Together AI receives US$305 million in financing to promote the development of GPU demand and in-depth inference models

In the AI ​​industry, Together AI recently announced a $305 million Series B round, which has attracted widespread attention. The company's rise is closely related to its newly launched deep-reaching model, DeepSeek-R1. Contrary to initial concerns, industry experts believe that advances in deep reasoning have not reduced the demand for infrastructure, but are constantly increasing this demand.

Since its inception in 2023, Together AI aims to simplify enterprise use of open source large language models (LLMs). Over time, the company gradually expanded its platform to provide a solution called "Together Platform" that supports deployment of AI in virtual private clouds and on-premises environments. In 2025, Together AI launched its inference cluster and autonomous intelligence (Agentic AI) capabilities, further enhancing the capabilities of its platform.

According to Vipul Prakash, CEO of Together AI, the parameters of DeepSeek-R1 are as high as 671 billion, which makes the cost of its operation of inference not to be underestimated. To meet the needs of more and more users, Together AI has launched the "Inference Cluster" service, providing customers with dedicated computing power from 128 to 2,000 chips to ensure the best performance of the model. In addition, DeepSeek-R1's request processing time is usually longer, with an average of two to three minutes, which also leads to an increase in infrastructure demand.

In terms of the application of inference models, Together AI has seen some specific usage scenarios, such as coding agents, reducing the illusion of the model, and achieving self-improvement of the model through reinforcement learning. These applications not only improve work efficiency, but also improve the accuracy of model output.

In addition, Together AI has acquired CodeSandbox to enhance its capabilities in autonomous intelligent workflows. This acquisition allows it to execute code quickly in the cloud, reducing latency and improving the performance of proxy workflows.

Faced with fierce market competition, Together AI's infrastructure platform is constantly being optimized, and the deployment of its new generation of Nvidia Blackwell chips will provide higher performance and lower latency for model training and inference. Prakash pointed out that compared with other platforms such as Azure, Together AI's inference speed has significantly improved, greatly meeting customers' needs for high-performance AI infrastructure.