By Mrinal Kochar, SVP of Engineering at Sandisk
India’s data centre story continues to evolve. For much of the past decade, the capacity build-out was fairly easy to understand. More people came online, data consumption increased, video streaming caught on, 4G and 5G networks expanded and businesses began generating more data. Data centre capacity grew alongside these shifts. However, Artificial intelligence (AI) is changing that equation considerably, including the underlying infrastructure.
This change matters as India’s AI infrastructure build-out gathers rapid momentum. According to Cushman & Wakefield, India’s operational data centre capacity reached approximately 1.8 gigawatts (GW) in the first half of 2026, with approximately 3.9 GW of upcoming supply expected to be delivered by 2030.
Traditionally, data centres have been built for high availability and transactional processing across a broad mix of workloads, including serving web content, managing databases, hosting applications, supporting cloud compute and so on. AI data centres, by contrast, place greater emphasis on parallel computing and high-throughput data movement.
In the era of AI, significant amounts of data are being generated and stored. In tandem, networks must move large quantities of data between accelerators, be it for training large language and multimodal models that can use petabyte-scale data sets, or running cloud-based AI compute for apps and services.
If the data cannot move fast enough from the storage to the GPUs, compute resources may be underutilized, reducing overall system efficiency.
Fast storage keeps the entire compute machinery moving, at the required pace. That has necessitated increased interest in fast data lakes that are backed by high-capacity and high-performance flash storage.
The Rise of ‘Fast Data Lake’ Architecture
Historically, data lakes were built as scalable and volume-driven storage repositories, for storing unstructured or semi-structured data troves which could be a mix of text, videos, audio, code and more. AI is shifting that paradigm, as datasets are no longer just archives, but can play an important role in machine learning and for analytics.
Modern AI data centres can use fast data lakes, designed on fast flash drives to reduce delays in data ingestion, preparation and access. While every piece of data doesn’t need to move to a fast data lake, AI workloads can still create a significant active tier within data lake. Datasets, frequently retrieved documents, embeddings, model weights, checkpoints, image and video assets, and intermediate datasets may need to move repeatedly between storage and compute; and traditional capacity-oriented data lakes may not provide the performance required for some of these workloads. This is where high-performance NVMe storage comes in, supporting capacity-intensive applications and workloads such as fast data lakes.
Additionally, power efficiency is an important consideration for enterprise SSDs, alongside storage performance. A May 2025 Sandisk white paper introduces Performance and Power Efficiency (PPE) as a metric for evaluating SSD performance relative to power consumption in AI-driven data centres.
As India’s data centre capacity expands in the coming years, the build-out is likely to place increasing emphasis on infrastructure capable of supporting data-intensive AI workloads, including fast data lakes. Storage performance and capacity will remain important as datasets become larger, models evolve, and AI workloads become more demanding. Flash storage can therefore play an important role in supporting the performance and efficiency of AI data centres.