India's AI Chip Shortage Reshapes Cloud Compute Market

India's AI sector is growing fast, but it keeps running into the same hard ceiling: not enough chips. The acute GPU shortage that defined the early generative AI boom has eased at the lower end of the market, but the scramble for next-generation accelerators is, if anything, more intense than before. Cloud providers, startups, and large enterprises are now operating in a tiered supply system where who you are matters as much as what you can pay.

The numbers behind the crunch are stark. A Jefferies report found that 8.9 gigawatts of global data centre capacity came online in 2025 against demand of nearly 21.1 gigawatts, leaving a shortfall of roughly 12 gigawatts. With hyperscalers projected to pour $770 billion into AI infrastructure in 2026, up 74% year on year, the gap is expected to widen before it narrows.

India Electronics and Semiconductor Association president Ashok Chandak describes the situation as a structural imbalance rather than a temporary squeeze. Delivery timelines for AI chips have improved since their worst point, he says, but global data centre demand continues to outpace supply by a wide margin.

Why New Chips Are the Hardest to Get

The bottleneck has shifted from older GPUs to the newest generation. NVIDIA is concentrating production on its latest architectures, and the Hopper generation is being wound down, with the H100 already declared end-of-sale, according to Karan Kirpalani, chief product officer at AI-native cloud infrastructure startup Neysa. Older chips are easier to find now. The newest high-performance GPUs are not.

Lead times reflect this directly. Sunil Gupta, cofounder and CEO of Yotta, says large deployments ran 6 to 15 months in 2024. The situation is more predictable now, he notes, but large builds still require early reservations and close coordination with hardware vendors and system integrators. Narendra Sen, cofounder of NeevCloud, adds that timelines once as short as two months have stretched to roughly four months for dedicated cluster setups, with some new orders for next-generation enterprise GPUs now pushed into 2027, with lead times ranging from 36 to 52 weeks.

The problem does not stop at the chip itself. Packaging technologies such as CoWoS (chip-on-wafer-on-substrate), high-bandwidth memory, and power infrastructure have all become separate chokepoints. For NVIDIA's B200 and B300 accelerators, Sen says, memory and networking components now carry longer lead times than the chips they support. The three dominant memory suppliers, SK Hynix, Samsung, and Micron, are not expected to add meaningful new capacity until 2027 or 2028, which means the crunch in AI server memory is unlikely to ease before then.

Geopolitics Sorts Buyers Into Winners and Waitlists

Export controls are reshaping who gets chips and when. Nearly 90% of advanced logic chip production sits in Taiwan, giving the global supply chain a narrow geographic base. That concentration, combined with US export restrictions, has pushed chipmakers to diversify into the United States, Europe, India, and other regions, even as those new facilities remain years from scale.

The immediate effect is a tiered market. Hyperscalers like Microsoft, Amazon, Google, and Meta have committed hundreds of billions of dollars through long-term purchase agreements, effectively locking up the bulk of NVIDIA's latest shipments. Sovereign AI programmes and strategic partners come next. Smaller enterprises are left competing for what remains. Sen of NeevCloud puts it plainly: export controls and compliance requirements are creating a system where vendors prioritise strategic markets first, and smaller buyers absorb the delay.

For India, this dynamic carries particular weight. The country imports almost all of its high-end chips and has no domestic alternative at scale. IESA's Chandak argues that GPU allocation has become a sovereign security issue and that India must move toward manufacturing its own compute. The government's IndiaAI Mission is already providing some GPU access to domestic startups, but it remains a partial solution against the scale of demand.

With supply constrained and procurement timelines stretching across quarters, Indian cloud providers have shifted away from the pay-as-you-go model. Gupta of Yotta says large enterprises, model builders, and AI platform companies are now planning compute requirements several quarters ahead. Forecasting demand accurately has become a competitive advantage in itself.

Scarcity is also changing how GPUs are actually used, and this may be the more durable shift. Voice AI company Murf.AI divides its compute between scheduled training runs and continuous real-time inference, reserving guaranteed capacity ahead of training instead of relying on spot markets, according to cofounder and CEO Ankur Edkie. Enterprise AI startup Nurix, cofounded by Mukesh Bansal, operates a fleet of 15 to 22 GPUs and segments workloads so that real-time inference runs during peak hours and fine-tuning happens in off-peak windows. It spreads workloads across hyperscalers and neoclouds, using NVIDIA H100s for heavy inference and older T4 and L4 chips for lighter models. CoRover goes further still: its composite AI architecture routes only the most compute-intensive tasks to large language models, allowing nearly 80% of its workloads to run without heavy GPU inference, with peak usage across providers reaching around 1,200 to 1,250 GPUs.

The common thread is that software optimisation and workload scheduling are replacing raw chip count as the primary lever of competitive advantage. Training is becoming a planned event rather than a continuous process. Inference is now the dominant compute workload. Multi-cloud sourcing is standard practice for anyone serious about reliability.

The structural compute deficit will not resolve quickly. New memory capacity is years away, geopolitical controls on chip exports show no sign of easing, and hyperscaler demand will continue to absorb the majority of new supply. For India's AI startups, the practical question is no longer just how many GPUs they can buy, but how efficiently they can extract value from the compute they already have access to. That shift from procurement to optimisation may define which companies scale and which stall.