SF Compute: Commoditizing Compute to solve the GPU Bubble forever

Episode Summary

Executive Summary: The episode centers on Evan Conrad’s thesis that GPU clouds are fundamentally different from CPU clouds: GPUs are capital-intensive, price-sensitive, and best treated like real estate/banking rather than software. He explains why CoreWeave’s long-term-contract model works, why SF Compute evolved into a spot market/brokerage, and how standardization, auditing, and future financial products could reduce risk for buyers and sellers.

Main Topics: Why GPU clouds are not CPU clouds (Priority: 5/5): Evan argues that GPU economics are driven by massive capital costs, depreciation, and customer price sensitivity, making the traditional high-margin software-cloud model ineffective. CoreWeave’s winning structure (Priority: 5/5): CoreWeave succeeded by signing long-term contracts, focusing on low-risk counterparties, and using cheap capital—more like a bank/real-estate business than a software company. SF Compute’s origin and evolution (Priority: 5/5): SF Compute began as an AI lab trying to buy compute, then became a broker to survive, and ultimately turned into a liquid GPU spot market with bids, asks, and short-duration reservations. Market structure, pricing, and utilization (Priority: 4/5): The discussion explains how spot-like pricing, contract length, and market volatility shape GPU utilization and why short-term, flexible purchasing can be economically efficient. Auditing, standardization, and reliability (Priority: 4/5): SF Compute differentiates itself by burn-in testing, active/passive checks, BMC access, and standardized contracts to make clusters more reliable and tradable. Futures and financialization of compute (Priority: 4/5): Evan frames futures as a risk-reduction tool that could stabilize the compute market, lower cost of capital, and eventually create a more mature financial layer for GPUs. Brand, culture, and hiring (Priority: 2/5): The company’s calm, anti-hype branding reflects its origin story and product philosophy; the episode closes with hiring pitches for systems and financial systems engineering roles.

Key Arguments: GPU customers are far more price-sensitive than CPU-cloud customers because every incremental GPU can directly translate into revenue, so software margins are harder to extract. CoreWeave’s model works because it sells long-term contracts to low-risk customers, which lowers financing costs and aligns with GPU depreciation risk. Hyperscalers are unlikely to make strong margins reselling NVIDIA GPUs because they either compete with their own customers or forgo better uses of capital. The best GPU business is to separate hardware ownership from software/services; mixing them tends to destroy margins and create risk. SF Compute’s market design lets buyers purchase short bursts, month-long capacity, or flexible reservations while sellers monetize idle inventory. A real spot market plus auditing and standardization can create a reliable index price, which could later support cash-settled futures. Futures are presented as a stabilizing mechanism, not speculation: they reduce risk for data centers, customers, and the broader AI funding ecosystem. Peer-to-peer distributed compute markets are viewed skeptically because speed-of-light and cluster-co-location constraints make fully distributed GPU compute inefficient. The company’s early survival pressure shaped its philosophy: low hype, low expectations, and operational rigor over flashy software promises.

Data Points: CoreWeave revenue concentration: 77% - Evan says Microsoft and OpenAI together account for 77% of CoreWeave revenue. Example hardware scale difference: CPU cloud: million dollars of hardware; GPU cloud: billion dollars of hardware - Used to illustrate why GPU clouds are much more capital intensive. Example pricing spread: $5/hour charged vs. $1.50/hour underlying GPU cost - Illustrative example of how high prices can be undermined by customer price sensitivity and competition. Training cost estimate: ~$500 million - Referenced as rough training cost for GPT-4 and o1 in the discussion of when custom chips become viable. Large-run chip-design threshold: $5 billion to $50 billion runs - At these scales, Evan cites Martin Casado’s point that custom ASICs may make sense. OpenRouter open-source inference footprint: ~10 H100 nodes - Used to argue that open-source inference demand is much smaller than closed-model demand. SF Compute early monthly obligation: ~$500,000 per month - Evan describes the company’s early year-long contract as a survival risk because monthly obligations matched their cash on hand. SF Compute cash on hand at the time: ~$500,000 - Matched the monthly obligation, creating a near-bankruptcy situation if the cluster wasn’t sold out. Typical contract duration in traditional GPU clouds: 1 year or longer - What vendors told SF Compute they needed to sign when trying to buy compute month-to-month. Potential burst purchase duration on SF Compute: 1 hour - Evan says the market can support thousands of H100s for an hour because the platform is built around hourly reservations.

Pivotal Quotes: "The best way to make money in GPUs was to do basically exactly what CoreWeave did." — Evan Conrad: Summarizing his thesis that long-term contracts and low-risk financing are the winning GPU-cloud strategy. "It's a bank. It's a real estate company." — Evan Conrad: His characterization of CoreWeave’s true business model beneath the cloud-provider label. "Futures are the way to chill out the entire industry." — Evan Conrad: His argument that financial instruments can reduce risk and stabilize compute markets rather than encourage speculation.

Implications: GPU infrastructure is evolving into a financialized, market-driven industry where risk management matters as much as raw compute. Listeners should expect more spot pricing, standardized contracts, auditing, and eventually futures-like products shaping AI compute access and cost.