Second Echelon Research

The market discovers first-echelon leaders instantly. NVIDIA went from $15 to $200+ as AI became obvious. By the time you see it on CNBC, the alpha is gone. The real opportunity is in the second echelon — companies that feed the same mega-trend indirectly, but haven't been fully priced in yet.

Step 1: Identify the mega-trend and its first-echelon leader (e.g., AI compute = NVIDIA)
Step 2: Map the supply chain — who supplies, enables, or benefits indirectly?
Step 3: Find companies priced for their legacy business while the AI/growth angle hasn't shown up in consensus estimates
Step 4: Verify quality using the Gross Margin 3-Question Test (see below)
Step 5: Buy before the market connects the dots

AI = GPUs (obvious) → but also memory (MU: +85% TTM), interconnects (CRDO: +226%), cooling (VRT: +29%), optical networking (CIEN: +26%), power generation (GEV), nuclear fuel (CCJ)

SpaceX (private, can't buy) → Rocket Lab (RKLB: only pure-play public launch company)

Photonics winners priced in (LITE +1,300%, COHR +298%) → substrate monopolists (SOI: 80% share), equipment monopolists (AIXA: 70-90% MOCVD share), CPO component makers (HIMX: Citrini's #1 pick)

AI agents creating 25x traffic → DNS monopoly (VRSN: 88% GM), CDN-to-compute (FSLY, NET), agent telephony (BAND: $1B market cap), stablecoin payments (CRCL)

Not all companies benefiting from a trend deserve investment. We apply a strict quality filter based on trailing gross margin trends. This framework, inspired by Terry Smith's quality investing approach, distinguishes structural problems from temporary dips.

Q1: Is gross margin declining?
No (stable or expanding) → PASS
Yes → Go to Q2

Q2: Is the decline structural or cyclical?
Structural (competition eroding pricing power, commoditization, no recovery timeline) → FAIL
Cyclical (product ramp, identifiable cause, management guides recovery) → Go to Q3
Mix shift (higher-growth but lower-margin segment growing faster) → Go to Q3

Q3: Is revenue growth accelerating despite the GM dip?
Yes + clear recovery catalyst → PASS (monitor quarterly)
No → FAIL

Additional filters:
- TTM (trailing 12 months) revenue growth ≥ 10% — NOT single-quarter YoY
- GM > 50% for tech/software, > 30% for other sectors
- Net income quality check — real operating income, not paper gains
- AI must be a revenue driver, not a threat to the business model

The first layer of second echelon identifies companies that directly supply the AI buildout but aren't the headline names. When NVIDIA sells a GPU, someone had to make the memory (MU), the interconnects (CRDO), the cooling system (VRT), the optical networking (CIEN), and generate the electricity (GEV, CCJ). These companies have real revenue acceleration and expanding margins, but trade at fractions of NVIDIA's multiple because the market hasn't connected them to the same trend.

The deepest supply chain layer. The optics winners (LITE +1,300%, COHR +298%) are priced for perfection. Citrini identified the next alpha one layer deeper: the substrate monopolists, equipment makers, and component suppliers that ALL optics companies depend on — regardless of which architecture wins.

Key insight: Whether pluggable transceivers or co-packaged optics (CPO) wins, the substrate (Soitec SOI wafers), the MOCVD tools (AIXTRON), the wafer bonding equipment (SUSS MicroTec), and the lens arrays (Himax) are ALL required. Architecture-agnostic = risk-free picks-and-shovels.

The photonics supply chain:
TSMC COUPE platform needs → SOI wafers (Soitec, 80% monopoly) → SiPh PIC fabrication (STM PIC100 with AWS, Nokia InP fab with NVIDIA $1B) → MOCVD tools to grow lasers (AIXTRON, 70-90% monopoly) → Micro lens arrays for CPO (Himax, patent-linked to TSMC) → FAU assembly (FOCI)

Why now: NVIDIA shipping first commercial CPO in H1 2026. Optical transceiver market exceeding $10B in 2026 (2x from 2024). Laser shortage with 1-year lead times. Every fab expansion = orders for the same 2-3 equipment/substrate suppliers.

AI agents generate 25x more network traffic than chatbots (Cisco estimate). The entire internet infrastructure must adapt. Citrini classifies the winners into three categories:

Infrastructure: The "east-west" traffic explosion. Traditional internet was north-south (user→server→response). Agentic traffic moves laterally between servers, APIs, and data centers. This reprices CDNs as "Compute Delivery Networks" (FSLY, NET, AKAM), optical backbone (CIEN), DNS resolution (VRSN — every API call starts with DNS), and API gateways (FFIV — owns NGINX).

Ecosystem: Companies offering services TO agents (B2A — business-to-agent). Agent telephony (BAND — PSTN infrastructure, carrier licenses), agent-to-human communication (TWLO — 15 years of customer data via Segment), digital agreements (DOCU — FedRAMP, ESIGN Act compliance), stablecoin payment rails (CRCL — $9T processed, Circle Nanopayments for $0.000001 transfers).

Governance: Observability (what is the agent doing?), security (is it allowed to do that?), and identity (who is it?). These converge into "Agentic Governance." Key players: PANW (acquired Chronosphere + CyberArk), DT/DDOG (observability), ZS (zero-trust inline security for every agent API call), SAIL (identity lifecycle for agent "employees").

The deepest layer of the entire framework. Software companies face existential AI disruption risk. Physical materials companies don't — you can't prompt your way out of a titanium shortage.

The thesis: Capital is rotating from bits (software, 30x revenue multiples) into atoms (physical materials, 5-8x EBITDA). The market is assigning an "AI Disruption Discount" to software while undervaluing the physical world that AI depends on. Every bottleneck in the AI buildout has been in the real world: power, cooling, materials, chemicals.

What makes an "atom" investable:
- Concentrated supply — monopoly, duopoly, or oligopoly structures
- Multi-year qualification cycles (2-5 years) creating deep switching costs
- Zero substitution paths — physics prevents alternatives
- Demand pulled forward simultaneously by defense budgets + electrification + aerospace + AI buildout
- Priced at commodity multiples (5-8x EBITDA) despite strategic value

The sectors:
Beryllium (MTRN — world monopoly), Titanium sponge (Japanese duopoly — 80% Western share), Tungsten (KMT — only Western vertically integrated manufacturer, prices tripled), Uranium conversion (SOLS — only US UF6 facility, contracts repricing 3x), Superalloys (ATI, HWM — jet engines + AI gas turbines), Electrical steel (JFE — monopoly on 6.5% silicon steel, transformer lead times 128 weeks), Carbon fiber (HXL — 93% aero/defense, SpaceX kicker), Mining equipment (Komatsu — same business as CAT at half the multiple), AI materials (ENTG — electronic chemicals, HOYA — photomask blanks, Asahi Kasei — HBM packaging chokepoints)