Korva rescues the ideas worth keeping from podcasts. Each one becomes a Card.
Without safety-related constraints, frontier labs could generate $100 million or more per megawatt and outbid other compute buyers.
An EUV lithography tool costing $400 million could resell for more than $1 billion because it is so scarce.
Frontier compute grows four- to fivefold yearly while equivalent capabilities need threefold less compute, implying roughly tenfold effective AI labor growth.
AI’s economics are bizarre: $10 of compute can become roughly $100 of model-provider revenue.
Elon’s compute business can sell capacity to major labs for $25–40 million per megawatt and recover capital expenditure within a year.
Anthropic’s inference revenue has reached $50 million per megawatt, against roughly $10–15 million in compute costs.
$6 billion in fab capital expenditure can produce one gigawatt of compute annually, potentially generating $100 billion in revenue.
The model layer went from negative gross margins to massive positive margins in about one year, with a path to $100 million per megawatt.
Even after assigning half the economics to costs and intermediaries, fab CapEx remains roughly 100 times below end revenue.
By late 2028, OpenAI and Anthropic could control most of the world’s usable floating-point compute.
Reaching 100 gigawatts of AI compute by 2028 requires labs to pay $25–50 million per megawatt and capture 70% of global compute.
By 2029, how much will AI borrowing raise interest rates, especially for major borrowers like Amazon?
AI labs face a stark choice: monetize compute through inference now or spend it on research with larger future returns.
Bring your next episode to Korva.
Turn podcast episodes into Cards you can save, explore, and find again.