Should your AI app rent its models or build its own?
Canva trained its own models and cut costs by about 90%. Harvey says rent them until the bill forces you to build. Our read: rent for hard tasks, own routine work at volume.
· 5 min read
Signalful Analysis
From interviews with Gabe Pereyra of Harvey, Misha Laskin of Reflection AI, Gabe Stengel of Rogo, Cliff Obrecht of Canva, Alex Mashrabov of Higgsfield, Mike Cannon-Brookes of Atlassian, Anish Acharya of Andreessen Horowitz and Daniel Dines of UiPath
Canva cut its AI costs by about 90% by training its own models. Harvey spends hundreds of millions of dollars a year running AI models and tells application companies to rent closed ones until the bill forces them to train their own. Opposite moves, one question: should an AI product run on rented models or its own?
Builders give three answers: rent the best model and ride each release, own the model for margin, or match the model to the task. A task worth thousands of dollars justifies the most expensive model on offer. Routine work at volume does not.
