Case Studies Through the Lens of the Layers

- Track 01 · Software & SaaS

Fast cycles, model layer dominates the story

Classic L2–L7 dynamics, foundation models commoditize the surface, distribution becomes the moat, and L8 (memory + workflow) is the unbuilt layer everyone is racing toward. Cycles measured in quarters.

L1 MIS-PACKAGED AS L79 min

Stack Overflow: When Your Community Becomes Training Data

Stack Overflow's traffic dropped roughly 35–50% after ChatGPT shipped. Fifteen years of community-built knowledge, packaged as L7b content and scraped into L2 training sets. The community that built the data captured none of the value; the model layer captured all of it. A textbook case of L1 data mis-packaged as L7 content.

~100M (2022)~55M (2025) ≈-45%

L1L2L7

Read

STRUCTURAL DIVERGENCE7 min

Apollo vs ZoomInfo: Same Layer, Opposite Strategies, Different Fates

Both sit at L1b, proprietary data. But Apollo went API-first and headless. ZoomInfo charges premium for an L7b UI wrapper. In an agent-first world, the UI tax is a liability. The data refinery wins.

$24B~$5B -79%

L1L7

Read

DEEP DIVE · CUSTOMER EXPERIENCE10 min

Sierra's Memory Moat: Why L8 Beats Salesforce's Agentforce

Sierra and Salesforce Agentforce look like the same product on stage, an AI agent that resolves customer issues. The Cube projection shows they are structurally opposite. Sierra was architected as L1c behavioral data + L5d operating playbooks + L8c network learning from day one: every resolution compounds into per-customer memory. Agentforce is L5 bolted onto Salesforce's existing L1, with no compounding loop. Same demo, opposite trajectories.

$0 (2023)$10B+ (2025) Compounding

L1L4L5L8

Read

MODEL LAYER TRAP9 min

Stability AI vs Midjourney: Why Open-Source L2 Couldn't Monetize

Stability AI open-sourced Stable Diffusion and watched the L2a it created become free infrastructure for everyone *except* Stability. Midjourney kept the model closed, built an obsessive Discord community, and compounded aesthetic memory at L8c. Same underlying technology, opposite layer architecture, 100× valuation gap. The cleanest L2-vs-L8 lesson in the open-vs-closed model debate.

$1B (2022)Restructured (2024) ≈-90%

L2L7L8

Read

THE FIVE ERAS · STRUCTURAL THESIS11 min

From Dashboard to Skill Hire: The Death of Per-Seat Software

Software has moved through five distinct eras of human–machine division of labor. We are mid-transition between Era 3 (The Dialogue, human directs, AI builds) and Era 4 (The Workspace, AI orchestrates, human supervises). Era 5 (The Skill Hire, the agent IS the worker) arrives by 2028. Per-seat pricing is structurally dead in Eras 4–5 because the seat itself goes away. Every product roadmap needs to be re-priced and re-architected along both the customer axis and the depth axis.

L5L6L7L8

Read

VERTICAL STACK7 min

Harvey AI Through the Layers

Harvey is built across four sublayers, L1b (licensed case law), L3a (compliance gates), L5b (legal reasoning scaffolds), L8d (institutional memory of matters). A useful case for mapping how a vertical-AI company actually stacks up, and where horizontal platforms can and can't reach.

- ~$1.5B (Mar 2026) Public reporting

L1L3L5L8

Read

CONSULTING × MODEL LAYER9 min

McKinsey + OpenAI (Lilli): When the Consulting Firm Owns the Memory, Not the Model

McKinsey didn't build a model. It built Lilli, an internal assistant trained on 100,000+ McKinsey documents, 70 years of proprietary studies, and the firm's named expert network. OpenAI provides L2. McKinsey owns L1 (the IP) and L8 (the firm's institutional memory). The consultant doesn't get disrupted by the model, the consultant rents the model and keeps the moat.

0 (Jun 2023)70%+ of firm weekly Compounding

L1L2L6L8

Read

VERTICAL STACK8 min

BloombergGPT: Why a 50B-Parameter Model Beats GPT-4 in Finance

Bloomberg trained its own 50B-parameter model on 40 years of proprietary financial data. Smaller than GPT-4. Better at finance tasks. The reason isn't the model, it's that Bloomberg owns the terminal (L4a API + L7c embedded), the data (L1b proprietary), the compliance posture (L3a regulated), and now the model (L2b specialized). Four sublayers in one regulated vertical.

L1L2L3L4

Read

L5 + L8 IN PRODUCTION8 min

Klarna: 700 Agents Replaced, $40M Saved, The First Honest Number on Agent Economics

Klarna's AI assistant handled 2.3M conversations in its first month, the workload of 700 human agents, with equal customer satisfaction and faster resolution. The headline is the cost. The structural story is that Klarna owned the customer data (L1c), the workflow (L5a), and the resolution memory (L8c). The model was a commodity input.

$0$40M Year 1

L1L5L8

Read

L7 ON RENTED L27 min

Devin at $2B: The Autonomous Coder With No Layer Beneath It

Cognition raised at $2B for Devin, the 'autonomous software engineer.' Impressive demo. Structural problem: Devin sits at L7c (embedded copilot surface) on top of someone else's L2a (Anthropic/OpenAI foundation models), with no L1, no L8, no L4. Meanwhile Cursor owns the IDE, the L4a railroad of where code is actually written.

- $2B Fragile

L2L7

Read

L4 DISTRIBUTION8 min

Perplexity vs Google: The Answer Engine vs The Default

Perplexity built a better answer experience. Google owns Chrome, Android, Safari's default, and the URL bar of the internet. Perplexity raised at $9B. Google shipped AI Overviews to 1.5B users in a quarter. The lesson: when the L4a owner ships your L7a, you don't win by being better.

L4L7

Read

L4 + L6 STACK8 min

Cursor at $9B: The IDE That Quietly Became the Most Important L4 in AI

Cursor isn't a model. It isn't an agent. It's the editor, the place developers spend 8 hours a day. By owning L4a (IDE distribution) and layering L6c (agent orchestration) and L8d (per-codebase memory) on top, Cursor has become structurally more defensible than the agents that run inside it.

- ~$9B Compounding

L4L6L8

Read

L3 TRUST PLAY7 min

Anthropic's Enterprise Wedge: Selling L3 When Everyone Else Sells L2

OpenAI sells the smartest model. Anthropic sells the most *governable* one. Constitutional AI, transparency, conservative refusals, SOC 2, FedRAMP, BAAs, Claude's enterprise pitch is L3a compliance + L3c safety bundled with competitive L2a. In regulated buyers' procurement, L3 is the decisive layer.

L2L3

Read

L1 + L4 STACK7 min

Adobe Firefly: The Only Image Model an Enterprise Can Legally Use

Midjourney is more aesthetic. Stable Diffusion is more open. Firefly wins enterprise because Adobe trained on licensed stock, and indemnifies customers against IP lawsuits. L1b (licensed data) + L3a (legal indemnity) + L4a (Creative Cloud distribution) is the only stack that survives an enterprise legal review.

L1L3L4

Read

L8 WITHOUT L28 min

Character.AI: The L8 Memory Moat That Couldn't Stand Without L2

Character.AI had real L8b memory, billions of personalized conversations, deep user attachment, the largest companion-AI user base. What it didn't have: an L2a it could afford. Compute costs ate the unit economics, and Google licensed the team and tech for ~$2.7B. A clean lesson in why L8b alone, without owned L2a or a profitable L4, is a vulnerable position.

$5B (last round)~$2.7B (license) Acquihire

L2L7L8

Read

L1 + L6 ENTERPRISE STACK8 min

Glean at $7.2B: The Enterprise Memory Layer Microsoft Was Supposed to Own

Glean indexes every document, message, ticket, and meeting inside a company, then makes it queryable by AI. That index is L1c behavioral data (proprietary to each customer), the orchestration is L6d context management across apps, and the cross-app memory is L8d institutional knowledge. Microsoft 'should' own this with Copilot. They don't, and Glean's $7.2B valuation says the market noticed.

$2.2B (2024)$7.2B (2025) +227%

L1L6L8

Read

SAASPOCALYPSE · SURVIVOR PATTERN9 min

Apollo: How Giving Up the SaaS Stack Became the Smartest Bet in B2B Data

Apollo spent a decade building the full GTM SaaS stack, sequencer (L5), dialer, full app (L7), workflow platform (L8). Then it noticed what every horizontal SaaS will eventually notice: when Claude and ChatGPT become the command center, marketers don't want to log into ten apps. Apollo's response was structurally radical: keep the L1b data moat (300M+ contact profiles), become the default MCP connector to the L2 layer, and let the surface that took a decade to build quietly recede. Free distribution from inside the AI command center. The SaaSpocalypse survival pattern, in real time.

$1.6B (2023 Series D)Strategic re-rate, L1+L2-connector Survived the SaaSpocalypse

L1L2L7

Read

WORKED EXAMPLE · GATEKEEPER ARBITRAGE8 min

Dripify vs LinkedIn: The L7 Arbitrageur Living Off an L1 Gatekeeper

LinkedIn owns the highest-value B2B identity graph on earth (L1b) and rate-limits it to protect the surface (L3 gate). That creates a massive gap between the marginal cost of automated outreach (pennies of cloud + proxy) and the perceived value of a single closed enterprise deal ($10K–$100K). Dripify lives entirely in that gap, a cloud-based L7 arbitrageur charging $39–$99/seat/month to bypass the gatekeeper's friction. The canonical illustration of Observation #6: the gatekeeper tax is always arbitraged.

L7 arbitrageur on rented L1Cat-and-mouse with L3 gatekeeper Perpetual margin under perpetual threat

L1L3L7

Read

- Track 02 · Vertical & Regulated

L8 sits above the model and slows commoditization

Healthcare, legal, finance. The model is necessary but not sufficient, clinical workflow, regulatory clearance, billing codes, and liability assignment form an L8 that takes years to build and is hard to replicate from above.

Structural reads through the 10-layer framework. The author is a product strategist applying the framework, not a domain expert in clinical medicine, law, or financial regulation.

VERTICAL & REGULATED · MEDTECH8 min

Tempus AI: When the Data Layer Sits Inside the Clinic

Tempus stitched together oncology-grade clinical and molecular data (L1b), trained vertical models on top (L2b/L3a), and embedded the output into oncologist workflow (L8d). The structural read: in regulated medicine, L8d (the clinical workflow + reimbursement path) sits above the model and slows commoditization meaningfully.

$8B (IPO Jun 2024)~$6B (May 2026) Range-bound

L1L3L8

Read

- Track 03 · Physical & Industrial

L-1 is the layer no AI-only entrant can replicate

Agriculture, robotics, autonomy, energy, manufacturing. The model is the easy layer. The hard layers are L-1 (physical assets + edge silicon + fleet density) and L8 (operating workflow + financing). Cycles measured in years, not quarters.

Structural reads through the 10-layer framework. The author is a product strategist applying the framework, not a domain expert in robotics, autonomy, or industrial systems.

PHYSICAL & INDUSTRIAL · AGROTECH9 min

John Deere: Why the Tractor Is the L-1 Moat

See & Spray puts vision models on the boom of a sprayer to herbicide weeds, not crops. The model is the easy part. The fleet of instrumented tractors with L0e edge compute, the L1d outcome-data flywheel, and the L8d dealer-financed install base, those are the layers no AI-only entrant can replicate.

$130B (early 2024)~$140B (May 2026) Stable / up

L-1L1L8

Read

PHYSICAL & INDUSTRIAL · AUTONOMY10 min

Tesla vs Waymo: Two Bets on Which Layer Wins Autonomy

Same end-state, robotaxis, two opposite structural bets. Tesla bets fleet-scale L0e edge compute (cameras + custom silicon on millions of cars) plus L1c behavioral data wins. Waymo bets dense L1b (HD maps + lidar-grade sensing in geo-fenced cities) plus L8d (operations + regulatory permits) wins. The framework reads them as different layer ownership, not different products.

Tesla: millions of FSD-enabled vehiclesWaymo: paid driverless in ~5 cities Different bets

L-1L1L8

Read