Neysa Rejects the AI Divide as it Takes on Big Cloud With 20,000 GPUs

Access to GPUs has emerged as a critical factor for companies building and deploying AI systems. According to the Indian government, the country's data centre capacity increased from about 375 MW in 2020 to around 1.5 GW in 2025, and is projected to reach roughly 8 GW by 2030.

Indian neo cloud provider Neysa is looking to capitalise on this growth. Its Velocis platform offers GPU-as-a-service, allowing businesses to train, fine-tune, and deploy AI workloads. Following its $1.2 billion funding round earlier this year and its entry into the unicorn club, the company is heavily investing in AI infrastructure across India, with a planned deployment of 20,000 GPUs.

“We’re aggressively scaling our silicon inventory across leading GPU platforms, and ensuring that we cater equally to the enterprise segment that deserves a world-class AI-PaaS offering,” CEO Sharad Sanghi tells AIM in an exclusive interview.

Sanghi co-founded Neysa in 2023 with Anindya Das. Before launching the company, he founded data centre services provider Netmagic Solutions, which was acquired by NTT Communications in 2012.

The company expects NVIDIA’s latest Blackwell GPUs, including the B300s, to lead the AI infrastructure push.

“Our Blackwells are already on the way. The IndiaAI Mission is a source of utilisation for us as well. We need those GPUs as quickly as possible to meet a lot of those requirements,” Karan Kirpalani, Chief Product Officer at Neysa, tells AIM.

Kirpalani also reveals that an IIT campus has placed an order for more than 200 servers. He adds that several AI-native startups, including Shodh AI, Navana.ai, Smallest.ai, GreyLabs AI, Protecto, and Kogo AI, are building and deploying their models on Neysa’s Velocis Cloud.

Neysa’s expansion comes amid a renewed interest in accelerating India's AI infrastructure. Companies including Yotta, Adani, RMZ, and AirTrunk are committing billions of dollars to AI-ready data centres, while the IndiaAI Mission is making domestic GPU capacity available to startups, researchers, and public institutions.

Recently, Amazon increased its AI infrastructure commitments in India and will now invest an additional $13 billion over the next four years to expand its cloud services. Meanwhile, Google is establishing a large-scale AI hub in Vishakhapatnam in collaboration with AdaniConneX and Nxtra by Airtel, and Microsoft has pledged $17.5 billion through 2029 to expand its AI and cloud footprint.

Frontier Models or Applications?

The debate over whether India should invest in building frontier foundation models or focus on applications has intensified in recent months. Sanghi rejects the idea that the two approaches are mutually exclusive.

“I believe that framing India’s strategy as a choice between building frontier foundation models or building applications is an analytical error,” he says.

In Sanghi’s view, strong technology ecosystems require progress at both layers.

“Frontier models matter because they expand what becomes possible. Applications matter because they turn that capability into economic value and real-world utility. The ecosystem needs both,” he explains.

Sanghi affirms Neysa’s execution layer looks to serve all kinds of AI deployments, whether it’s a team is training a massive, localised multilingual model like BharatGen’s Param 2, or deploying specialised agentic workflows. 

Inference is Key 

Sanghi says inference—not model training—is emerging as the defining challenge for organisations looking to scale AI.

“Training a model is a finite capital expense, but inference is a continuous operational expenditure that dictates the ultimate viability of an enterprise AI product,” he says. “As applications graduate from isolated pilots to core business workflows, inference becomes a recurring dependency layer.”

He adds that enterprises are no longer evaluating infrastructure solely on compute power. Instead, factors such as token economics, latency, data residency, and operational reliability are becoming critical.

Neysa’s Velocis AI platform, launched in 2024, addresses these challenges by combining orchestration, model serving, and observability into a unified AI platform. Sanghi says the company has already deployed optimised inference environments for workloads exceeding 140 billion tokens per month, helping enterprises lower infrastructure costs while maintaining production performance.

The platform is built for latency-sensitive workloads including voice AI, enterprise copilots, search, workflow automation, speech-to-text, text-to-speech, and OCR. It also aims to allay concerns around overseas data routing and data sovereignty by keeping inference and enterprise data within India.

Neysa has also recently partnered with AI inference startup Pipeshift to launch a production-grade, real-time inference platform deployed entirely within India. The offering combines Pipeshift’s managed inference platform with Velocis AI cloud, allowing enterprises to deploy open-source models such as Llama, DeepSeek, Mistral, Gemma, Qwen, and GPT-OSS in dedicated single-tenant environments through OpenAI-compatible APIs.

Challenging Hyperscalers

Despite the rise of Indian AI startups, most firms continue to rely on cloud providers such as AWS, Microsoft Azure, and Google Cloud. Sanghi believes AI workloads require a different infrastructure approach from the traditional cloud model.

“Traditional hyperscalers built highly successful platforms optimised for the CPU-driven web and application hosting era,” he says. “Production-scale AI introduces high-density architectural demands that legacy, virtualised environments were simply not designed to handle.”

To compete with global cloud providers, Sanghi argues that AI-focused infrastructure companies must offer both performance and operational advantages. In one case, Neysa delivered nearly 90% cost savings compared with proprietary models running on hyperscale clouds, Kirpalani says.

Through its Velocis Cloud platform, Neysa also provides dedicated bare-metal environments. “We provide dedicated, single-tenant bare-metal environments that eliminate the shared rate limits, noisy-neighbour bottlenecks, and cross-region routing overheads common in legacy clouds,” Sanghi says.

Sanghi also highlights the need to simplify AI operations for developers and data science teams. “By offering a seamless, single pane of glass experience that allows them to spin up GPUs as bare metal, virtual machines, or containers and publish models into inference with granular visibility and control, we’re expanding the frontier of choice.”

He believes the winner in the AI infrastructure race will be the platform that can deliver predictable economics and operational reliability.