The five layers of AI compute the market still mispriced — and the names we hold.

The cohort's cleanest single Asian-listed under-priced position — though after a +900% 1Y, the 'under-priced' descriptor is increasingly retrospective. HBM revenue contribution has crossed 40% of the DRAM segment with HBM gross margin materially above the blended DRAM cycle. The consolidated multiple still reads as if the firm were a conventional cyclical memory producer; the HBM share is structurally a different business with three-firm supply and named-customer concentration. Forward question: is this a hold/add or a take-profit?

Functionally a single-product monopoly that scales with EUV layer count. Every EUV mask used by TSMC, Samsung, and Intel passes through a Lasertec tool. The customer set is functionally three customers; the share count is small (~90M); the float is limited. Forward multiple sits well above the WFE peer set but the position is structurally a one-supplier monopoly with no qualified second source at production scale. Re-rates discontinuously on each EUV node-extension or layer-count disclosure.

HBM stacking and CoWoS-class packaging both require thinned wafers (sub-50µm) that only Disco's high-end grinders produce at yield. Functionally a single product line scaled to the AI-cycle backlog. The cohort's clearest example of a tool firm whose order book is a direct read on hyperscaler HBM and packaging demand — not on a generalized semiconductor capex cycle.

Dominant supplier of memory and SoC test handlers, with a structural share lead in HBM test (>90%) that the program treats as the cohort's single highest supplier-concentration metric. As HBM stack heights grow (8-Hi → 12-Hi → 16-Hi), test time per package scales roughly linearly — a multi-year tailwind that consensus revenue projections under-weight on the back-half of the HBM4 cycle.

Structurally under-priced relative to US Big Four WFE peers (AMAT, LRCX, KLAC). The discount composed of (a) Japan-listing illiquidity premium for US investors, (b) larger memory exposure depressing the trailing print through 2023-2024, (c) absence of US-investor narrative on the EUV coater/developer monopoly. Forward multiple is the cohort's clearest US-vs-Japan peer-relative arb on equal-quality WFE businesses.

ASMPT's TCB order book is the cleanest single publicly-disclosable read on TSMC CoWoS capacity additions. The HKEX listing produces a structural liquidity discount that depresses the multiple relative to where a sole-supplier monopoly to TSMC's packaging line would price on NASDAQ. Hybrid-bonder pipeline (for true 3D logic-on-logic stacking) is the firm's second-stage catalyst, currently under-recognized in the consensus forecast.

Revenue grew materially in 2024-2025 on the back of CoWoS-related orders. Consensus forward multiple reflects the historical molding-tool cycle rather than the AI-cycle capex tied to TSMC packaging. The firm is the cohort's clearest example of a small-cap Japanese tool firm whose product mix has fundamentally repivoted toward AI packaging but whose investor base has not yet repriced the business.

Datacom segment revenue mix has crossed 50% of consolidated revenue, with the EML laser business carrying a structural cost-advantage on the 1.6T generation transition. The consolidated multiple still embeds the legacy telecom-cycle mix and the materials-segment cyclicality. As cluster sizes scale from 10k-GPU to 100k-GPU configurations, per-GPU optical spend has roughly tripled; the firm's revenue trajectory has begun to reflect this but the forward multiple is anchored to the pre-AI-cycle business mix.

The cohort's clearest example of a foreign-listed firm with disclosed US-hyperscaler customer concentration that the US consensus does not read. Innolight holds large global share in 400G and 800G transceivers; the customer mix is roughly parallel between US hyperscalers and Chinese AI campus operators. The A-share listing produces a structural liquidity discount for US investors; the forward multiple sits well below where the equivalent US-listed firm would price on the same customer concentration. Geopolitical risk is real but is the disclosed reason for the discount.

The second of the Western optical-interconnect pair (with Coherent). Smaller, more transceiver-pure, and with a faster mix shift toward AI customers. The Cloud Light acquisition lifted the firm's datacom share materially; the consolidated multiple has not fully re-rated to reflect the post-acquisition mix. The position is somewhat more cyclical than Coherent on the back of less laser-chip vertical integration; the offset is larger AI-customer concentration and faster mix transition.

Electrical-Americas segment carries the cohort's most-disclosed direct-to-data-center exposure. Switchgear lead times are out 24-36 months; the firm books current-year orders against revenue that lands two-to-three fiscal years out. The forward multiple has rerated but anchors on FY1 consensus that has not yet absorbed the back-half of the 2026 hyperscaler-PPA cycle. FY2 and FY3 consensus carries materially more upside than the trailing print implies.

The cohort's longest-lead-time, highest-revenue-lag position. Industrial gas turbines for behind-the-meter generation at hyperscaler campuses carry three-year order backlogs; grid equipment for AI campus interconnection sits on similar lead times. The post-spin trading history is short and the consensus FY3 has compressed since the rerate — but the disclosed backlog and customer concentration sit at the cohort's upper bound for the power layer. The combination of long lead time + long contract duration + AI-cycle disclosed exposure is structurally rare.