AI Final

TAM (Total Addressable Market)

The "Universe": Total demand if everyone who could use the product did so.

SAM (Serviceable Addressable Market)

The "Scope": The slice of the market you can actually serve based on geography, regulation, and tech fit.

SOM (Serviceable Obtainable Market)

The "Target": The realistic share you can win considering competition and capacity.

Top-down Market Sizing

Starting with large market reports and applying filters

Bottom-up Market Sizing

Calculated as: Price x Customers x Adoption Usage

Bottom-up TAM Formula

TAM = eligible dev seats x $ -per seat per year

Hardware Layer

GPUs (NVIDIA H100), AI servers (Dell/HPE), and Networking (Cisco/Broadcom).

Software Layer

Foundation models (GPT, Llama), MLOps platforms (Databricks), and Copilots

Services Layer

Integration, security, and change management (Accenture, Deloitte, McKinsey).

The "Investment Loop"

A "frenemy" cycle where companies (NVIDIA, OpenAI, Oracle) act as each other's investors, landlords, and customers.

The AI Value Loop (6 Steps)

1. Data/Signals

2. Model Improvement

3. Productization

4. Deployment

5. Outcomes

6. Adoption

Jevons Paradox in Knowledge Work

As AI makes tasks easier, the total volume of tasks increases, which can lead to burnout instead of efficiency.

HITL (Human-in-the-Loop)

A framework where AI handles retrieval and synthesis, while humans handle prioritization and ethical judgment.

Generative vs. Agentic Workflows

Generative: Single response (drafting an email).

Agentic: Multi-step planning, tool use, and feedback loops.

Agentic Workflow Pattern

Sense → Plan → Act → Verify → Log

Attention Residue

The cognitive cost of switching tasks, which AI scheduling engines (like Motion) try to minimize.

CAC (Customer Acquisition Cost)

The total cost to acquire one customer (e.g., spend $1,000 on ads to get 20 customers = $50 CAC).

LTV (Lifetime Value)

The total profit a customer generates over their entire relationship with the company.

Churn vs. Retention

Churn: Customers who stop buying/leave.

Retention: Keeping customers coming back.

Upsell vs. Cross-sell

Upsell: Higher-end option ("Large" vs "Medium").

Cross-sell: Adding another item (fries with a burger)

Propensity vs. Uplift

Propensity: "Who is likely to buy?"

Uplift: "Who will buy because we intervened?" (Focuses on "persuadables")

The 4 Growth Levers

1. More Demand (Targeting)

2. Higher Conversion (Search/Recs)

3. Higher Basket/Margin (Pricing)

4. Faster Sales Velocity (Coaching)

The AI Pivot in Targeting

Shifting from broad demographics to behavioral + contextual clusters using embeddings.

Retail Search AI

Moving from Keyword Matching (Old) to Semantic Retrieval (New) to understand user intent.

Recommendation Approaches

1. Collaborative Filtering ("People like you...")

2. Content-based ("You like spicy...")

3. Session-based ("In this moment...") .

General Revenue Formula

Revenue = Traffic x Conversion x AOV x Frequency

Incremental Profit

= (Targeted x Uplift % x AOV x GM)- Campaign Cost

Retail Search AI Annual Value

Value = Search Sessions x (Change in search conversion) x AOV x Gross Margin

Forecasting Value (Inventory)

= (Stockout Reduction x Lost Margin Avoided) +(Overstock Reduction x Markdown Avoided)

Conversation Intelligence

AI (like Gong) analyzing objection patterns and talk-to-listen ratios to coach reps

The "CRM Tax"

The burden of manual data entry and follow-ups that AI automation (like Salesforce Einstein) aims to fix.

Dynamic vs. Personalized Pricing

Dynamic: Adjusting for supply/demand.

Personalized: Risky/Ethical issues (e.g., Delta won't use AI for personalized ticket prices based on personal data)

The Value Pyramid (4 Stages)

1. Efficiency (Automation)

2. Productivity (Throughput)

3. Growth (Revenue)

4. Strategy (Moats) .

Measurement Strategy

Use holdout or geo tests to measure causal impact; never ship models without a measurement plan.

Compute = Power

In AI, more compute translates to more strategic power; hardware is often considered the "real moat"

NVIDIA's Market Role

Dominates both the training and inference hardware markets.

The Hardware Layer

Controls supply (GPUs, TPUs, Networking, Chips)

The Cloud Platform Layer

Controls distribution via APIs, hosting, and fine-tuning (e.g., Azure, GCP, AWS) .

Foundation Models Layer

Owns the "Intelligence Layer" (LLMs, Multi-modal, Agents)

Open Source (Open Weight)

Anyone can see the "recipe" (code) and "ingredients" (data); treated as a public utility (e.g., Llama, Mistral)

Closed Source (VIP Club)

Secret proprietary "sauce" owned by corporations; treated as a premium, protected product (e.g., GPT, Gemini).

Privacy Advantage of Open Source

Models run on your own machine (e.g., a Mac Studio), so your data never leaves your room.

Convenience of Closed Source

The "Easy Button"; no setup required—just plug into an API and go.

Generative AI Stack

Hardware (GPUs, TPUs, Networking, Data Centers, Chips)

Foundation Models (LLMs, Multimodal, Agents)

Cloud Platforms (AZURE, GCP, AWS, APIs, Hosting)

Applications (Copilots, Agents, Search Enterprise AI, SaaS Tools)

Zero-Shot Prompting

Providing a task with no examples; typical accuracy for complex logic is <60%.

Few-Shot Prompting

Providing 2–4 examples to reach ≥90% format compliance and 85%+ accuracy.

Shot count vs. Token Cost

Adding examples (e.g., moving from 3 to 5 shots) has diminishing returns (only +1.5% accuracy) while significantly increasing API costs.

The "Master Prompt" Structure

1. Role,

2. Goal,

3. Context,

4. Input Data,

5. Output Format,

6. Rules,

7. Quality Check .

Constraint-based Prompting

Using strict rules like "Output ONLY JSOM" or "If missing info, output 'Unknown'" to prevent hallucinations.

ROI of AI Data Cleaning

AI shifts effort from manual entry to quality control; a task that takes ~40 hours manually can be done in ~2 hours with AI.

Hallucinations in Business

Confident but wrong/invented outputs, such as fake numbers, fake citations, or code that looks right but fails to run.

The PII Rule

Golden rule: PII count = 0. Never paste names, emails, student IDs, or passwords into unapproved AI tools.

Safe Transformation

Redacting identity while keeping structure

80/20 Rule in Coding

AI drafts ~80% of boilerplate code quickly; humans must lead the critical 20% (security, requirements, performance) .

Verification Target for Code

Aim for ≥90% of unit tests passing before trusting AI-generated code

Red Flag in AI Coding

If code does not run, treat it as a "prompt failure"; revise the prompt rather than blindly copy-pasting

Structured Data

Organized in rows/columns (SQL, Excel); represents 10-20% of enterprise data .

Unstructured Data

No predefined format (Emails, PDFs, Video); represents 80-90% of enterprise knowledge .

Semi-Structured Data

Uses tags or keys (JSON, XML, Logs); sits between structured and unstructured .

Garbage In → Garbage Out

The most important principle in AI; 60% of projects fail due to data readiness rather than model sophistication .

RAG (Retrieval-Augmented Generation)

The dominant enterprise AI architecture that grounds LLM responses in real retrieved data to solve the hallucination problem

The 5 Steps of RAG

1. Store Docs

2. Convert to Embeddings

3. User Question

4. Retrieve Chunks

5. LLM Synthesis

Embeddings

Numerical vectors that capture the semantic meaning of a document chunk

Samsung Data Leak (2023)

Engineers accidentally uploaded confidential source code and meeting notes to ChatGPT, leading to a company-wide ban .

EU AI Act

World's first comprehensive AI law using a risk-based classification system (Prohibited —> Minimal Risk) . Fines up 35M Euro

CCPA / CPRA

Grants California residents the right to opt out of automated profiling and delete data used in AI models .

Value vs. Feasibility Matrix

Quick Wins (High Value/Easy)

Strategic Bets (High Value/Complex)

Fill ins (Low Value/Easy)

Avoid (Low Value/Hard)

Prioritization Scoring Model

Business Impact (30%)

Data Quality (20%)

Technical Feasibility (20%)

Time to Value (15%)

Regulatory Risk (15%).

Priority Threshold

Projects should typically achieve a weighted score of ≥ 7.0 to be prioritized for an immediate roadmap.

Revenue Stream Mix

Products (38%)

Platforms/APIs (29%)

Professional Services (18%)

Data Monetization (15%)

The "Surprise" Costs

Data and Talent are the most underestimated costs, consistently consuming ~57% of a total AI project budget.

Build vs. Buy vs. Partner

Build: Best for core differentiators/IP ownership.

Buy: Best for commodity functions/speed.

Partner: Best for exploring new verticals/regulated entry.

AI Risk Taxonomy

Categorizes threats into four areas:

1. Technical (drift, hallucinations)

2. Business (ROI miss)

3. Ethical (bias)

4. Regulatory (EU AI Act)

AI ROI FORMULA

ROI = ((Net Benefit - Total Investment)/Total Investment ) x 100

Reactive vs. Proactive Support

Reactive: Waiting for the user to report a problem.

Proactive: AI identifies a struggle (e.g., failed checkout) and offers help before a ticket is created.

Deflection Rate

The percentage of customer queries resolved by AI/Self-Service without needing a human agent.

Sentiment Analysis

AI detecting the emotional tone of a message (e.g., Frustrated, Happy, Urgent) to prioritize the most upset customers.

Intelligent Routing

AI analyzing the intent of a message and sending it to the specific human expert best suited to solve it.

Zero-Touch Resolution

A query solved entirely by AI from start to finish with no human involvement.

KPIs (Customer Service)

CSAT (Customer Satisfaction) - A score (usually 1-5) given by the user after a support interaction.

FCR (First-Contact Resolution) -The percentage of issues solved in the very first interaction (AI or human).

AHT (Average Handle Time) - The total time spent by an agent or AI to resolve a single ticket.

NPS (Net Promoter Score) - Measures long-term loyalty by asking "How likely are you to recommend us?"

Cost Per Ticket

= Total Support Team Costs (Salaries + Tools)/Total Tickets Handled

Monthly Savings

(Tickets Deflected by AI) x (Cost per Human Ticket - Cost per AI Interaction)

Primary Goal of AI in Entertainment

To win and maintain attention in crowded markets and increase monetization efficiency.

Content Personalization

Using algorithms to shape what users watch and buy, improving experience at scale.

Monetization Strategy

AI increases efficiency by raising the relevance of ads and content to the individual viewer.

Creative Assistance

AI acts as an assistant for content recognition and accelerating creative workflows.

Primary Goal of AI in Sports

To analyze players, speed up coaching, and optimize game strategy to improve the chance of winning.

Tracking Data (AI IN SPORTS)

Makes physical attributes like spacing, movement, and defensive reactions measurable instead of purely visual

Decision Support

Moves from pure intuition to intuition informed by probabilities (e.g., predicting an opponent's most likely next action).

Core Functions (SPORTS)

Track, Coach, Scout, Strategize, and Perform.

AI Value Loop (Entertainment/Sports)

Step 1: Collect - Gathering behavior and tracking data (e.g., viewer clicks or player movement).

Step 2: Predict - AI Models generate probabilities or rankings based on the data

Step 3: Decide - Human leaders or automated systems choose an action based on predictions

Step 4: Value - The chosen action leads to a business or performance outcome

Industrial AI

Industrial AI focuses on hyper-efficiency, human safety, and supply chain resilience.

Digital Twins

Virtual simulations of physical systems used to "simulate before you spend" and optimize factory planning.

Predictive Maintenance

Using AI to prevent failures before they happen; for example, avoiding downtime events that can cost $50,000/minute

Computer Vision in Factories

Used for quality inspection and safety monitoring on the factory floor.

SDV (Software-Defined Vehicle)

A vehicle where features and functions are primarily enabled through software, creating new revenue streams.

Telematics

Long-distance transmission of computerized information, used in automotive for driver safety and in-cabin AI.

Edge Physics

The intersection of AI data and real-world physical constraints in autonomous vehicle operation.

S&OP (Sales and Operations Planning)

AI helps break silos by moving from batch-based to real-time planning for inventory and working capital.

Inventory Optimization

Using AI to reduce working capital; e.g., freeing $18M in inventory by improving forecast accuracy.