Coverage Philosophy
The library is organised into eight categories that intentionally mirror the most contested areas of the technology sector right now: generative AI infrastructure (compute silicon, memory & storage, infrastructure software), the hyperscalers funding it, early stage quantum, the security stack absorbing AI driven attack surfaces, the physical AI / robotics layer, and the underlying networking & server hardware that runs everything.
| Category | Why it's in scope |
|---|---|
| Quantum Computing | Long duration, pre revenue. Valuation needs a fundamentally different lens (TAM × probability × time, not DCF). High dispersion of outcomes. |
| Compute Silicon & Foundries | The full chip stack: merchant GPUs, custom XPUs, AI networking, foundries, and semiconductor capital equipment. AI workloads are the marginal driver but the cycle covers all of compute. |
| Memory & Storage Hardware | HBM is the supply constrained bottleneck of AI training; DRAM / NAND cycle dynamics are the most predictable in semis if you read the bit math. Storage hardware sits one layer above. |
| Hyperscalers | The capex source funding the entire AI infrastructure stack. Microsoft, Google, Amazon, Meta, Oracle together drive ~$420B+ of 2026 AI capex. Their disclosures set the demand backdrop for every silicon, memory, and networking thesis below; their own valuation is increasingly an AI platform question (Azure, GCP) on top of legacy software / consumer franchises. |
| Cybersecurity | Recurring revenue software with structural tailwinds; the right valuation lens varies by subsegment (endpoint vs network vs identity vs cloud). |
| Robotics & Autonomy | Earliest stage of the seven. Many names are still proving unit economics. Differentiating real product traction from narrative is the alpha. |
| Infrastructure Software | The application layer above the silicon. Includes observability, virtualization, databases, data clouds. Net dollar retention is the key metric. |
| Networking & Server Hardware | Datacenter switches and routers, server / appliance OEMs, and the PC supply chain. Cycle dynamics + customer concentration risk are the recurring debates; AI buildouts dominate the near term thesis for the networking subset. |
Valuation Framework
One of the biggest tells of weak research is forcing every name into the same valuation model. I use four primary frameworks, picked per company based on business stage and earnings quality:
| Framework | Best for | Used in |
|---|---|---|
| DCF + sensitivity grid | Companies with predictable cash conversion and a defined steady state. Required when terminal year economics are the dominant driver of value. Segment level DCF when the business is a multifranchise compounder where each segment has materially different cash dynamics. | Micron (memory cyclical); Microsoft (segment level: Productivity, Intelligent Cloud, More Personal Computing) |
| Sum of the Parts (SOTP) | Multiline companies where blended multiples obscure the underlying business mix. Each franchise valued on its own EBITDA × multiple. | Broadcom (AI semis × non AI semis × VMware software) |
| EV/Sales + scenario tree | Pre profit names where EBITDA is uninformative. Anchor on revenue, attach probability weighted scenarios for the path to profitability. | IonQ (quantum pre profit) |
| Comp matrix (EV/EBITDA, EV/Sales, P/E) | Cross checking primary models against peer multiples. Forces the primary model to defend its inputs against the market's pricing of similar businesses. | Across all dashboards (in development) |
Scenario discipline
Every dashboard frames Bull / Base / Bear with explicit price targets and probability weights. The weights live in data.js (pBull, pBase, pBear) and the shared engine renders the resulting probability weighted blended fair value directly under the summary KPI grid on every dashboard. Two outcome checks are visible on the same line:
- Blended FV vs current price: the live tactical edge.
- Blended FV vs headline 12-month PT: the internal consistency check. If these two values diverge by more than a few percent, the headline PT is anchored on the base case alone and not honoring the bull / bear tails, i.e., the scenarios are wrong or the weights are.
The reader does not have to take the headline PT on faith. They can see the math behind it and judge for themselves whether the probability assignments are credible.
Data Sources
- SEC filings (10K, 10Q, 8K): primary source for every reported financial figure. Footnoted inline where non obvious.
- Earnings call transcripts: management commentary on guidance, segment dynamics, and forward expectations.
- Investor day decks + supplementary materials: long range planning frames, TAM assertions, capital allocation roadmaps.
- Industry research (TrendForce, IDC, Counterpoint): for share data and market size triangulation where the company doesn't disclose directly.
- Hyperscaler capex disclosures: Microsoft, Google, Meta, Amazon, Oracle filings, the macro backdrop for AI semi and AI memory theses.
- Live quote feed: Finnhub via Cloudflare Worker proxy (key held server side). Stooq CSV as keyless fallback. Patches price driven ribbon values on each page load.
Every figure that isn't from the live feed is curated in a per-dashboard data.js file with an asOf field. Curated figures are reviewed on a quarterly cadence (after each earnings print) and on material developments.
AI Workflow
I use AI tools as an analyst force multiplier, not a substitute for judgment. Specific places they're integrated into this library:
- Earnings transcript digestion: LLMs extract guidance language, segment commentary, and management framing changes from each quarterly call. Output is a structured "what changed vs last quarter" delta that I review before updating data.js.
- 10K risk factor synthesis: LLMs map disclosed risks to my existing bear case scenarios. New risk language that doesn't map to a known scenario gets flagged for incorporation.
- Cross dashboard "Ask the Thesis" widget: embedded directly in dashboards. The user describes a scenario in natural language; the model returns a structured impact analysis against the dashboard's existing thesis and PT math. Implemented as a Cloudflare Worker proxy holding the Anthropic API key server side (same architecture as the live quote feed).
- Document QA across filings: embeddings based search over the corpus of 10Ks and transcripts I'm tracking, used during model updates to find specific past disclosures.
What I deliberately do not use AI for: the rating decision, the price target, and the scenario probability weights. Those are author judgements. The AI tools accelerate the inputs to those decisions; they don't make the decisions.
Update Cadence
- Quarterly: after each earnings print, the affected dashboard's
data.jsis refreshed with reported figures and guidance. Scenarios are stress tested against the new data. PT is updated if the print materially changes the model inputs. - On material developments: guidance preannouncements, M&A, major customer wins/losses, regulatory actions. These trigger an out of cycle refresh.
- Live (intraday): price derived ribbon values (last, YTD, market cap, forward P/E) automatically update from the live quote feed on each page load.
The "Data as of" stamp on each dashboard, and on every hub card, carries a colored freshness badge so a reader can see at a glance how current the underlying snapshot is:
- Fresh: snapshot is less than 30 days old. Within the current reporting cadence; figures are the latest reviewed.
- Aging: 30 to 90 days. Pre next earnings window; still reliable but flagged for upcoming refresh.
- Stale: more than 90 days. Beyond one fiscal quarter; verify against the latest filing before relying on any figure.
Live price driven values (last trade, YTD %, market cap, forward P/E) are independent of this badge, they refresh on every page load via the quote feed. The badge applies only to curated, point in time figures.
Disclaimers
Not investment advice. The research in this library is the author's personal analytical work, prepared for portfolio and educational purposes. Nothing in it constitutes a recommendation to buy, sell, or hold any security. Past performance does not predict future results; forward looking statements are subject to material risk and uncertainty.
All third party trademarks, company names, and product names referenced in the library are the property of their respective owners. The author has no affiliation with any company covered unless explicitly disclosed within a specific dashboard.