gavin baker

SK Hynix benefits from the same LTA dynamic as Micron. Hyperscalers breaking LTAs would risk future HBM allocations, making memory revenue more durable than market assumes.

Ramp's spend data reveals AI-native companies allocating 20-25% (up to 50%) of total compensation budgets to token inference, correlating with faster growth. This labor-substitution dynamic supports the long-term compute demand thesis.

Base10 is among the inference clouds monetizing open-source inference, part of the under-the-radar demand acceleration that public markets miss.

OpenAI has accelerated its model releases and growth, countering the narrative that open source competition hurts frontier labs. The combination of OpenAI acceleration, Anthropic strength, and open source explosion is driving net token demand acceleration.

CoreWeave has demonstrated ability to energize gigawatts rapidly. Its contracted compute base rolling off into higher spot prices creates a repricing tailwind that accelerates revenue and cash flow, reducing credit dependence for the broader buildout.

Nvidia trades at its lowest forward PE in a decade because the market believes it is significantly over-earning, but contracted compute is rolling off at a massive discount to spot prices, driving accelerating operating cash flows at hyperscalers that can fund the buildout without excessive debt. Nvidia's credit-wrapper model, equity stakes in labs, and financing/land/power advantages create a self-reinforcing moat.

Stripe's internal data (alongside Cognition and Ramp) shows companies with highest AI token spend grow meaningfully faster, reinforcing the view that AI drives productivity not just cost-cutting.

Harvey (pre-acquisition) and Cursor leaned heavily into customizing open-source models via inference clouds like Fireworks, achieving better performance at half the cost of pure frontier models. This router architecture makes AI natives more defensible and drives more total compute.

Benchmark's investment in StarCloud (orbital compute, partnering with SpaceX for Starlink laser links) is a sanity check that orbital compute is credible. If Benchmark, SpaceX engineers, and Elon are all 'crazy' on this, the probability of it being real increases.

China's DUV machine is a phase transition (prop plane to jet turbine) but 25 years behind EUV. Market overreacted; any impact on ASML orders is 5+ years out. Learning-by-doing cycles cannot be teleported. Decoupling is self-reinforcing but gradual.