Neysa & Pipeshift Launch Realtime Inference in India
A New Approach to AI Inference in India
India is an inference-first market. Voice agents, copilots, vision-based search, agents, reasoning workflows. The applications driving AI adoption here are inference-heavy by design, and the country runs more inference traffic per user than almost any other market in the world.
Yet inference itself has lived somewhere else – on closed-source LLM stacks built for other markets. The result is a cost base that swings on someone else’s pricing decisions, latency dependent on network paths nobody here controls, and a data residency posture that increasingly fails procurement at regulated customers.
Today, Neysa and Pipeshift are launching real-time inference for open source LLMs, deployed entirely inside India. A fully managed inference platform built for the workloads, economics, and sovereignty requirements that production AI in India now demands.
The cost of running production AI in India
- Token pricing breaks unit economics: Providers raise prices, drop them, retire models, and introduce tiers on their own schedule. None of it is predictable enough to plan a P&L around.
- Latency depends on network paths nobody here controls: Indian-region LLM API endpoint alternatives on general-purpose clouds do not contractually guarantee that inference stays in the country.
- Sovereignty is now regulation, not preference: India’s DPDP Rules are rolling out through 2027, and the RBI’s ‘Master Direction on Outsourcing’ already binds financial services. Meaning, your data has to stay within the country.
- You don’t own the model: A closed provider can deprecate it, throttle your traffic during a spike, or shift pricing without warning. You cannot fine-tune on your own data.
Collected together, they make the case for a different stack.
What’s changed
Two shifts have happened in tandem.
Open source has caught up. Llama 4, Qwen 3.6, DeepSeek V4, Mistral Large 3, Gemma 4, GPT-OSS. For most production workloads, these models match or come close enough to the best closed models that the cost gap becomes the deciding factor.
Indian AI has hit production scale at the same time. A new generation of voice agents, copilots, and reasoning workflows is hitting traffic levels where the structural trade-offs of the closed-API stack matter every day.
The models are now good enough to take the bet on an open source stack, and the traffic is now large enough to make the bet worth taking.
How real-time inference on Neysa works
Our engineering team, working alongside Pipeshift’s, benchmarks open source LLM against your evals, writes custom CUDA kernels, picks the right inference engine between vLLM and SGLang, tunes the parallelization strategy for your latency target, and stands up a dedicated, single-tenant endpoint in the region you need it.
The platform handles autoscaling, failover, kernel updates, and model swaps as new open source releases arrive. You drop the endpoint into your existing stack through an OpenAI-compatible API.
Pricing is hybrid. A reserved baseline handles your steady traffic. On-demand scales through spikes. You commit to a monthly spend pool, and as new silicon ships, your endpoint moves to it with no contract renegotiation.
“There is a clear line between AI that works in a demo and AI that works in production. Crossing that line takes more than a good model. It takes infrastructure that holds latency under load and keeps costs predictable at scale. That is the line our partnership with Neysa helps Indian companies cross.”
Arko Chattopadhyay, Co-founder and CEO, Pipeshift
What this looks like in production
The clearest case to look at is voice AI, where the trade-offs of the closed-API stack hit hardest. Nurix AI runs voice agents in production at scale. Their bar is sub-second latency on every interaction, because anything slower is a pause the user hears. On the closed-API stack, that bar was hard to hold from India. On Neysa, with Pipeshift’s inference layer tuned to their workload, their TTFT dropped by 3x compared to their previous setup.
“We needed sub-second LLM latency for voice agents in production, and real-time inference from Neysa and Pipeshift cut our TTFT 3x versus our prior setup in India. Their team’s support and quick resolution time has helped make seamless rollouts to production.”
Pushkar Patel, Nurix AI
The same stack works for teams with different priorities. ZingHR runs open-source LLMs for HR workloads where sensitive customer data has to stay inside their environment. The deployment had to be fast, cost-efficient, and fully in-country.
“We needed secure, high-performance open-source LLM deployments within our own infrastructure, and the execution was seamless from day one. The inference speeds were consistently fast, the deployment was highly cost-efficient, and most importantly, sensitive customer data always remained within our environment.”
Lokpal Vora, ZingHR
For teams running heterogeneous workloads, LLMs alongside speech-to-text, TTS, vision, or custom containers, all of it runs on a single cluster in a single region. The round trips disappear. The egress markup disappears with them.
“India has produced some of the most sophisticated AI applications in the world. What it has lacked is the infrastructure layer to take those applications to production at full performance. Neysa is that layer. Our partnership with Pipeshift brings real-time inference to Indian companies for the first time as a fully managed service, and the results our customers are already seeing speak for themselves.”
Karan Kirpalani, Chief Product Officer, Neysa
Get started
The fastest way to evaluate real-time inference on Neysa is to put your current setup in front of us.
Bring your test set, your latency target, and your monthly spend ceiling. We come back with a benchmarked endpoint and a deployment plan that meets all three. The typical path from first call to a production endpoint is under two weeks.
The inference layer India has been waiting for is live. We are looking forward to seeing what gets built on it.
Explore more: here
