Blaze brings pay-per-token Qwen inference to India

Blaze packages Qwen for consumption-based inference

E2E Networks has launched Blaze, a managed inference service that gives developers and enterprises access to Qwen and other open-weight models through a single OpenAI-compatible API. Customers pay only for the tokens used, while E2E handles model hosting and scaling. The service runs on NVIDIA GPU infrastructure hosted in India and is aimed at production workloads where inference cost, latency and data residency matter.

Qwen is one of several model families available through Blaze, alongside NVIDIA Nemotron, Sarvam, DeepSeek, GLM and Gemma. The service supports use cases including coding agents, voice and chat applications, document processing, optical character recognition, image and video generation, and other language applications. Its positioning is therefore broader than a Qwen distribution agreement: it is a managed access layer spanning multiple open-model ecosystems.

Blackwell expansion supports an inference strategy

E2E said its AI factory capacity is expanding to approximately 14.6 MW. That includes a 1,024-GPU NVIDIA Blackwell cluster operating since May and an additional 5.4 MW of planned Blackwell capacity, including HGX B300 systems. The expansion would move the company from approximately 9.2 MW to 14.6 MW, an increase of nearly 59% based on the disclosed capacities.

The additional infrastructure is intended to serve work spanning foundation-model training, post-training, production inference and agentic AI. However, the reviewed reporting does not disclose a firm commissioning timetable, the required investment or the number of GPUs planned for the added capacity. It also does not establish that the expansion will translate into a specific procurement order.

Hosting and serving are the service differentiators

Blaze incorporates open-source components from NVIDIA DSX OS, while NVIDIA Dynamo provides distributed model serving, including cache-aware routing and separated prefill and decode processing. NVIDIA Grove and KAI Scheduler coordinate inference workloads across E2E's Kubernetes infrastructure. BlueField DPUs, DOCA and Infra Controller are intended to provide hardware-level isolation between customers.

The architecture also uses key-value cache reuse and custom NVIDIA CUDA kernels to reduce repeated computation and support long-context and coding-agent workloads. For voice and chat applications, E2E says the service is designed to reduce time to the first generated sentence. These features address the cost and latency trade-offs that can arise as Indian developers move beyond pilot deployments, although the reviewed reports provide no comparative performance benchmarks.

API inference complements private deployment

Blaze extends E2E's infrastructure offering into managed API-based inference while leaving private and customized deployment to its TIR platform. TIR supports model development, training, fine-tuning and self-hosted inference on managed Kubernetes infrastructure built around NVIDIA Blackwell clusters. Customers can therefore use Blaze for consumption-priced access to hosted open models or TIR when they need greater control over their own or fine-tuned models.

E2E Networks positions the two services as a path from model customization to production inference within India. Hosting data locally may help organizations address data-residency considerations, but local placement alone does not establish regulatory compliance; access controls, data handling and security arrangements remain workload-specific. The strongest verified industry signal is therefore a broader managed-inference offering that incorporates Qwen, rather than disclosed enterprise demand or a new Qwen-specific procurement lead.

Industry and product implications