Menlo Ventures partner Deedy says inference providers are quickly reaching $1 million in monthly revenue

Menlo Ventures partner Deedy (@deedydas) posted on X on October 25, 2026, saying that a large number of inference providers have emerged, and some can reach $1 million in monthly revenue in a very short time. The economic model he described was: even if B300s are rented on demand at $8 per GPU hour, deploying open-source models such as GLM 5.3, Kimi K3, and DeepSeek 4.1 Flash/Pro, at reasonable throughput (100 tok/s), each GPU hour can generate more than $10-20 in token revenue; DeepSeek's DeepInfer shows that during non-peak hours it can reach $38 per GPU hour.

Deedy said that gross margins depend on utilization and the efficiency of the tech stack. Gateways such as OpenRouter and Vercel can aggregate demand. He said some providers discount tokens in exchange for long-term commitments, and he has heard of cases in which Chinese model tokens are sold back to Chinese labs. Some Chinese open-source model labs require revenue sharing from providers that reach a certain scale, especially for pre-release access, which raises the profitability threshold.

He observed that the barrier to entry has fallen, with small teams deploying open-source models on rented clusters through “vibe code.” Nearly every week, dozens of such companies emerge, with revenue growing rapidly but gross margins remaining low, and they raise financing on that basis. Because services are becoming commoditized, this pushes up compute prices and depresses token prices. In theory, neoclouds with the lowest compute costs (such as Crusoe, CoreWeave, Nebius, and GMI) and their inference businesses may have an advantage in a price war, followed by larger players (Modal, Fireworks, Baseten); but in practice the market is growing rapidly, and new entrants (such as inference.net, DeepInfra, Wafer, Relace, and Morph) are also growing quickly.

Deedy listed risks including whether open-source models will continue to remain competitive, whether their creators will further seek rent, whether model form factors will change compute demand, how different chips will affect economics, and who will be able to secure compute resources when supplies are tight. He attached a schematic showing how gross margins change with utilization under different model/chip combinations, and concluded that many inference providers are currently growing rapidly.

Source:

1. Title: Post by Deedy on the economics of inference providers Publisher: Deedy (@deedydas), partner at Menlo Ventures Date of publication: 2026-10-25 URL: https://x.com/deedydas（相关帖） Nature of source: personal original post Verification status: actually read in this round, verification date 2026-10-26 Supporting content: the above economic estimates, company names, observations, and risk list all come from the post; the figures are estimates or quotations provided by Deedy and were not independently audited