Investment theses, due diligence findings, and evolution logs for each position in the Tiger Portfolio.
Astera Labs is building the connectivity platform layer for AI data centers. Their product portfolio spans PCIe retimers (Aries), CXL memory controllers (Leo), smart fabric switches (Scorpio), and optical engines — covering the full spectrum of data movement bottlenecks inside GPU clusters. This multi-product breadth is rare for a company this young and gives them multiple shots on goal as AI infrastructure scales.
The Amazon relationship is the centerpiece: a $6.5B warrant deal (3.26M shares at $142.82, performance-vested, expires 2033) signals deep strategic commitment, not a transactional supplier arrangement. Microsoft is confirmed for CXL memory deployment. In 2026, Scorpio smart fabric switches are expected to expand from 1 to 3+ hyperscaler customers — a critical inflection for product diversification.
The stock has declined 49% from its ATH of $262.90, driven by valuation compression rather than fundamental deterioration. Revenue growth remains strong at +92% YoY with 75.7% gross margins. The risk is customer concentration (top 3 = ~80% of revenue) and a rich valuation even post-selloff, but the depth of hyperscaler commitment and multi-product expansion provide a credible path to sustained growth.
Astera's moat lies in the breadth of their connectivity platform — PCIe, CXL, fabric switches, and optical engines — designed to work together inside AI clusters. CXL is a nascent standard where Astera has first-mover advantage with Leo controllers already deployed at Microsoft. The Amazon warrant structure creates deep financial alignment that competitors can't easily replicate. Design wins at hyperscale are sticky: once embedded in a rack architecture, switching costs are high due to qualification cycles and system interdependencies.
Credo Technology essentially created the Active Electrical Cable (AEC) category and dominates it with an estimated 88% market share. AECs are the copper-based alternative to optical cables for short-reach interconnects inside AI clusters — cheaper, lower power, and increasingly preferred by hyperscalers for rack-level connectivity. As AI clusters scale, the density of short-reach connections grows exponentially, and Credo sits at the center of this trend.
The growth story is extraordinary: +272% YoY revenue growth driven by hyperscaler adoption. Amazon is the #1 customer (not Microsoft, as commonly assumed), with concentration improving from 86% to 42% as Microsoft, xAI, and Meta have ramped. A 5th hyperscaler is now ramping. This rapid customer diversification significantly de-risks the revenue base.
The company is founder-influenced: co-founders Lawrence Cheng (CTO) and Job Lam (COO) remain active while professional CEO Bill Brennan manages operations. Product expansion into line card retimers, optical DSPs, and PCIe retimers provides growth beyond the core AEC franchise. The risk is that 93% of revenue still comes from just 4 hyperscaler customers, making Credo highly dependent on AI capex cycles.
Credo created the AEC market and commands 88% share (650 Group estimate). Their moat is built on proprietary mixed-signal SerDes IP that enables high-speed, low-power copper connectivity — extremely difficult to replicate. Once a hyperscaler qualifies and deploys AECs in their rack architecture, switching to a competitor requires re-qualification of the entire cable plant. The technical complexity of high-speed SerDes design (112G PAM4 and beyond) creates a natural barrier to entry, with only Broadcom as a credible potential competitor.
Fabrinet is the invisible manufacturer behind the AI optical revolution. They don't design transceivers — they build them, with extreme precision, at low cost, in Thailand. Their defining relationship is with NVIDIA: Fabrinet manufactures NVIDIA's self-designed optical transceivers, capturing ~35% of revenue from this single customer. As NVIDIA pushes to self-produce 50%+ of its optical transceiver needs (up from 15-20%), Fabrinet is the primary beneficiary.
The business model is uniquely capital-efficient: $0 R&D spend (customers fund it), customers provide equipment, capex runs 2-3% of revenue (vs. 8-12% for Coherent), zero debt, and $969M in cash. Gross margins of 12.4% look low but are structural for contract manufacturing — they're actually 2-3x peers like Jabil. This is a toll-booth business: Fabrinet doesn't take technology risk, they take execution risk.
The "100% share" narrative needs clarification: it refers to NVIDIA's self-designed 1.6T transceivers ONLY, not the total transceiver market. Coherent designs its own transceivers; Fabrinet builds NVIDIA's designs. They operate in different parts of the value chain. The key risk is NVIDIA concentration: if NVIDIA decides to in-source manufacturing or switch to another contract manufacturer, Fabrinet loses its primary growth engine.
Fabrinet's moat is operational excellence in precision optical manufacturing at scale, combined with a Thailand-based cost advantage that's difficult to replicate. The zero-R&D, customer-funded model means Fabrinet never takes technology risk — they take execution risk, and they execute consistently. The NVIDIA relationship has deepened over multiple product generations, creating qualification-based switching costs. Competitors would need to build equivalent cleanroom capacity and demonstrate yield performance over years to become credible alternatives.
nVent Electric is the cleanest risk profile in the Tiger Portfolio. As NVIDIA's liquid cooling partner with a $2.3B backlog (3x YoY growth), nVent is positioned at the intersection of AI infrastructure and physical cooling — a problem that only gets harder as GPU power density increases. Unlike connectivity plays, cooling is non-negotiable: every AI data center needs it, and the thermal envelope keeps expanding.
What makes nVent unique in this portfolio is diversification. No single customer exceeds 12% of revenue — a stark contrast to the 35-80% concentration seen in ALAB, CRDO, and FN. This reduces binary risk and provides more stable revenue visibility. The company spun off from Pentair in 2018, and CEO Beth Wozniak has been leading since inception.
The trade-off is growth: at +35% YoY, nVent isn't a hypergrowth name, and the 28x forward P/E reflects expectations for steady rather than explosive growth. The stock is testing $115-121 resistance. This is a portfolio anchor — lower upside but significantly lower downside than the other names.
nVent's moat is built on deep thermal engineering expertise combined with the NVIDIA partnership for liquid cooling solutions. Data center cooling is becoming more complex as GPU power density increases — this isn't a commodity business but requires specialized engineering for each deployment. The $2.3B backlog provides revenue visibility, while customer diversification (no customer >12%) reduces concentration risk. As data centers shift from air to liquid cooling, nVent's installed base and qualification relationships create switching costs.
Tower Semiconductor is the leading specialty foundry for silicon photonics (SiPho), the critical optical interconnect technology enabling 800G and 1.6T data transceivers in AI data centers. With a confirmed NVIDIA collaboration, $920M committed to 5x capacity expansion by Dec 2026, and 70%+ of future SiPho capacity already reserved through 2028, Tower is a picks-and-shovels AI infrastructure play with a defensible analog/specialty niche that TSMC and Samsung don't prioritize.
Revenue is inflecting — SiPho grew 115% YoY in 2025 to $228M — and the 2028 target model implies a path to $2.84B revenue at 40% gross margins, nearly doubling from here. The margin story is the most compelling part: net margin went from 11% to 18% through 2025 as SiPho mix increased, with incremental margins near 49%.
CEO Russell Ellwanger has led Tower for 21 years, transforming it from a near-bankrupt single-fab Israeli chip maker into a $15B+ global specialty foundry. The SiPho/NVIDIA play is his best strategic move yet. Key risks include the $920M CapEx bet, Intel Fab 11X collapse, and elevated valuation at ~70x TTM P/E — though on the 2028 target model, forward P/E compresses to ~20x.
Tower's moat is built on deep silicon photonics process IP developed over 10+ years, combined with capacity lock-up (70%+ of SiPho capacity reserved through 2028 with customer prepayments). The NVIDIA collaboration on 1.6T optical modules validates Tower's platform at the highest level. Semiconductor process qualification takes 12-18 months per customer, creating high switching costs. With $920M committed to 5x capacity expansion, Tower is building a supply-side moat that competitors would need 2-3 years and hundreds of millions to match. TSMC, Tower, and AIM Photonics control >60% of photonics wafer contracts, but Tower is the de facto leading commercial SiPho foundry.
Arista Networks is the dominant Ethernet networking platform for hyperscale data centers and the primary beneficiary of the AI infrastructure buildout. Ethernet has captured 65%+ of new AI back-end deployments by early 2026, displacing InfiniBand, and Arista leads in 400G/800G branded DC switching. With a $9B revenue base growing 29% YoY, 64% gross margins, and $10.74B in cash with zero debt, this is software-like economics built on hardware.
The AI networking supercycle is the primary growth engine: $1.63B in AI revenue in 2025, doubling to $3.25B guided for 2026. Meanwhile, campus networking ($815M→$1.25B) opens a new $20B+ TAM where Arista has barely scratched the surface. CEO Jayshree Ullal has a flawless track record — originally guiding $10B revenue by 2028, she'll hit it 2 years early.
Key risk is customer concentration: Microsoft (26%) + Meta (16%) = 42% of revenue. Meta already proved this risk in 2020-21 when it skipped the 200G generation. Valuation at 52x trailing P/E is demanding and any growth deceleration gets punished. But with founder ownership at 22%+, $6.8B backlog, and GPU-agnostic positioning, Arista is a rare compounder with both secular tailwinds and a widening moat via EOS.
Arista's moat is EOS (Extensible Operating System) — a single binary image across all platforms that dramatically simplifies operations vs. Cisco's fragmented NX-OS/IOS stack. Once deployed at scale (10,000+ switches), rip-and-replace costs are prohibitive. The merchant silicon strategy (Broadcom ASICs) delivers faster time-to-market and lower cost structure, enabling 64% gross margins while aggressively pricing. Being the incumbent at Microsoft, Meta, and other hyperscalers creates a reference architecture flywheel — new cloud builders default to Arista because that's what the cloud titans run. NOS agnosticism (letting hyperscalers run their own OS) paradoxically strengthens the relationship.
Coherent Corp is the vertically integrated optical transceiver player — they design AND manufacture their own transceivers, including InP (Indium Phosphide) vertical integration. This is a fundamentally different model from Fabrinet (which only manufactures for others) and gives Coherent more control over their technology roadmap and margins. The 4x book-to-bill ratio signals strong demand visibility.
However, Coherent is currently classified as Watch/Weakened due to technical concerns. The chart has thin structure above $121, meaning there's limited price support if the stock breaks out and then reverses. Revenue growth of +17-22% YoY is respectable but modest compared to ALAB (+92%), CRDO (+272%), or even FN (+36%). Gross margins of 36.9-39% are decent but not exceptional.
The thesis here is wait-and-watch: Coherent has real technology advantages, but the risk/reward isn't as compelling as the core holdings. The weakened technical setup means entries need to be more patient and selective.
Coherent's moat is vertical integration: they control the entire stack from InP wafer fabrication through transceiver design and assembly. This gives them technology differentiation and (eventually) margin advantages as they optimize across the stack. However, vertical integration also means higher capex (8-12% of revenue vs. Fabrinet's 2-3%) and more execution complexity. The moat is real but comes with higher capital requirements.
Quanta Services is the largest electrical contractor in the United States with 68,000 workers and a monopoly-like scale in electrical infrastructure. Every AI data center needs to be connected to the power grid, and Quanta is the company that does the physical electrical work. The $39.2B backlog is massive and provides multi-year revenue visibility.
However, Quanta is classified as Watch/Weakened because the stock is near all-time highs and trading at 76x P/E — an extraordinary valuation for what is fundamentally an electrical contracting services business with ~15% gross margins. The services model inherently limits margin expansion, and the near-ATH price means the risk/reward is unfavorable for new entries.
This is a "right company, wrong price" situation. The thesis on Quanta's business quality and market position is strong, but the valuation requires patience. A meaningful pullback (20-30%) would make this a much more interesting entry point.
Quanta's moat is pure scale: 68,000 workers, decades of utility relationships, and the largest fleet of specialized electrical equipment in the country. Building a competitor would require years and billions in investment — it's practically impossible to replicate this workforce and relationship base. Every major data center build, grid upgrade, and renewable energy project needs electrical contractors, and Quanta is the dominant player. Regulatory relationships and safety track records add further barriers.
Upstream EML (Electro-absorption Modulated Laser) manufacturer — the optical source inside every high-speed transceiver and co-packaged optic (CPO). 86% cloud revenue. Two emerging revenue categories: (1) Optical Circuit Switch (OCS) — intelligent optical switching for AI scale-out networks, $400M+ backlog converting H2 2026; (2) CPO laser chips — LITE supplies the EML source to COHR, FN, and others building next-gen modules. Baker-perfect alignment: optical interconnects are one of his two clearest AI infrastructure calls. Q3 FY2026 guidance: $805M (~85% YoY). Promoted to watchlist Feb 19, 2026 — technically extended (RSI 74, +188% above 200d), watching for consolidation to SMA20 ($467) before entry.
MEDIUM. Three primary OCS customers represent revenue concentration. However, EML laser manufacturing has 18–24 month qualification cycles — switching costs are structural. Vertically integrated competitors (e.g., TSMC SiPho route) pose a 3–5 year risk, not near-term.
Hyperscaler OCS rollout delay beyond 2026; competing EML process certified at TSMC before LITE fab expansion completes; CPO deployment pushed past 2027 by hyperscalers.
171-year glass science moat plays AI infrastructure buildout through passive optical fiber and cable — the physical conduit for every AI cluster's networking traffic. Meta $6B multiyear deal anchors US capacity expansion with revenue assurance structures. Optical segment growing 24% YoY. Springboard plan targets $11B incremental revenue by 2028 ($27.4B total). Differentiated from COHR/FN/LITE: GLW makes the pipe, not the signal — passive infrastructure that benefits from any optical architecture (GPU or ASIC, pluggable or CPO). Poorna conviction pick. Promoted to watchlist Feb 19, 2026 — technically extended, watching for pullback before entry.
LOW–MEDIUM. Meta $6B deal has revenue assurance structures (customer prepayments). Glass manufacturing IP is 171 years deep — no credible vertical integration threat. Main risk is GLW's diversification: Display (TV panels) and Auto segments are cyclical and can offset optical gains. AI thesis is real but diluted across total company.
Meta deal delay or cancellation; Display segment deterioration exceeds optical gains; on-chip photonics eliminates external fiber runs (decade-long risk); enterprise wireless substitution for short-reach connectivity.
Pure-play data center power and cooling infrastructure — 100% AI/DC revenue vs NVT's ~30% DC exposure. Vertiv makes UPS systems, power distribution units (PDUs), thermal management, and liquid cooling purpose-built for hyperscale AI data centers. $15B backlog doubled YoY; 252% Q4 organic order growth. Direct Baker alignment: liquid cooling is mandatory infrastructure for Blackwell (30kW → 130kW per rack) and VRT's full product suite is the answer. Superior to NVT on AI-pure-play: VRT's entire business benefits from AI capex surge, not just one segment. Promoted to watchlist Feb 19, 2026 — technically extended, watching for pullback before entry.
NVT: diversified industrial (~30% DC, ~70% industrial/enclosures). VRT: 100% DC-focused. The two serve the same physical layer but VRT is higher-beta AI and NVT provides portfolio stability. Both can coexist — VRT for the pure-play AI cooling thesis, NVT as the more conservative infrastructure anchor. Entry at different times depending on which technical setup arrives first.
MEDIUM. Hyperscaler customer concentration (large orders from MSFT, Google, Amazon, Meta). Baker's "Blackwell ROI air gap" — 3 quarters of training capex without inference revenue — could cause temporary order pause. Schneider Electric and Eaton compete in power distribution. Liquid cooling competition from CoolIT, Vertiv has first-mover advantage but market will broaden.
Hyperscaler capex pause affecting liquid cooling orders; supply chain execution failures on $15B backlog conversion; hyperscaler vertical integration of power/cooling (currently unlikely — too capital-intensive and outside core competency).