Why India’s AI Boom Is Running On A Waiting List

The way India buys AI compute is being rewritten. The acute GPU shortage that dominated the early GenAI boom has eased from its most extreme levels, but the market is still far from relaxed. 

Industry insiders say the scramble has now shifted from older-generation GPUs to acquiring the latest-generation AI chips, which remain difficult to source. To remedy this, cloud providers are now moving to reserve capacity months in advance and leaning on mixed fleets of old and new hardware to keep workloads moving. 

At the same time, geopolitics, export controls and the concentration of semiconductor production in a handful of regions have created a system in which large strategic buyers are prioritised, while smaller enterprises are waiting in a queue. 

As if this were not enough, the supply chain bottleneck has spread into the less visible layers of the chip manufacturing stack, meaning that all parts of the AI system may not be available at the same time.

The pressure is also altering behaviour on the demand side. AI companies are responding to chip scarcity by buying more GPUs and using them more efficiently and effectively. Training is becoming more scheduled, inference is becoming the dominant workload, and software optimisation is emerging as a competitive advantage. 

This has turned compute into a strategic resource rather than a simple purchase. The effect is especially sharp in India, where almost all high-end chips are imported. As demand continues to outrun supply, how is this gap for new chips reshaping the way cloud providers and startups source, allocate and consume compute? Let’s find out…

Sourcing Problem: Waiting Time Stretches For New Chips

At the top of the chain, the core problem is that the GPU supply is not keeping pace with demand. As per a Jefferies report, 8.9 GW of global data centre capacity became operational in 2025 against demand of nearly 21.1 GW, a shortfall of about 12 GW. 

With hyperscalers expected to infuse $770 Bn (up 74% YoY) in this sector in 2026, the crunch is only expected to deepen.

However, according to India Electronics and Semiconductor Association (IESA) president Ashok Chandak, delivery timelines for AI chips have improved since the peak of the shortage, but global data centre demand continues to massively outpace supply. He frames it as a structural imbalance rather than a passing squeeze.

This “imbalance” is showing up in the kind of hardware that buyers can secure. In the words of AI-native cloud infrastructure startup Neysa’s CPO, Karan Kirpalani, NVIDIA is concentrating on its newest architectures and the Hopper generation is being retired, with the H100 already declared end-of-sale.

Consequently, older-generation chips are becoming easier to source, while the procurement of newer high-performance GPUs remains severely constrained.

Cloud providers note that this tension is showing up directly in delivery schedules. Yotta’s cofounder and CEO Sunil Gupta said that lead times for large deployments ran anywhere from 6-15 months in 2024. In contrast, he believes that the current situation is more predictable but “far from gone”, adding that large builds still require early reservations and close coordination with OEMs, system integrators and GPU vendors.

On the flip side, Narendra Sen, cofounder of NeevCloud, adds that lead times that were historically as short as about two months have stretched to roughly four months for dedicated cluster setups. 

The bottlenecks are not just limited to chip fabrication. The issues have also moved downstream into packaging technologies such as CoWoS (chip-on-wafer-on-substrate), the limited supply of high-bandwidth memory (HBM), and the power infrastructure needed to run large clusters.

For newer accelerators like NVIDIA’s B200 and B300, Sen said, memory and networking components now carry longer lead times than the chips they support. Simply put, AI chips may be available much before the memory and networking parts needed to build the full AI system. 

The new factories of three dominant memory suppliers (SK Hynix, Samsung and Micron) are also not expected to add much capacity until 2027 or 2028, so few expect the crunch to ease before then. 

Since large cloud companies lock in long-term orders that smaller players can’t match, the same pressure squeezing GPUs is now hitting the memory every AI server depends on. This deepening infrastructure crunch has made securing chips not just a financial race, but a high-stakes game of prioritisation.

The Broken, Tiered Market

As the dogfight for GPUs continues, experts argue that geopolitics has become one of the most powerful market forces. The reason? Concentration. Nearly 90% of advanced logic chip production (2 nm, 3 nm, and 5 nm nodes) sits in Taiwan, while China dominates the critical minerals space. This has pushed chipmakers to diversify capacity into the US, Europe, India and other Southeast Asian nations even as they deal with other geopolitical squabbles. 

“Export controls and compliance requirements are reshaping GPU allocation, creating a tiered system. Vendors are prioritising strategic markets, sovereign AI programmes, and hyperscalers, which often reduces availability for smaller enterprises. Because of these supply chain complexities, tariffs, and material shortages, lead times for next-generation enterprise AI GPUs now range between 36 and 52 weeks, with some new orders being pushed into 2027,” said  NeevCloud’s Sen.

On top of this, hyperscalers like Microsoft, Amazon, Google and Meta are committing hundreds of billions of dollars towards AI infrastructure, effectively locking up the overwhelming majority of NVIDIA’s latest shipments through long-term purchase agreements.

This has left many smaller enterprises gasping for AI compute. To remedy this, many companies are now making capacity planning a core part of their business. 

“AI infrastructure has become a strategic resource rather than an on-demand purchase. Large enterprises, model builders and AI platform companies are now planning their compute requirements several quarters ahead,” said Yotta’s Sunil Gupta.

For Indian infrastructure providers, accurately forecasting demand has become a competitive advantage. Cloud providers are reserving capacity years in advance and using older-generation hardware until newer clusters arrive, shifting away from the pay-as-you-go model to shield themselves from supply disruptions.

For India, which imports almost all of its high-end chips, this tiering carries strategic weight. IESA president Ashok Chandak argues that GPU allocation is a sovereign security issue, and the country must manufacture its own compute.

Doing More With Less

Scarcity is changing not only how GPUs are bought but also how they are used. AI companies increasingly see utilisation, not procurement, as their biggest competitive advantage. Training workloads are becoming scheduled events, inference is becoming the dominant guzzler of compute, and software optimisation is replacing brute-force scaling. 

Voice AI company Murf.AI illustrates the pattern. Cofounder and CEO Ankur Edkie says the startup divides compute into scheduled training and continuous real-time inference for its live voice traffic, with most day-to-day capacity going to inference. Rather than chasing raw GPU count, Murf matches capacity to demand by reserving guaranteed capacity ahead of scheduled training runs instead of relying on the unpredictable spot market.

Mukesh Bansal-cofounded Nurix, an enterprise AI startup, has also adopted a similar approach, albeit at a smaller scale. Operating a fleet of 15 to 22 GPUs, Nurix segments workloads so that real-time inference runs during high-throughput hours while fine-tuning runs in off-peak windows. 

To avoid being hostage to any single provider’s availability, the company spreads its workloads across hyperscalers and neoclouds, using NVIDIA H100 for heavy inference and older architectures like the T4 and L4 for lighter models.

Similarly, CoRover uses a composite AI architecture that routes only the most compute-intensive workloads to LLMs, enabling nearly 80% of its tasks to be handled without GPU-heavy inference. Cofounder Ankush Sabharwal notes that the startup follows a hybrid deployment strategy, sourcing GPU capacity from multiple providers, including Google Cloud, Yotta, NxtGen and the IndiaAI Mission, with peak usage reaching around 1,200-1,250 GPUs.

All said and done, this structural compute deficit could decide the fate of India’s AI ambition. With chips scarce and processing rationed, can homegrown AI startups move on from how much compute they can purchase to aggressively optimise what they already have?

The post Why India’s AI Boom Is Running On A Waiting List appeared first on Inc42 Media.