Soaring AI Costs Push Startups Toward Cheaper Open-Weight Models
Startups are increasingly ditching expensive proprietary AI from OpenAI and Anthropic in favor of cheaper open-weight models. The trend accelerated after legal tech firm Harvey saw its gross margins collapse from roughly 50% to negative 50% following a March update that spiked customer usage of its AI agents.
In response, Harvey built its own model using China-based Moonshot AI's Kimi K3, which reportedly performs close to Anthropic's best offerings at a fraction of the cost. The move, along with other adjustments, has restored Harvey's margins to positive territory. Healthtech startup Abridge is following a similar path, building a custom foundation model for clinical settings trained on Nvidia's open models.
Investors including Sequoia Capital and General Catalyst are backing the shift, which cuts one of startups' biggest costs while giving them more control over their technology. The trend threatens to erode revenue at OpenAI and Anthropic as both prepare for anticipated IPOs, and it also serves as a hedge for startups worried that frontier models could face restrictions or slowed development.