Bottom-up models, technology primers, and investment memos covering the buildout of compute, networking, power, and the application layer that consumes them.
Working index — small audience, unlisted.
Working index of research outputs. Each card links to a self-contained piece. Models open as live-data summary pages; primers and memos open as the full reading text with inline citations.
Numbers throughout chain back to the source model (the bottom-up token model, which now incorporates the DC-capex archetype build). Citations chain back to verbatim primary sources, verified line-by-line.
What a token actually is, how it's priced, and where the margin lives. The unit-level mechanics of inference economics — input vs. output, the deflation curve, and the agentic multiplier.
Read primerOptical and electrical interconnect economics through 2030. Goldman's $154B optical TAM unpacked, with services unbundled — network test, hyperscale fit-out, fiber MSPs all surface as exposed names.
Read primerIs multi-DC AI interconnect a distinct, investable equipment category? A $10.4B (2026) market growing to a $17.5B 2029 base inside a $13.9B–$23.4B bear-to-bull bracket — sized layer by layer, with conviction anchored on coherent pluggables and DWDM line systems.
Read primerAs AI inference pushes outward, does it drive a secondary optical wave at edge DCs, CDNs, regional colos and tier-2/3 carriers — and who captures it? A descriptive map of what's real today: the measured money is hyperscaler scale-across (captured by Ciena/Nokia/Cisco + Marvell/Coherent/Fabrinet); non-scale vendors win only the fragmented regional/metro tail, with zero wins at the hyperscaler tier. Ribbon's "67% rural DCI" claim, stress-tested.
Read primerHow the AI datacenter buildout is being financed. Neocloud debt, hyperscaler bond issuance, ABS structures, Stargate financing — and where the credit cycle gets uncomfortable.
Read primerWhat "agentic AI" actually means in practice, and the workforce-replacement framing that underpins the token-demand model. Foundational primer for the bottom-up token model.
Read primerThe agent-to-agent communication layer — MCP, A2A protocols, orchestration patterns, and what changes when agents talk to other agents rather than to humans.
Read primerA market deep-dive that maps where the optical dollar actually lands across the six-layer chain and reads the expectations already baked into prices. The spine is "two ends and a pool": value accrues at the two ends — the scale system (Ciena, capturing the regional/metro socket the sub-scale vendors were meant to own) and the merchant components — and in the durable optical line-system pool, while IPoDWDM hollows out the standalone-transponder middle. The optical transport market hit $16B in 2025 (+10% YoY) on DCI demand (+40%). Then a crowded-versus-overlooked investible-pockets screen across the public and private names — a value-accrual map plus a valuation read, no forward model and no per-name buy/sell call. With model-derived charts.
Read reportWhat's actually inside each datacenter archetype, line by line. Cumulative spend by category, 2026–2030. The investable-theme view sits on top of this: networking surfaces as the asymmetric exposure.
Read memoAn answer-shape brief on one question: is NVDA actually cheap right now? Yes on the evidence — ≈31x trailing earnings is about half its five-year average (~63x), the next-twelve-month multiple is the lowest since June 2019, and NVDA is the cheapest of the compared AI-semi set (AMD, Broadcom, TSMC) on trailing earnings outright and growth-adjusted (forward PEG ≈0.76 vs 0.92–2.15; BofA frames 0.28 on 2027 estimates). Two honest caveats: the discount fades over a 20-year window, and the low multiple prices consensus estimates that assume revenue nearly doubles again — the market charging for estimate risk, not a free discount. 57 citations, 42 primary snapshots, dual-family audited.
Read briefThe HBM leader, priced for permanent peak-cycle margins. On forward earnings it screens among the cheapest in memory (6.6x) while price-to-book sits near a five-year high — the classic cyclical-at-peak tension the report sets out to reconcile. Every durable-earnings method — a light DCF across bear/base/bull and a normalized price-to-book — clusters below today's market price; even Morgan Stanley's Overweight target has been run past by the stock. A valuation-risk and asymmetry call, not a short: the near term is hugely cash-generative and largely locked (HBM sold out ~3 years, Q2 guided up, ~35T won net cash). With model-derived charts.
Read reportComparative deep-dive on Japan's three "electric-wire" majors — Furukawa (5801.T), Sumitomo Electric (5802.T), Fujikura (5803.T) — as one AI-datacenter-optical theme expressed through three very different vehicles. The purity spectrum (81% / 42% / 18% of group operating profit), the valuation trap (Furukawa is nearly as expensive as Fujikura on half the AI purity), and why an equal-weight basket is the disciplined way to own the theme — full theme beta, diversified single-name risk. With model-derived charts.
Read reportDoes a test-and-measurement business capture the AI-networking buildout? Keysight is a toll on the complexity of interconnect — test demand leveraged to standards churn (800G→1.6T→3.2T, UEC / UALink), not capex dollars. The Spirent twist (the data-center-Ethernet test line went to Viavi for $399M), the agentic-traffic runway, and why at ~$350 the re-rating is already paid for. High-quality hold / buy on weakness, with model-derived charts.
Read reportA two-name comparative deeper report. Both are sub-scale tier-2 optical/access vendors with real-but-minority AI-DC interconnect exposure (RBBN ~3% of revenue, ADTN ~11%); the AI re-rating accrued to scale leader Ciena (~11.6× EV/Sales) while RBBN (0.97×) and ADTN (1.25×) got none. ADTN has the stronger claim but carries the ~$351.7M DPLTA squeeze-out overhang.
Read reportSingle-name research on Corning, the 175-year-old materials company now a leading passive-optical supplier to the AI data-center buildout: the Enterprise-optical ramp (+61%), the Meta + NVIDIA agreements, and why ~$183 prices in the Springboard plan.
Read briefA single-name deeper report on Circle, the USDC stablecoin issuer: the reserve-income float model, the Coinbase distribution tax that absorbs ~half the rate sensitivity, and why ~8.6× EV/Sales still isn't cheap versus Coinbase after a 66% drawdown. Cautious / below-consensus, with model-derived charts.
Read report