Mitesh Agrawal

Mitesh Agrawal is CEO of Positron AI, an AI inference hardware company. Formerly Lambda’s chief operating officer, he moved from GPU cloud infrastructure to leading a company offering alternative transformer-inference accelerators. His public arguments emphasize energy efficiency, performance per dollar, and giving customers more control over their infrastructure.[2][5][3]

On January 27, 2025, Agrawal announced his departure from Lambda, saying the preceding week had been his last there and that he was joining Positron. He described the opportunity as bringing an alternative accelerator to the transformer-inference market, calling the team “a big part of why I joined.” Positron announced his appointment as CEO on February 3, succeeding founder and former CEO Thomas Sohmers.[3][5]

In his appointment statements, Agrawal interpreted DeepSeek’s results as evidence that inference-time compute was becoming a scaling factor for AI. He argued that efficient inference would matter both for serving users and for building and refining subsequent models: “This is where performance per dollar and performance per watt become absolutely critical.” He also advocated freedom from vendor lock-in and hardware that puts customers in control of their infrastructure.[5]

Agrawal also appeared alongside Thomas Sohmers in a Latent Space interview about Positron’s accelerators.[1]

Before joining Positron, Agrawal worked on Lambda’s cloud business. In a Crescent Cove interview, he recalled: “We launched cloud services, and I led that effort.” His 2024 publications at Lambda explain the infrastructure access and workflow problems he addressed publicly during that period.[4][6][7]

Agrawal authored Lambda’s June 3, 2024 announcement of 1-Click Clusters. He argued that teams running experiments for a few weeks needed flexible cluster access rather than long contracts that left GPUs idle between runs. The offering combined 16–512 NVIDIA H100 GPUs, 400 Gb/s InfiniBand networking, and a one-week reservation minimum. Preinstalled ML software and shared storage were intended to let engineers concentrate on ML workflows instead of infrastructure management. The accompanying Open-Sora fine-tuning example used a four-node, 32-GPU cluster; that work belonged to Lambda’s ML Research team, with Chuan Li explicitly credited.[6]

In his August 15, 2024 Hermes 3 announcement, Agrawal connected training infrastructure with model access. Nous Research had fine-tuned Meta’s Llama 3.1 405B on an eight-node Lambda cluster, and Lambda offered limited-time free access through an OpenAI-compatible Chat Completions API and a chat interface. The model used Neural Magic’s FP8 method for quantization.[7]