Gavin Baker - Watts and Wafers - [Invest Like the Best, EP.473]

Frontier AI labs like Anthropic are limited by compute availability, not demand. Their true earning power, or "Unconstrained Revenue," is likely 2-3x their reported ARR, a critical metric for valuation when considering their growth if supply constraints were removed.

Anthropic's superior capital efficiency, evidenced by its significantly lower cash burn to achieve a revenue scale comparable to OpenAI, indicates a structurally lower cost per token. This highlights a key competitive differentiator in the capital-intensive AI model race.

In the current AI landscape, economic value is overwhelmingly created by companies possessing the highest ratio of utilized GPUs per employee. This trend suggests that access to and efficient use of computational power, rather than human capital alone, is the primary driver of value, at least at the infrastructure layer.

The concept of "data centers in space" is often misunderstood. It's not about launching massive buildings, but rather individual, 3,000-pound server racks connected via lasers into a virtual data center. This reframing makes the ambitious idea far more practical and achievable with current technology.

Separating inference into "prefill" (memory-bound) and "decode" (bandwidth-bound) tasks is a game-changer for hardware longevity. It allows older GPUs to be used for prefill tasks indefinitely, extending their useful economic life from 3-4 years to 10-15 years, a boon for data centers and their financiers.

To avoid being crushed by incumbents, AI startups must operate on ideas that are both non-obvious ("different") and difficult to execute ("hard"). If a startup's core idea becomes obvious to the world before it achieves significant scale, larger companies with more resources will inevitably co-opt the market.

Elon Musk's ability to raise vast sums of capital stems from a 20-year commitment to making investors money. He achieves this by consistently setting fundraising valuations at a fair or even understated level, treating investor returns as a "sacred covenant" that builds long-term trust.

Beyond technology, Elon Musk's strategy for the TeraFab chip plant involves a deep cultural and talent play. He plans to build a "Taiwan town" and similar communities in Texas to recruit the world's best semiconductor engineers by recreating their home environments, a unique advantage over incumbents.

Unlike the dot-com era's debt-fueled fiber overbuild, the current AI boom is constrained by wafer supply, controlled primarily by TSMC. Their disciplined capacity expansion, despite immense demand, prevents a speculative oversupply of GPUs, effectively acting as the single most important governor against an AI bubble.

The "bitter lesson" states that more compute always beats better algorithms. While this has held true, it may be temporarily violated by the arrival of ASI. An ASI's first goal would be to become smarter and more efficient, potentially creating algorithmic breakthroughs that temporarily outpace the benefits of raw compute.

The move from flat-rate subscriptions to pay-per-use models for frontier AI is a pivotal growth catalyst. Similar to how early cellular plans with overage fees drove massive revenue, this shift unlocks uncapped spending and is predicted to push labs like OpenAI and Anthropic to over $200 billion in ARR.