India has plenty of companies building AI applications, but very few are focusing on the hardware powering them. Netweb Technologies is one Indian company building that infrastructure, from high-performance computing systems to GPU servers.
Founded by Sanjay Lodha in 1999, Netweb Technologies started out building high-performance computing (HPC) systems for scientific and research applications.
In 2004, it deployed the Kabru supercomputer at the Institute of Mathematical Sciences in Chennai. It later supplied systems for projects including the PARAM Yuva-II supercomputer at the Centre for Development of Advanced Computing (C-DAC) and PARAM Ambar at the Indian Space Research Organisation (ISRO) for atmospheric research.
Over time, Netweb expanded into AI systems, data centre servers, high-performance storage, and private cloud solutions.
In 2023, it deployed AIRAWAT, India’s AI supercomputer, and entered a manufacturing partnership with NVIDIA for server designs based on its Grace and Grace Hopper platforms. The company went public on the BSE and NSE in July 2023.
NetWeb recently raised approximately ₹1,200 crore through a Qualified Institutional Placement (QIP). As of November 12, its shares have been up 35% year-to-date, trading at around ₹4,160 apiece.
In an exclusive interview with AIM, Swastik Chakrabarti, Vice President of Technology at Netweb, says that the company is one of nine global manufacturers of NVIDIA GPU servers and the only one in India. The company develops custom server designs that undergo NVIDIA certification.
He cites the company’s work on power delivery for NVIDIA’s NVL72 architecture as an example. Netweb developed an alternative design that converts AC power to DC internally within the server, which Chakrabarti says received NVIDIA’s approval.
The company also benchmarked its servers against international original equipment manufacturers (OEMs) through MLPerf, a benchmark suite for measuring AI system performance in training and inference.
Chakrabarti says Netweb’s servers were listed alongside systems from HP, Dell, Lenovo, Cisco, Google, and Supermicro in MLPerf results. Beyond NVIDIA, Netweb is working with AMD on its MI350 and MI450X accelerators, as well as Qualcomm on neural processing units (NPUs).
The company is also expanding its software capabilities to help customers deploy AI infrastructure according to their business needs, rather than buying hardware without a clear use case.
Its software offerings include Tyrone cluster management tools, Kubyts for container management, and Skylus for private cloud and hyperconverged infrastructure. The company also offers Skylus.ai for deploying and managing GPU-based AI infrastructure, alongside specialised AI computing and storage platforms
Agent Guard
The company is developing an agent security framework called Agent Guard, which requires AI agents to obtain time-limited access tokens before accessing enterprise data.
As enterprises deploy AI agents that can access databases, data warehouses, and other internal systems, Netweb is developing controls to restrict their access. Chakrabarti says Agent Guard would require both human users and AI agents to register with an identity and access management system.
Agents would need to request a token each time they access a dataset or other resources. “These tokens will be valid for 5–10 minutes and can be used only once,” he says.
The system would log every data access request alongside the token used to authorise it. The approach aims to ensure that agents operate within defined permissions, with human approval built into the access process.
Netweb is also working on post-quantum cryptography-ready solutions that store authentication keys in a compliant vault, rather than leaving them exposed to potential theft.
Government Partnerships
Indian government customers account for 55% of Netweb’s clientele, while enterprises contribute the rest. Chakrabarti says enterprise demand for GPU servers was growing faster, with the company serving sectors including banking, financial services and insurance (BFSI), manufacturing, retail, and healthcare.
Netweb Technologies counts several Indian government agencies and public-sector organisations among its customers, including C-DAC, ISRO, the Defence Research and Development Organisation, and RailTel.
Chakrabarti says government organisations are using AI for document summarisation, procurement and request-for-proposal processing. Also, retrieval-augmented generation (RAG) pipelines and optical character recognition (OCR) are helping process lengthy documents and extract relevant information.
AI agents are also being used to assist with RFP workflows, including specification generation and vendor validation. Other applications involve monitoring the Government e-Marketplace portal to identify procurement activity and classify requirements.
In healthcare, it is exploring the use of AI chatbots as assistants for doctors. Chakrabarti also cites more advanced applications, including using AI agents to help direct medicines towards specific cancer cells.
Multi-Tier GPU Infrastructure
Netweb is developing a multi-tier token generation engine to help customers use existing GPU infrastructure instead of replacing it whenever a newer generation becomes available.
The system can distribute workloads across accelerators such as NVIDIA V100 and A100 GPUs, H100 and H200 GPUs, and the newer B200 generation.
Chakrabarti says the choice of hardware should depend on the performance requirements of each workload rather than the availability of the latest processors. For instance, a document summarisation task that can take 30 minutes to complete may not require a B200 or H200 GPU if an older A100 can meet the deadline.
The approach allows organisations to reserve higher-performance accelerators for latency-sensitive workloads while running less demanding tasks on older hardware.
Netweb is also exploring storage optimisation and CXL memory to improve the performance of existing systems. Chakrabarti says storage optimisation could deliver 10–30% better performance from the same hardware, depending on the workload and configuration.
The company is also examining ways to improve prefill and decoding performance in AI inference. Prefill processes the input prompt, while decoding generates the model’s response. Improving these stages can reduce response times.
Sovereign Server?
Netweb sources many of its server components from other companies and countries. The company is open to sourcing more components from Indian suppliers as the domestic semiconductor ecosystem grows, paving the way for fully sovereign servers.
That said, Chakrabarti says quality, availability, and delivery capacity would determine whether local components could be integrated into its systems.
Beyond processors, memory and storage, servers require components such as capacitors, inductors, field-effect transistors (MOSFETs), and power-conditioning units.
“We will be more than happy to have those components sourced locally once they match the standards,” he says.
He adds that suppliers must also be able to meet demand at scale. A component manufacturer capable of delivering only a few thousand units may not be sufficient when the requirement runs into millions.
