How to Make Money with AI

If you’re here from the lecture, welcome! Add me on LinkedIn, and tell me what you build. I’m always happy to chat about AI, startups, and business models.

The slides and source code are available under CC BY-SA 4.0 license. Every figure in the deck is cited in the references at the end.

Before you build on AI, ask:

1. Does a heavy user make you money, or lose it?
2. What’s left of you the day the lab ships your feature?
3. Can you swap models when the landlord raises the rent?
4. Does your edge survive a model 10× cheaper?
5. Who pays the bill when the investor money runs out?

Reading List

The most useful sources from the lecture, in the order we covered them.

The gold rush and the value chain

• Gartner’s AI spending forecast, the source of the US$1.48 trillion we started with
• AI’s five-layer cake, Nvidia’s own view of the stack
• Emerging Architectures for LLM Applications from a16z, for what sits between the model and the app

Chips and infrastructure

• How Much Does an NVIDIA NVL72 Cost? from Wing VC, a bill of materials for one rack
• Inside the world’s most powerful AI datacenter, Microsoft’s tour of Fairwater
• OpenAI’s Stargate: where the US sites stand and Frontier LLM training runs can’t get much longer from Epoch AI, which is the best single place for data on compute
• Inside Memphis’ fight against xAI, for what a datacentre costs the town next door
• Neoclouds: AI’s $150 Billion Middle Layer That Owns Almost Nothing, on who carries the risk between the chips and the labs

Models

• OpenAI’s key concepts, for what a token is and why you pay per token
• Towards Understanding Sycophancy in Language Models from Anthropic, on why the model tells you what you want to hear
• AI model price charts, the cost-versus-intelligence charts from the lecture, regenerated as new models launch

Pricing and unit economics

• The New Business of AI from a16z (2020), the classic on why AI margins look nothing like software margins
• Margins by Sector from Aswath Damodaran, to compare AI against every other industry
• Character.AI’s first-day declaration in its Chapter 11 case: a beloved product’s unit economics, in the company’s own words under oath
• OpenAI is losing money on its pricey ChatGPT Pro plan, the heavy-user problem at the top of the market
• Three companies repricing in public: Cursor, Replit and GitHub Copilot
• The Path Forward for Legal AI from Contrary Research, on pricing a vertical AI product

Your supplier is your competitor and landlord

• Windsurf says Anthropic is limiting its direct access to Claude, and Cognition’s account of what happened next
• Cursor’s popularity has come at a cost from Newcomer, on paying the landlord who competes with you
• Get the PDF outta here, the day one ChatGPT update wiped out a category
• Introducing Claude Design, and what it did to Figma’s share price

Liability and extraction

• Air Canada’s chatbot lawsuit, where the bot’s promise was the company’s promise
• A car dealership added an AI chatbot to its site, and an AI coding tool wiped a company’s database
• Anthropic’s $1.5 billion copyright settlement, a legal bill the size of a training run
• IMDA’s Model AI Governance Framework for Agentic AI, Singapore’s voluntary approach

The moat

• LLMflation from a16z, on the ~10× yearly fall in the price of a given capability
• The State of Generative AI in the Enterprise from Menlo Ventures, on why total spend rises anyway
• OpenRouter’s rankings, to watch which models people actually use
• 7 Powers by Hamilton Helmer, the framework behind the “moats that hold” slide
• Meta partners with Midjourney, the profitable one selling to the giant instead of being eaten by it

Who makes the money?

• OpenAI and NVIDIA’s 10-gigawatt partnership, the “up to $100B” deal that became a $30B stake
• Nvidia weighs a $250 billion guarantee for OpenAI’s Ohio campus, the chip layer underwriting its own demand
• The Economist on Nvidia funding OpenAI’s GPU purchases, source of the “$10bn for every $35bn” estimate

More from me

• Early Stage AI Startups, last year’s BSN4811 lecture, for startup stages and VC funding
• My AI Platforms Lecture, including open-source as a defensive and offensive maneuver
• Chesterton’s Fence, or in short: “Do not remove a fence until you know why it was put up in the first place.”

As always, turn your critical thinking skills on and carefully engage with the claims each source makes. Several of the figures in this lecture come from deals still in negotiation; check what happened since.