If vLLM already solved LLM serving, why did SGLang appear?
AI in Agriculture
India has fed itself through some remarkable moments of reinvention.
The Green Revolution continues to be the most cited one. Starting in Punjab in 1965, India went from chronic food shortages to near self-sufficiency in under a decade. Within five years, the country was producing enough grain for its entire population – roughly 70% of it coming from Punjab alone. Other states watched and copied. The formula was simple: more fertilizer, better seeds, more irrigation.
And it worked – for a while.
The Math Has Changed
By 2050, India’s population will be around 1.68 billion people. Feeding them will require approximately 70% more food than what the country produces today. While the Indian agricultural land has the physical potential to deliver about 50% more – it still leaves a gap of 20% that inputs alone can’t close.
Meanwhile, urban migration is accelerating. India’s urban population is expected to surpass its rural population by 2040. More people in cities means more land consumed by construction and fewer people left to farm. And the workforce that grows India’s food is shrinking while the population that needs it keeps growing.
Pushpendra Singh, founder of Annam AI, put it plainly at the India AI Impact Summit in Delhi: “What are the options? Technology, technology, and technology. There is no other way.”
Enable AI for Real-World Agriculture
Run agricultural AI at scale with infrastructure built for complex, data-intensive workloads.
The Green Revolution Left a Bill Nobody Wants to Pay
The input-heavy model that drove India’s agricultural growth also left a legacy most people don’t talk about enough.
Decades of over-fertilization and pesticide use have contaminated water and soil across large parts of the country. Residues move through crops and into produce. The consequences are visible today – in contaminated water sources, degrading soil quality, and yields that are quietly declining in regions that were once the country’s most productive.
You can’t fertilize your way out of a structural problem. The next phase of agricultural growth can’t come from the same levers that drove the last one. The opportunity lies in better decisions, not more inputs.
Agriculture Is Unlike Other AI Problems
Most AI development happens in relatively controlled environments. Data flows in from structured systems, variables are limited, and infrastructure (electricity, connectivity, devices) is largely taken for granted. Agriculture is a different beast.
Farms span vast with fragmented geographies. Soil conditions, water availability, pest behavior, and microclimate variations differ not just between states but between adjacent fields. Nearly 40% of India’s agricultural commodity is wasted every year – not from bad farming, but from cold chain failures and poor last-mile transportation. And many farmers till date don’t even have smartphones.
Roughly 70% of the variables in agricultural AI are outside human control. That alone makes it a fundamentally different engineering problem from anything being built in enterprise software.
Pushpendra puts it candidly, “We are actually at the level of Nokia 3310. And we need to reach the level of iPhone 17 Pro – and we’ll reach that level in 2030, not 2026.” The basic job right now is making the 3310 work reliably.
Before Intelligence, There Has to Be Data
The prerequisite for meaningful agricultural AI is data that largely doesn’t exist yet at the right granularity.
Useful agricultural intelligence has to be hyperlocal. That means field-level observations, localized weather patterns, crop health indicators, soil readings, and the kind of contextual knowledge farmers carry but rarely gets captured in structured form.
Consider India’s Kisan Call Center – the primary farming advisory service for millions of farmers. Analysis of over 35 million questions from the system found that not a single question or answer was complete. A farmer calls with a problem – a BSc agriculture graduate answers from memory. If they don’t know, the call escalates to a senior with more years of
experience but the same knowledge gaps. The intelligence underwriting India’s agricultural advisory system at scale is largely individual, inconsistent, and unreliable.
Closing that gap requires more than better models. It requires a data foundation strong enough to power them, and the infrastructure to train, retrain, and run those models as conditions change season to season. This is where teams working on agricultural AI increasingly turn to purpose-built GPU infrastructure like Neysa Velocis. Processing hyperlocal datasets at scale, running continuous model updates, and supporting real-time inference demands that general-purpose cloud wasn’t designed to meet efficiently.
Neysa Velocis: The Infrastructure Behind Green Intelligence
Neysa provides the AI infrastructure foundation that lets agricultural AI teams stay focused on the actual problem. Annam AI for instance, runs their models on Neysa Velocis – solving the challenges Indian farmers face today, from climate uncertainty to resource scarcity. Whether the objective is training on hyperlocal crop data, running inference at scale across multiple languages, or building real-time advisory systems that have to work under time pressure, the underlying infrastructure has to be purpose-built for the workload.
Agriculture doesn’t need another generic cloud. It needs compute that performs reliably at the scale the industry demands.
Watch the full conversation with Pushpendra Singh, founder of Annam AI, at the India AI Impact Summit here.
