MiniMax bets on model ownership

MiniMax is betting that owning the model layer is the only way to keep AI output cheap enough and good enough to win multimodal workloads. Instead of paying another lab every time a customer generates text, speech, or video, it trains its own models and sells them directly. That can turn gross margin into a product advantage, but it also forces heavy spending on GPUs, training runs, and research talent before revenue fully catches up.1 2 3 4

• The clearest contrast is with distribution layers like Veed and Segmind. They aggregate outside models and monetize workflow or orchestration, which lowers R&D burden but leaves core model cost and quality in someone else’s hands. MiniMax keeps that control in house, which matters more when customers buy multiple modalities together.1 5 6
• There is precedent for the margin upside. Canva moved from plugging in third party AI tools toward owning more of the stack through Leonardo AI and its own design model, so it could bake generation directly into editing workflows instead of paying a vendor at arm’s length on every request.7 8
• The cost side is real. In frontier AI, companies that train their own models need enough capital and infrastructure to survive repeated model cycles. MiniMax had raised an estimated $1.15B and reached an estimated $800M revenue run rate by August 2026, which helps explain why it can afford a full stack strategy while many smaller apps cannot.2 3

Going forward, multimodal AI should split more sharply between model owners and model resellers. If MiniMax keeps model quality competitive, its direct control over training, serving, and pricing can compound into lower task costs and broader enterprise expansion across text, audio, and video from the same account.1 2 3 4