommendation — every idea requires fundamental model build before sizing.

---

Prepared May 29, 2026. This is a research starting point, not investment advice. Verify all multiples and prices on a live terminal before any action.