E2E Networks Launches Blaze Pay-Per-Token AI Inference Service on 14.6 MW Sovereign AI Factory

E2E Networks Unveils Pay-Per-Token AI Inference Platform

E2E Networks has launched Blaze, a pay-per-token AI inference service designed to provide developers and enterprises with streamlined access to multiple open AI models through a single OpenAI-compatible API1

1

2

2

. The platform runs on NVIDIA GPU infrastructure hosted in India and targets production AI workloads requiring both cost efficiency and low latency. Customers pay only for the tokens they consume while E2E Networks manages model hosting and scaling infrastructure.

The launch addresses a critical inflection point as Indian AI companies transition from pilot projects to production deployments. According to Tarun Dua, Founder and Managing Director of E2E Networks, "There is customer demand for frontier-class open weight models running on an NVIDIA Cloud Partner with NVIDIA reference design clusters as Sovereign Compute. Blaze provides this on E2E Cloud in India"1

1

. This shift has intensified pressure on enterprises to optimize token efficiency as AI agents drive up consumption and AI-assisted development costs mount.

Sovereign AI Factory Expansion Reaches 14.6 MW Capacity

E2E Networks is expanding its Sovereign AI Factory capacity to approximately 14.6 MW, positioning India as a viable alternative for data sovereignty requirements1

1

. The infrastructure includes a 1,024-GPU NVIDIA Blackwell cluster that became operational in May, alongside an additional 5.4 MW of Blackwell capacity featuring HGX B300 systems2

2

. This substantial infrastructure investment reflects growing demand for compute resources that remain within Indian borders, particularly from banking, healthcare and public sector organizations with strict data governance mandates.

Blaze provides access to open-weight models including NVIDIA Nemotron, Sarvam, DeepSeek, Qwen, GLM and Gemma1

1

. The platform integrates open-source components from NVIDIA DSX OS, the modular software layer within the NVIDIA DSX AI Factory Platform. NVIDIA Dynamo handles distributed model serving with KV cache-aware routing and disaggregated prefill and decode processing, while NVIDIA Grove and KAI Scheduler coordinate inference workloads across E2E's Kubernetes infrastructure2

2

.

Hardware-Level Isolation and Production-Ready Architecture

Blaze employs NVIDIA BlueField DPUs, DOCA and Infra Controller to enforce hardware-level isolation between customers, addressing critical security requirements for regulated industries1

1

. This architecture proves essential for banks, healthcare providers and government entities that cannot compromise on data sovereignty or multi-tenancy security.

The platform targets inference-heavy applications where latency and cost per token directly impact product viability. Use cases include voice applications, coding agents, document AI, optical character recognition, image and video generation, and language processing2

2

. For latency-sensitive voice and chat workloads, Blaze delivers fast time to the first sentence, while leveraging key-value cache reuse to reduce inference costs for open-weight models.

Vishal Dhupar, Managing Director for South Asia at NVIDIA, noted that "Access to AI computing and inference services is essential to advancing India's AI ecosystem. NVIDIA's accelerated computing platform provides a foundation for cloud providers like E2E Networks to make AI more accessible to India's developers and enterprises"2

2

.

Integrated Path from Training to Production Inference

Blaze operates alongside TIR, E2E's existing AI platform, creating a comprehensive workflow from model development to production deployment1

1

. While Blaze handles API-based, token-efficient inference for open-weight models with optimized performance and seamless scaling, TIR enables self-hosted inference and model pre-training and post-training on NVIDIA Blackwell GPU clusters through managed Kubernetes. Customers can serve their own or fine-tuned models on TIR's managed infrastructure with integrated NVIDIA Dynamo, offering greater control over infrastructure and model customization.

This dual-platform approach matters as enterprises seek flexibility in how they deploy AI. Organizations can experiment and fine-tune models on TIR before moving to production inference on Blaze, or maintain hybrid deployments that balance control with operational efficiency. The token-based pricing model on Blaze allows companies to scale inference costs directly with usage, avoiding the capital expenditure associated with dedicated GPU infrastructure while maintaining performance guarantees through E2E's reference design clusters.