SF Compute: Commoditizing Compute
Latent Space: The AI Engineer Podcast
2025/04/11
SF Compute: Commoditizing Compute
SF Compute: Commoditizing Compute
Latent Space: The AI Engineer Podcast
2025/04/11
This podcast delves into the intricacies of the GPU market, focusing on CoreWeave's unique business model and its resemblance to a real estate venture. Evan Conrad from SF Compute discusses how his company transitioned from an AI lab to a cloud provider by selling GPU clusters. The conversation also examines the challenges and opportunities within the GPU industry, including pricing dynamics, customer needs, and future demand forecasts.
Evan Conrad explains how SF Compute navigated the complexities of the GPU market by starting as an AI lab and pivoting to become a cloud provider. The discussion highlights CoreWeave’s financial strategy, which resembles a real estate business due to long-term contracts with stable customers. Key topics include interest rate risks, utilization rates, and supply chain issues affecting GPU availability. SF Compute addresses these challenges by offering short-term GPU purchasing options and ensuring high cluster reliability. The podcast further explores the inefficiency of peer-to-peer GPU networks, emphasizing the importance of interconnected clusters. Customer stories illustrate the value of providing significant burst capacity to grad students and startups. Additionally, the speakers touch on VC-provided GPU clusters and credit risk arbitrage, along with innovative pricing models like replacing preemptible pricing with short reservations. Future plans involve financializing GPUs through futures contracts to stabilize the market, contrasting with current AI industry hype. Finally, branding and hiring strategies at SF Compute are discussed, highlighting their focus on low-key marketing and recruiting skilled systems engineers.
05:19
05:19
CoreWeave doesn't resemble a cloud provider or software company financially.
08:02
08:02
Chart explains GPU market profits through interest rates and cost depreciation.
11:38
11:38
CoreWeave's business model is more viable compared to selling high-priced short-term contracts.
16:34
16:34
Treating GPU clouds as real estate businesses can be profitable.
25:24
25:24
Short-term GPU burst capacity allows flexibility for users with varying needs.
28:17
28:17
Price drops in H100 GPUs reflect complex market dynamics.
30:00
30:00
Evan predicts a GPU shortage by winter due to future chip roll-out.
34:01
34:01
Peer-to-peer GPU markets are less efficient than interconnected clusters due to physical constraints.
36:50
36:50
Offering big burst capacity is particularly rewarding for difficult customers
38:23
38:23
VCs provided GPU clusters to startups due to credit risk challenges.
46:35
46:35
The financialization of GPUs could lead to their listing on exchanges.
50:17
50:17
Automated refunds switch customers to other resources during hardware issues.
57:53
57:53
Derivatives on mortgages are 12 times larger than the mortgages themselves.
58:00
58:00
Creating futures contracts for compute reduces risk.
1:06:23
1:06:23
Every year, someone tries to crack email but gives up.
1:08:43
1:08:43
Intercom is better positioned to use AI for email in support use-cases.
1:09:07
1:09:07
SF Compute seeks Linux-savvy engineers, not just Rust experts.
1:09:53
1:09:53
The role aims to prevent money loss and improve pricing for vendors and buyers
1:10:57
1:10:57
Tiger Beetle ensures systems do not lose money
