Gavin Baker - AI Market Jitters - [Invest Like the Best, EP.485]

July Felt Like 2022 But Metrics Accelerated
- July saw AI activity accelerate across private metrics despite public sell-offs and weak sentiment.
- Gavin observed GPU availability, rental pricing, DRAM spot, and token growth all accelerating in Silicon Valley anecdotal telemetry. (Time 0:02:27)
Contracted Compute Trades Below Rising Spot
- Many long-term GPU contracts are trading far below current spot, creating a repricing dynamic as contracts roll off.
- Gavin reports anecdotes of rental clusters paying 50–60% more than months earlier when contracts expire. (Time 0:04:49)
Rising Yields Make Debt-Funded Buildouts Risky
- Real yields and credit spreads widened in July, pushing CDS higher across the tech complex.
- Gavin warns that if the AI buildout needs material debt financing this spread widening would be a large negative risk. (Time 0:10:19)
Open Source Is Dark Matter For Public Markets
- Open source model adoption is ‘dark matter’ to public markets because it’s hard to measure yet materially increases token demand.
- GLM 5.2 and Kimi K3 accelerated open-source adoption, shifting inference mix toward cheaper tokens. (Time 0:17:41)
LLMs Have Become The Market’s Shared Narrator
- Many market participants route news through the same LLM tools like Claude, creating homogenized trading reactions.
- Gavin says Claude acts like a Walter Cronkite for the stock market, amplifying shared interpretations. (Time 0:21:21)
Continual Learning Won’t Remove Hardware Demand
- Continual learning and sample-efficiency breakthroughs could change training demand but likely won’t eliminate semiconductor need.
- Gavin notes labs claim progress, but even with efficiency, training’s share of semiconductor demand won’t go to zero. (Time 0:25:00)
Routers Drive Multi-Model Efficiency And More Tokens
- The future will be multi-model orchestration with routers sending requests to cheaper open-source models and frontier models as needed.
- Gavin explains routers can lower customer spend while increasing total token consumption because tokens cost similar compute. (Time 0:27:42)
Token Spend Reaches High Shares Of Payroll
- Tokens are becoming a meaningful share of compensation spend at AI-native companies, sometimes 20–30% or higher.
- Gavin cites examples: some firms at ~20%, Dylan Patel claims ~30%, and an extreme anecdote at 50%. (Time 0:31:50)
Breaking Memory LTAs Can Blow Up Your Franchise
- Memory suppliers and hyperscalers are signing LTAs that trade short-term upside for supply certainty, creating strong game theory deterrents to breaking deals.
- Breaking an LTA risks losing future allocations and potentially your franchise in tight markets. (Time 0:34:17)
Nvidia’s Credit Wrapper Expands Its Revenue Per Gigawatt
- Nvidia introduced a credit-wrapper plus revenue-share model that functions like vendor-aligned financing and gives it upside via equity and royalties.
- Gavin argues this raises revenue per gigawatt and creates a de facto cloud-like recurring business for Nvidia. (Time 0:37:45)
Regulation Is The Biggest Existential Risk
- Regulation is the single biggest risk to AI’s trajectory, driven by politics, misinformation, and local actions like data center moratoria.
- Gavin urges better industry storytelling to correct myths about water use, power costs, and local job benefits. (Time 0:51:47)
SpaceX Could Rapidly Shift Compute Supply Dynamics
- SpaceX’s compute push (Grok 4.5 and Cursor) could rapidly add material market capacity at low cost, but public estimates may understate or misprice that potential.
- Gavin notes claims like 8 GW in 18 months are ambitious yet would reshape market supply if true. (Time 0:59:09)