The lesson
Turn an infrastructure opinion into a demand-shape argument, so the reader can see why the old buying model breaks.
The essay opens with Gen AI products creating constrained GPU demand, then asks whether startups should keep buying long-term GPU reservations. It reframes the market around inference, under the heading "How do you make money from AI models?", then argues through "Inference workloads are volatile" and "It gets more complicated – the 24h cycle" with simple request-pattern charts. After widening the claim with "Training is also volatile!" and "There’s many other volatile workloads!", it moves to "How can GPU providers provide on-demand GPUs?" and answers with pooling ideas like "Demand pooling" and "Supply pooling."
Don't copy Modal's GPU market thesis or its specific workload examples. Keep the structure of showing the demand pattern first, then arguing for the buying model that pattern requires.
Tell us what you want to make. We’ll talk through how to put these ideas to work for your brand.