Perplexity AI has launched TransferEngine, an open-source tool designed to enable GPU-to-GPU communication across different cloud providers, addressing the costly issues of vendor lock-in and the need for cutting-edge hardware. This tool allows enterprises to run massive trillion-parameter models, such as DeepSeek V3 and Kimi K2, on older H100 and H200 GPUs without sacrificing performance. Previous solutions often forced companies to choose between cloud providers due to incompatible networking protocols, leading to slower performance. TransferEngine overcomes this by acting as a universal translator for data transfer, achieving 400 gigabits per second throughput on both Nvidia and AWS systems. The technology uses Remote Direct Memory Access (RDMA), allowing data to move directly between GPUs, bypassing the main processor. This efficiency is especially needed for large models that can't fit on single systems, and the tool has already been implemented in Perplexity's AI search engine. It's capable of handling high-speed data transfer for disaggregated inference and supports rapid weight updates for models in under 1.3 seconds. TransferEngine has also matched the performance of proprietary solutions while being compatible with AWS's Elastic Fabric Adapter, a significant breakthrough in cloud model deployment. Perplexity's choice to open-source this technology sets it apart from competitors like OpenAI, which keep their implementations private. By releasing the complete library, including code and benchmarking tools, Perplexity aims to establish a standard in the industry and invites community contributions to enhance the tool further. The company is committed to optimizing TransferEngine for AWS, reflecting a strategic move to benefit from collective advancements in cloud technology.