Analysis · AI & Deep Tech

Neysa: AI chip shortage shifts to newer GPUs

What they do

Global data centre demand of 21.1 GW in 2025 far exceeded supply of 8.9 GW, per Jefferies

Hyperscalers lock up most NVIDIA shipments, leaving smaller Indian enterprises waiting months for chips

NVIDIA is concentrating on newest architectures, retiring Hopper generation and declaring H100 end-of-sale

India's AI compute market faces a structural GPU shortage that has eased from peak levels but remains severe.
Global data centre demand reached 21.1 GW in 2025, but only 8.9 GW of capacity became operational, a shortfall of about 12 GW.
Hyperscalers like Microsoft, Amazon, Google and Meta have locked up the overwhelming majority of NVIDIA's latest shipments through long-term purchase agreements.
Lead times for next-generation enterprise AI GPUs range between 36 and 52 weeks, with some new orders pushed into 2027.
Older-generation chips like the H100 have been declared end-of-sale as NVIDIA concentrates on newer architectures.
Bottlenecks have moved downstream into packaging technologies like CoWoS, high-bandwidth memory (HBM), and power infrastructure.
Export controls and geopolitics create a tiered system where strategic buyers and sovereign AI programmes are prioritised over smaller enterprises.
Indian infrastructure providers now plan compute requirements several quarters ahead and shift away from pay-as-you-go models.

Nearly 90% of advanced logic chip production (2 nm, 3 nm, 5 nm nodes) sits in Taiwan, creating geopolitical risk.
New factories of memory suppliers SK Hynix, Samsung and Micron are not expected to add much capacity until 2027 or 2028.
Hyperscalers are expected to invest $770 billion in AI infrastructure in 2026, up 74% year-on-year.
Ashok Chandak of IESA called GPU allocation a sovereign security issue, urging India to manufacture its own compute.

India's AI compute market consists of cloud providers, data centre operators, and AI startups that rely on imported high-end GPUs.
Companies like Yotta, Neysa, and NeevCloud provide GPU cloud services to Indian enterprises and AI model builders.
Infrastructure providers use mixed fleets of old and new hardware to keep workloads moving.
Training workloads have become scheduled events, while inference (running trained models) is now the dominant consumer of compute.
Software optimisation has emerged as a competitive advantage over brute-force scaling.

Chipmakers are diversifying capacity into the US, Europe, India and Southeast Asia amid geopolitical tensions.
India's semiconductor mission aims to build domestic chip manufacturing capabilities.
Smaller enterprises are making capacity planning a core part of their business strategy.

Product: AI cloud infrastructure platform providing GPU-backed compute capacity and AI inference services
Customers: Large enterprises and organizations requiring substantial AI compute for training and inference
How it makes money: Infrastructure-as-a-Service model providing GPU computing resources and inference capabilities
What sets it apart: Fully deployed within India, competing against global hyperscalers with domestic infrastructure
Operations: Scaling infrastructure with planned deployment of over 20,000 GPUs across the country