AI Summary
→ WHAT IT COVERS Evan Conrad, founder of SF Compute, traces how an accidental GPU sublease became a compute infrastructure business. He explains GPU cloud economics, the "offtake engine" model, where the real AI bubble sits, and why driving prices down — not capturing margins — is the core strategy. → KEY INSIGHTS - **Offtake contracts:** Before any large GPU cluster gets financed, a long-term offtake agreement must exist — a signed customer contract that the financier lends against. Without offtake, speculative cluster builders get wiped out, as happened repeatedly during the H100 cycle when idle capacity drove prices to $0.40/hour on the spot market. - **Where the AI bubble actually sits:** The risk is not in GPU clouds themselves, but in the equity of AI startups that raised large rounds, paid upfront for multi-year GPU contracts, and have no exit if their product fails. Those companies cannot sell, recap easily, or recover sunk compute costs — making their equity near-worthless on failure. - **GPU vs. CPU margin reality:** AI companies cannot achieve SaaS-style margins because every inference or generation consumes compute proportional to output volume. Unlike software sold once and scaled infinitely, AI workloads scale costs linearly with revenue, forcing GPU buyers to treat compute as a commodity and prioritize price above all managed-service features. - **The Marriott model for compute:** SF Compute designs cluster bills of materials, builds infrastructure using external capital from funds seeking AI exposure, then operates those clusters as a property manager would — splitting revenue with capital partners rather than owning assets outright. This lowers SF Compute's risk while giving financial partners direct compute exposure without intermediary cloud markups. - **Liquidity as bubble prevention:** AI companies locked into long-term GPU contracts currently have zero recourse if demand shifts — the contract simply destroys the company. SF Compute's order book creates a secondary market where companies can sell back unused contracted capacity, recovering partial value rather than facing binary outcomes of full deployment or total loss. → NOTABLE MOMENT Conrad describes how SF Compute discovered that no AI lab customer would ever pay a software markup on GPU hours — not because the product lacked value, but because the entire raised round was already allocated to compute costs, leaving literally no budget for additional services. 💼 SPONSORS None detected 🏷️ GPU Infrastructure, AI Compute Economics, Venture Investing, Cloud Markets, AI Bubble Risk