AI summary by TwinMind
The Coming AI Startup Bust
Asianometry
13m 18s
AI Startup Bubble Overview
- AI startup bubble currently forming with unusual valuations that don't make economic sense
- Products built in highly competitive areas with daunting economics
- Many startups will fail but this is normal market process
- Bust ahead but overall AI investment boom benefits are real
Extreme Valuation Examples
- Sierra (Brett Taylor co-founder): seeking $4+ billion valuation
- Conversational AI product with minimal revenue (~$1M estimated)
- 200x revenue multiple
- Perplexity AI: raising $500M at $8-9B valuation
- $50M annual revenue pace
- 160-180x price-to-revenue multiple
- Poolside: $500M raise at $3B valuation
- AI-powered software developer platform
- Imbu: raised $200M+ at $1B valuation in Sept 2023
- Zero revenue at time of funding
- Glean: $250M raise, valuation doubled from $2B to $4.5B in 6 months
- $55M annualized revenue = 82x multiple
- OpenAI comparison: $157B valuation, $4-5B revenue = 42x multiple
Return Analysis Problems
- Optimistic exit scenarios: $40-50B valuations (Atlassian/Snowflake level)
- $4B raise → 8-10x return sounds good but:
- 10-12 years average SaaS time to market
- Annual return only 18-25%
- Not high considering risks involved
Celebrity Founder Phenomenon
- High valuations/low revenue indicates betting on team/AI celebrity status
- Examples:
- Ilya Sutskever (Safe Superintelligence)
- Mira Murati (former OpenAI CTO raising for new startup)
- Fei-Fei Li (World Labs): $100M at $1B valuation
- Streaming bubble parallel: Ryan Murphy, Shonda Rhimes, JJ Abrams deals
- Warner Brothers paid $250M for Abrams deal (Sept 2019)
- Logic: deliver Game of Thrones-style hits for subscriber growth
- Few years later: some hits produced but deals wouldn't be renewed today
Economic Challenges for AI Applications
Cost Side Issues
- AI apps more expensive to train/run than traditional cloud apps
- Higher usage = higher costs for provider
- Must either dilute quality or pass costs to users via price increases/per-use models
- Specialized labor costs (e.g., legal AI firms need elite in-house lawyers)
Revenue Side Problems
- Basic thinking: capture cost savings by replacing jobs/salaries
- AI agents likely need human-in-the-loop, blunting monetary gains
- Even successful replacements face competition from cheaper AI products
- Margins get competed away
Technology Competition Dynamics
- OpenAI started with 2-year lead, still strong revenue but tech lead shortened
- All researchers talk, access same hardware, follow similar precepts
- New innovations (like GPT-o1) incorporated by competitors within months
- Every AI product ~6 months apart from competitors
Overcrowded Categories
AI Coding Assistants Example
- Poolside + other $100M+ funded players:
- Magic, Codium, Cognition, Devin, Augment, Tab9
- Plus GitHub Copilot (market leader)
- Hard Disk Drive parallel (1977-1984):
- 43 manufacturers raised $400M from VCs
- Over half funding in last 2 years
- Industry grew to 70 firms, too similar to each other
- Too modular/easy to replicate
- PC market pullback 1983-1984 killed many startups
Macro Investment Context
- Mega-cap companies pouring billions into chips/data centers
- American companies spending $26B/year on data centers (up 60% from last year)
- Equals spending on retail stores + restaurants combined
- TSMC Q3 2024: seeing "more demand than ever dreamed about"
Potential Outcomes
Optimistic Scenario
- Beginning of magnificent investment run
- Revolutionizes work, brings prosperity across classes
- Creates glorious future
Pessimistic Scenario
- Pull-forward demand, data center investment pauses 3-4 years
- Similar to post-COVID boom pattern
- Difficult but not devastating
Bubble Comparison to Historical Examples
- Past major busts: Japanese real estate (1980s), telecom fiber (1990s), US housing (2000s)
- Those were destructive because involved high leverage
- Current AI boom: Amazon/Microsoft/Google funding from monopoly profits
- No "AI bonds to grandmas and pension funds" yet
- Larger economy should ride out AI startup bust fine
Impact on Different Stakeholders
VCs and Investors
- Sophisticated investors who know the risks
- Can engineer soft exits (example: Inflection AI)
- Raised millions for chatbot competitor to ChatGPT/Claude/Gemini
- Failed but Microsoft deal helped staff/investors recover money
- Already rich and diversified, can handle losses
Startup Employees
- Most vulnerable group: ordinary folks with families
- Parallel to streaming content bubble fallout
- Hundreds in film industry (especially LA) left in lurch after bust
- Advice for startups: manage money prudently, extend runway, focus on revenue streams ASAP
Market Testing Benefits
- Capital flows help evaluate new businesses through market testing
- Examples of successful market testing:
- Prosumer chatbot space: Anthropic's Claude succeeded, Inflection's Pi failed
- Money shifted to vector embedding models, then AI agents
- If AI agents fail, flows will reroute (possibly to robotics)
- Process develops products, gains traction
- Most fall by wayside but few will stick and change the world
