AI summary by TwinMind
Is the AI Boom Real?
Asianometry
13m 43s
Sam Altman's Chip Venture Overview
- 3-week travel through Japan and US studying AI/chip boom landscape outside Taiwan
- Sam Altman (OpenAI) chip venture news evolution:
- Initial reports during OpenAI ouster: raising money from Middle East investors for "just billions"
- Conversations with TSMC people in Abu Dhabi and US government
- Wall Street Journal report: eye-catching $1-7 trillion figure
- Silicon Valley consensus: number too high, possibly negotiating tactic
- Once talking trillions, hundreds of billions feels less substantial
- The Information context: Altman privately clarifies figure represents total investments needed
- Includes real estate, power for data centers, chip manufacturing
- Spread over multiple years as ecosystem concept
Industry Financial Context
- 2023 total semiconductor sales: ~$520 billion
- 2023 total capital expenditures (per SEMI trade group): ~$140 billion
- Even spread over 5 years with diluted expenses = COVID-like step function upward
- Semiconductor industry conservative, many "old heads" seen boom/bust cycles
- Sam not first to claim use case will "change the world"
Scaling Laws Foundation
- Concept from 2020 OpenAI paper "Scaling Laws for Neural Language Models"
- Core principle: more data + compute = better results (less loss)
- Enabled GPT series development
- Ilya Sutskever (OpenAI co-founder) insight: neural networks weren't good because too small
- 1000 neuron network can't do anything regardless of learning quality
- Much larger networks can do unprecedented things
- GPT progression: GPT-3 big → GPT-4 bigger/better → GPT-5 will be even bigger/better
- No indications scaling laws breaking apart
Scaling Laws vs Moore's Law Parallels
- Different content but similar industry effects
- 1980s-1990s: Moore's Law became semiconductor industry rallying cry
- Metronome keeping time from Santa Clara to Tokyo to Hsinchu
- Scaling laws potential similar impact on AI industry
- Simple, understandable rallying cries driving R&D roadmaps for years
Arguments Against Scaling Laws
- Data limitation: pulled all existing internet data
- Workarounds possible if money available
- Physics problems cropped up with Moore's Law, industry found solutions:
- High-K metal gate, FinFET, DUV lithography, dry etch, ion implantation, SEM
- Key question: whether money exists to continue driving investment
- Reference: Dario Amodei's posts "Will Scaling Work" and podcast
NVIDIA Competition Landscape
- Multi-pronged attack on NVIDIA's AI accelerator profits
- Question: Is NVIDIA fortress as assailable as appears?
- Jensen's response: aggressively ramping annual accelerator lineup updates
- Reminiscent of old NVIDIA graphics card wars (new product every 6 months)
- NVIDIA maintains rollout speed by shipping before testing
- Use latest computer software design/emulation tools
- Model and ship new GPU designs without first fabbing physical chip
- Jensen quote: "if we taped out chip, got it back from fab, started working on software, found bugs, taped out again - we would have been out of business"
NVIDIA Strategy Challenges
- Tape out "perfect chips" inevitably causes customer problems
- Deployment issues with drivers/affiliated software won't work well initially
- NVIDIA brand can take hit, AI chip startups cannot
- Takes several years to build team good enough to ship working product
- High iteration speed rough on customers
- No choice but buy what's available
- Stings spending tens of thousands only to have vastly better version shortly after
Tech Giants as Main Threat
- NVIDIA less concerned about startups/established silicon players
- More worried about tech giants (Microsoft, Google, etc.)
- Tech giants driving current AI investment spend
- Most incentive to cut NVIDIA out of margins using custom ASICs
- Historical precedent: Apple 2E third generation powered by custom ASIC
- Modern example: Google TPU (fourth iteration)
- Makes up big part of Google's compute advantage over OpenAI/other AI labs
Microsoft's Vertical Integration Push
- CEO Satya Nadella Q2 2024 earnings: most Azure AI Services usage from inference vs training
- Inference activities easier to push to custom ASIC
- Working on vertically integrating database stack parts
- Recent report: developing network card to shuttle data between servers
- Vertical integration implies either:
- AI database stack scale so large/costly every penny counts
- Industry growth petered out, competing on price
- Both scenarios feel weird to consider
NVIDIA's Medium-Term Outlook
- Takes time for tech giants to ramp up AI chip designs
- First Google TPU was very bare bones product
- Short term: things remain as they are
- Medium/long term: NVIDIA must push hard to win performance-cost crown
- Cost benefits of vertical integration significant enough to matter
AI Boom Reality Assessment
- Key question: Is OpenAI boom "real with capital R"?
- Massive AI progress past year (e.g., OpenAI Sora text-to-video model)
- Technology impressive but doesn't seem like hit product by itself
- Need embedding into compelling products
- Who making money from actual consumers?
Consumer Adoption Comparison
- Moore's Law investments driven by real downstream demand:
- Military → consumer electronics (radios, calculators) → PC → smartphones → cloud computing
- All things ordinary people wanted/used
- Unlikely truly large AI investments happen unless ordinary consumers buy services majorly
- Financial Times report: ChatGPT hit $2 billion run rate revenue December 2023
- Only solid news indicating people paying real money for LLM service
iPhone Growth Benchmark
- iPhone released June 2007
- 2009: 20 million iPhones sold, $13 billion revenue
- Growth trajectory: 40 million → 72 million → 125 million following years
- iPhone = greatest consumer technology product in history
- ChatGPT still one of fastest growing consumer tech products
- Altman: ChatGPT generates 100 billion words/day
- Facebook took time to produce meaningful revenue, invested in audience building
- ChatGPT possibly doing same but AI has shorter leash due to costs
Microsoft Copilot as Key Indicator
- Industry watching Microsoft Copilot subscription service closely
- $20/month offering for people and enterprises
- If catches on: "really off to the races, sky is the limit"
- If not: need to adjust mental models
AI in Existing Products
- LLM revolution most "real" when embedded in known products
- Google Performance Max ads: AI automates creation, deployment, targeting
- Apparently generates tens of billions of dollars
- Apple disabled cookie tracking 2021 with App Tracking Transparency
- Meta initially said would cost company $10+ billion in sales
- Meta clawed back targeting accuracies using AI, collecting data from conversions API
- Maybe killer app for AI is just "more ads"
AI as Enabling Technology
- Ben Thompson and others: AI is enabling technology
- Makes rich richer rather than creating new class of rich people
- Unlike PC or electrification
- If this is all AI is, then "bit disappointing"
- Nevertheless makes AI boom real
