Together AI targets inference layer with new model partnerships

Together AI is positioning itself as a specialized cloud provider focused on the inference layer, a strategy driven by the belief that this is the only point where AI companies deliver tangible value and generate revenue. Dan Fu, VP of Kernels at the company, stated that inference is the critical consumption endpoint for AI, noting that it is the stage where users actually interact with the technology. To support this focus, Together AI has established partnerships with model providers Moonshot, Minimax, and Cursor to serve their specific inference needs. The company is also broadening its training capabilities to assist general businesses with post-training and fine-tuning models, aiming to build infrastructure analogous to traditional cloud services but optimized for AI workloads. Fu emphasized that the company’s strategy involves scaling compute supply to produce more tokens, supported by internal tools like custom Perfetto and trace viewers that allow engineers to visualize hardware-level performance.

Beyond infrastructure, Fu highlighted how AI coding models have accelerated the development of GPU kernels and inference stack components over the past year. However, he argued that the effectiveness of these tools is directly proportional to the engineer’s domain expertise. Without a deep understanding of hardware and software stack details, AI-generated code may be incorrect or poorly designed due to edge cases. Fu described AI as an amplifier for existing skills rather than a replacement, noting that the team has invested in creating Markdown-based agent read-me files to help AI agents pick up work effectively. He observed that the role of the engineer is shifting toward orchestrating agents and managing the cognitive load of high-velocity code production, where greater domain expertise allows for faster utilization of AI tooling.