E2E Networks launches Blaze for pay-per-token AI inference

A new model for AI production

E2E Networks has introduced Blaze, a new inference service designed to help businesses move their artificial intelligence projects into actual production. Inference is the process where a trained AI model generates an output, such as text or images, based on a user's input. Many companies are currently struggling to move past the pilot phase because of the high costs and technical hurdles involved in scaling these systems.

The platform provides a pay-per-token system, which means customers only pay for the specific amount of data processed by the AI. This model is intended to help enterprises manage the rising costs associated with AI agents and data sovereignty requirements. By using a single API, developers can access a wide variety of open-weight models without having to build separate connections for each one.

As reported by Analytics India Magazine, the launch comes at a time when Indian AI companies are seeking more cost-efficient ways to run large-scale workloads. Blaze is specifically built to support production-grade tasks like coding agents, document AI, and voice applications. It aims to provide a smoother transition for companies that need to deploy reliable AI features quickly.

Infrastructure and model variety

Blaze is powered by significant hardware investments, including NVIDIA GPU infrastructure hosted within India. E2E Networks is expanding its AI factory capacity to approximately 14.6 MW to meet this demand. This includes a 1,024-GPU NVIDIA Blackwell cluster that has been operational since May, along with an additional 5.4 MW of Blackwell capacity featuring HGX B300 systems.

The service supports a diverse range of open-weight models, which are AI models whose internal parameters are available for public use. These include well-known models such as NVIDIA Nemotron, Sarvam, DeepSeek, Qwen, GLM, and Gemma. By offering these models through a single platform, E2E Networks allows users to choose the best tool for their specific needs while keeping the underlying hardware management in the hands of the provider.

Tarun Dua, Founder and Managing Director of E2E Networks, noted that there is significant demand for frontier-class open-weight models running on a cloud partner with NVIDIA reference design clusters. He stated that Blaze provides this on the E2E Cloud in India. This setup is designed to function as Sovereign Compute, which refers to computing resources that are managed and located within a specific country's borders to meet local regulations.

Technical optimizations for speed

To handle latency-sensitive applications like voice and chat, Blaze incorporates several technical optimizations to improve the speed of the first sentence. One key method used is key-value (KV) cache reuse. This technique helps reduce the overall cost of inference by remembering parts of previous computations so the system does not have to re-calculate them every time.

The platform integrates several software components from the NVIDIA DSX AI Factory Platform. Specifically, it uses NVIDIA Dynamo for distributed model serving, which handles KV cache-aware routing and separates the prefill and decode processing. Prefill is the stage where the model processes the initial input, while decode is the stage where it generates the output word by word.

Workload coordination is managed through NVIDIA Grove and the KAI Scheduler, which organize tasks across E2E’s Kubernetes infrastructure. Kubernetes is a popular open-source system used to automate the deployment, scaling, and management of containerized applications. These tools work together to ensure that the AI models can handle high volumes of traffic while maintaining consistent performance for the end user.

Security and enterprise features

For organizations in highly regulated industries, Blaze includes specific hardware-level isolation features. The platform utilizes NVIDIA BlueField DPUs, DOCA, and an Infra Controller to create secure boundaries between different customers. This level of isolation is particularly important for sectors like banking, healthcare, and the public sector, where data privacy is a primary concern.

The service is designed to integrate easily with all other E2E Cloud platform services. This allows enterprises to build complex workflows that combine AI inference with other cloud capabilities. By providing a managed environment, E2E Networks takes on the responsibility of hosting and scaling the models, allowing the customer to focus on building their applications.

Blaze is positioned alongside E2E’s existing AI platform, TIR. While Blaze focuses on production inference, TIR supports self-hosted inference, model pre-training, and post-training on NVIDIA Blackwell GPU clusters. This combination gives customers a full path from the initial training of a model to its final deployment in a production environment. Customers can also use TIR to serve their own custom-tuned models through managed infrastructure.