E2E Networks launches Blaze AI inference service with Nvidia

Moving from pilots to production

E2E Networks has officially launched Blaze, a new inference service designed for open-weight AI models. Inference is the process where an AI model actually generates a response or performs a task based on a user's input. This service is specifically built to help developers and enterprises move their projects beyond the testing phase and into real-world use.

The service operates on a pay-per-token basis, which means users are billed only for the specific units of text or data the model processes. This model is becoming increasingly important as Indian AI companies scale their operations. As these companies deploy AI agents, the amount of data processed can grow rapidly, making cost-efficiency a primary concern for business viability.

Tarun Dua, the founder and managing director of E2E Networks, noted that Blaze is intended for running production inference. The goal is to provide a reliable environment where companies can manage the high costs and technical demands of large-scale AI. By offering a single endpoint for multiple models, E2E Networks aims to simplify how developers access these powerful tools.

Nvidia-powered infrastructure

Blaze AI is built on a foundation of Nvidia GPU infrastructure and several open-source components from the Nvidia DSX OS. The DSX OS is a modular software layer that sits within the Nvidia DSX AI Factory Platform. By using these components, E2E Networks can provide a high-performance environment that is optimized for the specific needs of AI workloads.

Several specific technologies from Nvidia are integrated into the Blaze service to manage how data moves and how models are served. For example, Nvidia Dynamo handles distributed model serving, which allows the system to manage complex tasks like disaggregated prefill and decode. This helps the system handle different parts of the AI generation process more efficiently across multiple GPUs.

The platform also utilizes Nvidia Grove and the KAI Scheduler to coordinate how workloads are distributed across E2E's Kubernetes fleet. This coordination ensures that the hardware is used effectively to meet user demands. Additionally, Nvidia NIXL is used to facilitate fast data transfer between different GPUs, which is essential for maintaining the speed of the AI's responses.

Latency and cost optimizations

For many AI applications, the speed at which a model responds is just as important as the quality of the answer. Blaze is specifically designed for inference-heavy products where low latency is a requirement. This includes applications like real-time voice assistants, coding agents, and document AI. For voice and chat, the service focuses on providing a fast time to the first sentence to ensure a smooth user experience.

The service also focuses on reducing the cost of running open-weight models through high kv cache reuse. A kv cache is a technical method used to store previous parts of a conversation so the model does not have to re-calculate them every time. By reusing this cache, Blaze can significantly lower the cost per token for developers who are running long or complex interactions.

The platform is also well-suited for diverse media generation tasks. This includes image and video generation as well as Optical Character Recognition (OCR) and general language applications. By optimizing for these specific types of workloads, E2E Networks aims to provide a versatile tool for a wide range of industrial and consumer AI products.

Sovereign compute and security

A major component of the Blaze launch is the focus on sovereign compute, which refers to running AI on infrastructure located within a specific country's borders. This is a high priority for many Indian companies that need to comply with local regulations. E2E Networks is positioning its India-hosted AI factory as a primary location for these types of sensitive production workloads.

Security is a core feature of the platform, particularly for industries with strict data rules like banking, healthcare, and the public sector. Blaze uses Nvidia BlueField DPUs, Nvidia DOCA, and the Nvidia Infra Controller to enforce hardware-level isolation between different customers. This means that even though multiple companies might use the same infrastructure, their data remains strictly separated at the hardware level.

Vishal Dhupar, the Managing Director for South Asia at Nvidia, stated that access to AI computing and inference services is essential to advancing India’s AI ecosystem. By providing a secure and local way to run frontier-class models, E2E Networks and Nvidia are attempting to create a foundation for domestic AI growth. This infrastructure allows companies to adopt advanced AI without compromising on data privacy or regulatory requirements.

Expanding capacity in India

To support the growing demand for Blaze AI, E2E Networks is significantly expanding its physical infrastructure in India. The company's AI factory capacity is growing to approximately 14.6 MW. This expansion is designed to accommodate the increasing needs of companies moving toward production-scale AI deployments.

A significant part of this expansion includes a 1,024-GPU Nvidia Blackwell GPU cluster that has been live since May 2026. The Blackwell architecture represents the latest generation of high-performance computing from Nvidia. This cluster provides the raw power necessary to run the large, complex models that modern enterprises are beginning to request.

In addition to the Blackwell cluster, E2E Networks is adding another 5.4 MW of capacity. This extra space will include HGX B300 systems and other Blackwell GPUs to ensure there is enough room for growth. By scaling the physical hardware alongside the Blaze software service, E2E Networks is preparing for a future where high-performance AI inference becomes a standard requirement for Indian businesses.