AI summary by TwinMind
Where is the AI Boom Going?
Asianometry
16m 31s
Conference Background & Hot Chips Overview
- Attended Hot Chips semiconductor design conference at Stanford
- Focused on one-on-one conversations rather than formal presentations
- Key takeaways from presentations:
- Meta and Tesla making progress on cheap ASICs for specific AI applications
- New co-packaging optics (CPO) technology emerging as future for inter-chip data transmission
- Transceivers becoming declining technology
- AMD and OpenAI presented minimal new developments
AI Revenue & Market Assessment
- Microsoft Co-pilot performance mixed, not "setting world on fire"
- Estimated ~$1 billion annualized revenue across all Co-pilot products
- Product rebranding scheduled, already generating industry jokes
- Two years post-ChatGPT launch, no major revenue challenger emerged
- Leading AI companies by revenue (excluding ChatGPT/OpenAI):
- 11 Labs (voice generation, NYC): ~$100 million revenue
- Character.AI: highest usage but monetization challenges
- Core users are teenagers with no money
- High AI model costs led to recent layoffs
- Licensed models and sold part of team to Google
Successful AI Business Models
- Strongest product-market fit: companies leveraging LLM core advantages vs replacing humans
- Most promising applications:
- Time-saving for workers handling large information volumes (text/audio)
- Document parsing (Google Gemini demo: 75-page service agreements)
- Developer tools (Cursor AI text editor for code assistance)
- Administrative tasks, customer service, order taking
- Future potential: drive-through order systems for fast food
API Pricing War
- Dramatic price reduction in LLM API tokens:
- Original GPT-4 API (March 2023): $36 per million tokens
- GPT-4o (September 2024): $4 per million tokens
- 9x reduction over 17 months (80% annual reduction)
- Price cuts driven by software improvements and better resource utilization
- Occurred within same generation of Nvidia GPUs
- Historical parallel: Bob Noyce at Fairchild cutting transistor prices from $5 to $0.25, ICs from $20 to $2
Foundation Model Limitations & GPT-5 Anticipation
- Industry consensus: GPT-4 and current models still very limited
- Entire industry waiting for GPT-5 as next breakthrough
- OpenAI viewed as "black box" even by competing AI research labs
- Comparison to IBM PC era: OpenAI defining standards, others following
- Historical parallel suggests eventual need for companies to move beyond OpenAI leadership
Alternative Approaches & Scaling Challenges
- Few companies trying non-scaling approaches:
- Sakana AI (Japan)
- Safe Super Intelligence
- Bay Area remains "scale-pilled" - committed to continued scaling
- Scaling requires immense financial/material resources
- Only few tech giants can afford continued investment
- Growing strain indicators among tech giants regarding capital expenditure levels
- "Game of chicken" scenario: giants must spend tens of billions to avoid being left behind
AI Chip Market Dynamics
- Current ranking: Nvidia dominates, AMD distant second
- Nvidia's future dominance depends on:
- Tech giants' AI chips emerging as major players (Google TPU, Amazon AWS chips)
- AI chip startups breaking out (unlikely given 55+ startups at Hot Chips)
- Semiconductor development challenges:
- 5-6 month design and fabrication cycles
- Cannot rapidly scale like software companies
- Most AI chip startups likely to fail or be acquired as acqui-hires
Intel Crisis & Recovery Scenarios
- Disastrous quarter announcement, 30% stock crash
- Mid-September board meeting considering:
- Capital expenditure cuts
- Business divestments
- Potential company split (like AMD decade ago)
- Recovery timeline: 18A node (2027-2028) as potential savior
- Depends on process node quality and customer adoption
- Long timeline for Intel to "muddle through"
- Chips Act Part 2 expected but funding amounts unclear
Autonomous Vehicles Success Story
- Waymo achieved widespread deployment in San Francisco and Los Angeles
- Transitioned from exclusive to ubiquitous presence
- Won autonomous vehicle wars through brute force approach:
- Mapped roads comprehensively
- Built powerful data collection apparatus
- Identified best performing driving models
- Set up depots and ride-sharing service
- Public treating self-driving cars as mundane rather than revolutionary
Future Outlook
- Industry focus shifting from AI models and chips to data centers themselves
- Semi Analysis selling data center model, potentially bigger than AI accelerator model
- Next two years of scaling work already planned
- Uncertainty beyond that timeframe regarding scaling law capabilities
