AI Startups & AI in Markets

Some thoughts on where the money is, and where it probably isn’t

Danilo Freire
danilo.freire@emory.edu

Department of Data and Decision Sciences
Emory University

Nice to meet you!

A bit about me

Visiting Assistant Professor in the Department of Data and Decision Sciences

MA from the Graduate Institute Geneva, PhD from King’s College London, Postdoc at Brown University, Senior Lecturer at the University of Lincoln, UK

Research interests: computational social science, experimental methods, policy evaluation

I teach DATASCI 185: Introduction to AI Applications, DATASCI 350 - Data Science Computing, and DATASCI 385 - Experimental Methods

Some thoughts on AI startups

The $600 billion question

One frame to hold onto tonight

  • Sequoia’s David Cahn, 2024: AI’s $600B Question
  • The hardware build-out implies $600B in yearly revenue
  • Actual AI revenue is a tiny fraction of that
  • The frame for tonight: where’s the cash coming from?
  • Ask this about every startup I mention

The gap between AI capex and AI revenue. Source: David Cahn, Sequoia Capital (2024)

Where the money is going

Record volumes, extreme concentration

  • February 2026: $189B in venture funding, an all-time record
  • 83% went to three firms:
    • OpenAI ($110B)
    • Anthropic ($30B)
    • Waymo ($16B)
  • 2024 corporate AI: $252B (Stanford AI Index)
  • US: $109B. China: $9.3B. UK: $4.5B
  • If you are not one of those three, your fundraising looks nothing like this

Sam Altman, CEO of OpenAI, whose single $110B raise accounted for over half of February 2026’s record month. Photo: TED / Wikimedia Commons (CC BY 4.0)

The three layers of AI

Different layers, wildly different economics

  • Models: OpenAI, Anthropic, Google, xAI, Meta, DeepSeek, Kimi, Z.ai
    • Capital-heavy, very few winners, most still losing money
  • Infrastructure: NVIDIA, the cloud, vector DBs, evaluation tools
    • Quiet, profitable, picks-and-shovels
  • Applications: vertical tools on top. Cursor, Windsurf, Harvey, Perplexity
    • Low cost to start, brutal competition, thin moats
  • So far, most of the value has gone to layers 1 and 2
  • Most startups live at layer 3, and most die there
  • Good primer: Sequoia’s AI in 2025

The wrapper problem

Why “we use GPT-4” is not a business

  • A “wrapper” calls someone else’s model with a nice interface
  • My acid test: if OpenAI shut down its API tomorrow, would the company still exist?
  • Three structural problems:
    • Every call costs money, so margins are thin
    • The underlying model gets cheaper and better every quarter
    • If the idea is obvious, 50 other teams are already building it
  • Wrappers can make money. They rarely become real companies

NVIDIA H100, the chip your API calls are running on. Photo: Geekerwan / Wikimedia Commons (CC BY 4.0)

Inside Y Combinator

What the top accelerator tells us about the AI wave

Garry Tan, CEO of Y Combinator. Photo: Web Summit / Wikimedia Commons (CC BY 2.0)

What investors actually want

Four questions to ask about any AI company

  1. Proprietary data. Does using the product produce data that improves it?
  2. Workflow lock-in. Once a customer is using it, how painful is switching?
  3. Distribution. Who already has the customers? Sales, partners, brand, virality?
  4. Unit economics. Does each customer earn more than they cost in compute?

If a founder cannot answer all four clearly, that is your answer

Case study: winners

Three companies doing it right (so far)

Cursor (Anysphere) code editor

Harvey legal AI

Perplexity AI search

  • $20B valuation (Sep 2025)
  • 40x in 18 months from $500M
  • ~22M monthly users, ~$200M ARR
  • The free tier is the moat

The pattern: pick a domain, own the workflow, compound the data

Case study: losers

Brilliant teams, real tech, no business

  • Inflection AI (chatbot “Pi”)
    • Raised ~$1.5B, burnt through most of it
    • Microsoft paid $650M to license the tech, hired the 70-person team
    • Lesson: a model alone is not a business
  • Adept (AI agents)
    • Raised ~$415M, Amazon hired the founders in June 2024
    • Lesson: if your moat is talent, you can be hired away
  • Humane AI Pin (wearable device)
    • Raised $230M, sold to HP for $116M in Feb 2025
    • Lesson: novel form factors rarely beat existing distribution

A 10-minute startup evaluation

Questions to ask, in order

  1. Google the founders. Are they domain experts, or random pivoters?
  2. What data do they have that nobody else does?
  3. What happens to their margins when GPT-5 is half the price?
  4. Ask for a live demo. Break it with your own input
  5. Look at who the customers are, not how many
  6. Search Product Hunt. 30 clones means they are not special
  7. Who would buy the company if the founders quit tomorrow?

Red flags in a pitch

What to listen for when a founder speaks

  • “We use Claude” with no plan beyond that
  • Vague answers about where the data comes from
  • No discussion of unit economics
  • Demo-driven, no real customers
  • They talk about the model more than the customer
  • Paul Graham, 2025: “Not every new company needs to be about AI”
  • Founders matter more than the idea

Paul Graham, co-founder of Y Combinator. Photo: Wikimedia Commons (CC BY-SA 2.5)

AI and the markets

A tour of the quant world

The firms actually using AI for trading

NYSE trading floor. Photo: Kevin Hutchinson / Wikimedia Commons (CC BY 2.0)

What LLMs are actually used for in finance

Mostly not for picking stocks

  • Sentiment analysis: news, earnings calls, analyst reports, X, Reddit
  • Document processing: summarising 10-Ks, extracting data from filings
  • Alternative data parsing: unstructured text → trade signals
  • Compliance and risk: flagging suspicious trades, monitoring communications
  • Pictet AM says up to 50% of alpha in their AI strategies comes from ML “conditioning” of existing factors
  • Survey: From Deep Learning to LLMs in Quant Investment

The alpha decay problem

Why no AI strategy stays ahead for long

  • An alpha is an edge that earns returns above the market
  • Once discovered and widely traded, it disappears
  • Every serious quant fund is in a permanent arms race
  • LLMs are now used to generate candidate factors automatically
  • The truth to carry with you: there is no permanent edge in markets
  • Today’s miracle is tomorrow’s consensus, and the day after that it is in an ETF

So why does Renaissance still win?

What thirty years of returns can teach us

  • Medallion reportedly returned ~30% in 2024, when many AI-heavy funds struggled
  • Three reasons it keeps working:
    1. Closed to new money since 1993, so it stays small
    2. Extreme secrecy, so strategies do not leak
    3. No AI hype: physicists and statisticians, not ML influencers
  • Beating the market is extremely hard, but possible. It just looks like forty years of quiet, boring work 😅

Medallion Fund annual returns vs S&P 500, 1988-2021. Chart: Of Dollars And Data. Data: Bradford Cornell

Retail AI tools for investing

What you have, what it can’t do

  • eToro global survey: 19% of retail investors used AI in 2025, up from 13%
  • Common tools: ChatGPT, Claude, Perplexity Finance, Gemini, Magnifi, Composer
  • What ChatGPT cannot do well for stock picks:
    • No real-time data unless you connect a tool
    • Hallucinates on smaller companies
    • Trained to please you, not to disagree
    • Fitted to recent history, bad in downturns
  • StockBench (2025): most LLM agents failed to beat buy-and-hold

Your AI investing checklist

Before you trust any AI finance tool with real money

  1. Does it pull live data or use a stale training cut-off?
  2. Does it cite sources you can check?
  3. Ask it to argue the opposite side of a company you like
  4. Test it on a small company you know well. Watch for hallucinations
  5. Ask for confidence intervals. A good tool admits uncertainty
  6. Write down your own reasoning before acting on its answer
  7. Never give an AI tool access to your brokerage account

My personal veredict

What AI can and cannot do for an individual investor

Can do well

  • Read filings and summarise earnings calls
  • Compare companies on fundamentals
  • Scan news for sentiment shifts
  • Draft research notes
  • Catch arithmetic mistakes
  • Explain jargon in plain English

Cannot do well

  • Predict prices reliably
  • Replace your judgement
  • Beat a low-cost index fund long term
  • Tell you when it is wrong
  • Know what happened yesterday

A useful mental model: AI is a research analyst who is fast, tireless, slightly drunk, and occasionally lying

Three takeaways

The whole talk compressed into three sentences

  1. AI startups: ignore the model, look at the moat. Ask about data, workflow, distribution, and unit economics. If any is missing, walk away
  2. AI in markets: it works, mostly as a research layer. The firms that actually beat the market are small, secretive, and decades in. Your ChatGPT is not one of them
  3. For you: the cheap, obvious applications are saturated. Real opportunities are where you personally have data, distribution, or domain knowledge nobody else has

Thank you! 🙏🏻

Questions and further reading

To start the discussion:

  • Which AI startup has the most defensible moat? Why?
  • $10,000 in one AI company: which, and what’s the exit?

Further reading