India’s artificial intelligence ambitions are creating demand for advanced processors, but buying powerful AI chips is only one part of building a reliable computing ecosystem. Industry executives are warning that insufficient storage, slow networking and inadequate supporting infrastructure can prevent expensive graphics processing units (GPUs) from operating efficiently. As Indian companies invest in AI computing capacity, the challenge is shifting from acquiring hardware to ensuring that every part of the system works together.
The issue is becoming more important as businesses deploy AI for software development, enterprise automation, research and customer-facing applications. Neysa, an AI cloud infrastructure provider, had 12,500 GPUs operational and planned to add another 8,000, according to its chief executive, Sharad Sanghi. Industry leaders say the return on these investments will depend on data movement, network speed, power availability and workload planning—not simply the number or generation of chips installed in a data centre.
Key Takeaways
- Beyond chips: AI infrastructure needs fast storage, high-speed networking, reliable power and efficient data pipelines.
- GPU utilisation: Storage and input/output bottlenecks can leave expensive processors waiting instead of completing useful work.
- Neysa expansion: The AI cloud provider reported 12,500 GPUs live, with plans to add another 8,000.
- Networking matters: Distributed AI training requires computing clusters to exchange large amounts of data efficiently.
- Power planning: Higher-density AI servers require data centres to prepare for increasing power and cooling demands.
- Workload selection: Businesses may not need the newest GPUs for every AI task; existing processors can be sufficient for some workloads.
- India’s opportunity: Developing the supporting infrastructure could help the country make better use of its growing AI computing capacity.
Why Buying AI Chips Is Not Enough
GPUs are designed to perform large numbers of calculations in parallel, making them useful for training AI models and running inference workloads. However, processors cannot deliver their full potential if the data they need is not available quickly enough.
AI systems depend on a chain of interconnected components. Storage systems hold training datasets, model parameters and other information. Networks move data between servers and computing clusters. CPUs coordinate tasks, while power and cooling systems keep the hardware operating within safe limits.
If one part of this chain becomes a bottleneck, the entire system can slow down.
Sharad Sanghi, co-founder and chief executive of Neysa, told The Economic Times that enterprise customers sometimes buy the latest GPUs without making corresponding investments in storage and networking. The resulting input/output bottlenecks can reduce GPU utilisation and weaken the financial returns from expensive hardware.
This matters because AI processors are costly assets. Businesses must pay for the hardware or cloud capacity, electricity, cooling, data-centre space and supporting software. When processors spend too much time waiting for data, those costs continue even though the system is producing less useful work.
What GPU utilisation means
GPU utilisation measures how actively a processor is performing work over a period of time. A low utilisation rate can indicate that a computing system is waiting for data, running an inefficient workload or operating below its available capacity.
Sanghi cited an example in which a business using only 30% of its GPU capacity could struggle to justify the cost of its AI infrastructure.
That example does not mean every company should target 100% utilisation at all times. Some workloads need spare capacity to handle demand spikes, while maintenance and other operational requirements can affect utilisation. The goal is to achieve appropriate, sustained usage for the workload without compromising reliability or performance.
For businesses, measuring utilisation alongside the cost per completed task can provide a clearer picture of whether an AI investment is delivering value.
Storage: The Often-Overlooked Part of AI Infrastructure
AI systems process large volumes of information. During model training, they repeatedly access datasets and intermediate results. During inference, they need to retrieve model data, process incoming requests and return responses quickly.
If storage cannot supply data at the required speed, GPUs may remain idle while waiting for the next batch of information.
Different workloads require different storage solutions. High-bandwidth memory (HBM), located close to the processor, supports rapid access to data used in computation. Solid-state drives can store active datasets and frequently accessed files, while hard disk drives can be useful for large volumes of less frequently accessed information.
These technologies perform different roles and cannot simply be substituted for one another. Effective infrastructure planning involves matching storage capacity, throughput, latency and cost to the needs of each application.
India’s data-centre ecosystem also relies heavily on imported enterprise storage hardware, including components and systems supplied by international manufacturers. Domestic system integration and software-defined storage capabilities can help businesses assemble and manage infrastructure, but they do not eliminate every hardware dependency.
As AI adoption grows, storage procurement and design will become increasingly important. A company that invests heavily in computing without ensuring sufficient storage performance may find that adding more GPUs does not produce a proportional increase in useful output.
Networking Can Keep Expensive GPUs Waiting
Networking is another critical part of the AI infrastructure stack, particularly when a model runs across multiple servers or data-centre locations.
Large AI workloads can be distributed across clusters containing many processors. Those processors must exchange data and coordinate their calculations. If the network cannot move information quickly enough, individual GPUs may wait for other parts of the system to catch up.
Gautam Billa, vice-president of solutions engineering at Ciena, told The Economic Times that slow connections can increase GPU idle time during distributed model training.
High-performance AI networks therefore need to support high bandwidth, low latency and reliable communication between computing nodes. The precise requirements depend on the model, workload and architecture.
The challenge extends beyond model training. AI agents may operate continuously, call external tools, retrieve information and exchange results between different components. As organisations deploy more such systems, their networks must accommodate sustained traffic rather than only occasional bursts of activity.
For Indian companies building AI services, network planning will be important both within individual data centres and across geographically distributed infrastructure.
Power and Cooling Will Shape Future AI Data Centres
Power availability is another factor that determines whether AI chips can operate efficiently at scale. Advanced AI servers concentrate substantial computing capacity—and therefore electricity demand and heat—within relatively small spaces.
Data-centre operators must plan for the power required by individual servers, racks and entire facilities. Cooling systems must then remove the heat generated by that equipment while maintaining reliable operations.
Greg Funk, vice-president of modular power converters at Vertiv, told The Economic Times that power distribution requirements become particularly important as rack power density rises. He discussed higher-voltage distribution, including the possible use of 800-volt direct current systems, as operators prepare for future generations of AI hardware.
Such changes are infrastructure-planning options rather than a universal requirement for every data centre. The appropriate design depends on equipment specifications, facility layout, safety standards and future expansion plans.
Planning ahead is nevertheless important. A data centre that installs expensive GPUs but cannot supply sufficient power or cooling may struggle to activate additional capacity when demand grows.
India’s AI expansion therefore depends on more than the availability of processors. Electricity transmission, reliable power distribution, cooling capacity and suitable data-centre locations will all influence how quickly new infrastructure can become operational.
India Is Expanding Its AI Computing Capacity
The infrastructure challenge comes as India increases access to advanced computing resources through government programmes and private investment.
The IndiaAI Mission aims to expand affordable computing access for startups, researchers, academic institutions and public-sector organisations. The government has also announced plans to increase GPU availability through its common compute facility.
In August 2026, the Ministry of Electronics and Information Technology said that 15 compute service providers had been empanelled and 237 projects approved for subsidised compute support, with 93.18 lakh GPU hours sanctioned. These figures illustrate the effort to broaden access to computing resources, although access to GPUs alone does not guarantee that every project has the storage, networking and power infrastructure required for its workload.
The private sector is expanding as well. Neysa’s reported plan to add 8,000 GPUs to its existing 12,500 operational units reflects demand for AI computing capacity among Indian and international customers.
The next phase of development will involve ensuring that this additional hardware can be used efficiently. That means integrating processors with storage, networking, software and data-centre infrastructure instead of treating each investment as a standalone purchase.
Choosing the Right Chip for Each AI Workload
Another way to improve returns is to avoid buying more computing power than a task requires.
Not every AI application needs the newest and most expensive processor. Some inference workloads can run on older GPUs, while simpler tasks may be handled by CPUs or other specialised processors.
Sanghi told The Economic Times that companies should select hardware according to the specific workload and performance required. He cited an unnamed payments company that runs batch AI jobs on spare CPU servers, making use of existing capacity rather than purchasing additional GPUs for every task.
This approach can reduce capital expenditure and improve the utilisation of equipment already in place. It can also help businesses separate workloads that require high-performance accelerators from those that do not.
The decision should be based on measured performance, reliability and total cost—not solely on the age or headline specifications of a processor. Newer chips may be justified when they materially reduce processing time or operating costs, but they are not automatically the most economical choice for every application.
What Businesses Should Measure Before Expanding AI Capacity
For organisations investing in AI infrastructure, a broader assessment can help prevent expensive bottlenecks.
| Area | What businesses should evaluate | Why it matters |
|---|---|---|
| GPU utilisation | Time spent performing useful computation | Reveals whether expensive hardware is being used effectively |
| Storage | Throughput, latency and capacity | Prevents processors from waiting for data |
| Networking | Bandwidth, latency and reliability | Supports communication between servers and clusters |
| Power and cooling | Available capacity and future requirements | Helps avoid deployment limits and overheating |
| Workload selection | Required performance for each task | Prevents unnecessary spending on premium hardware |
| Cost per task | Total infrastructure expense against useful output | Connects technical performance to business value |
These measures are useful because the number of installed GPUs alone does not reveal how efficiently an AI system operates. A smaller, well-designed cluster may deliver better value for a particular workload than a larger system constrained by data movement or power limitations.
Companies should also test their systems under realistic workloads before expanding capacity. Benchmarking can reveal whether additional spending should go toward more processors, faster storage, better networking or software optimisation.
The Bigger Picture
India’s AI ambitions are increasingly dependent on the infrastructure surrounding the chips themselves. Advanced GPUs provide the computing power needed for sophisticated AI workloads, but storage, networking, power and cooling determine whether that power can be used efficiently. As businesses expand their AI deployments, weaknesses in any of these supporting systems can increase costs and delay the benefits of investment.
The opportunity is to build a more complete AI infrastructure ecosystem, combining computing hardware with efficient data movement, reliable facilities and workload-aware system design. Government initiatives and private investment are expanding access to computing capacity, but long-term competitiveness will also depend on how effectively that capacity is operated and maintained.
Looking Ahead
India’s next phase of AI infrastructure development will require companies to evaluate total system performance rather than focus exclusively on processor purchases. Storage upgrades, faster networks, power planning and better workload allocation can help organisations improve utilisation before committing more capital to additional GPUs. For cloud providers, these improvements could also help make AI computing services more reliable and economically attractive.
As AI workloads become more complex and agent-based applications generate sustained traffic, infrastructure planning will become even more important. The success of India’s AI expansion will depend not only on how much computing capacity the country adds, but also on whether businesses can turn that capacity into useful, affordable and reliable AI services.
Frequently Asked Questions
1. Why does India need better storage and networking for AI chips?
AI processors need fast access to data and reliable communication with other components. Inadequate storage or networking can leave GPUs waiting, reducing the amount of useful work they complete.
2. What is GPU utilisation?
GPU utilisation measures how actively a graphics processing unit is working. Low utilisation can indicate that a processor is waiting for data, handling an inefficient workload or operating below its available capacity.
3. How many GPUs does Neysa have?
Neysa’s CEO told The Economic Times that the company had 12,500 GPUs operational and planned to add another 8,000.
4. Why is power infrastructure important for AI data centres?
Advanced AI servers can consume substantial electricity and generate significant heat. Data centres need suitable power distribution and cooling systems to operate this equipment reliably.
5. Does every AI application need the latest GPU?
No. The required hardware depends on the workload. Some tasks can run effectively on older GPUs or CPUs, while more demanding applications may justify newer accelerators.
6. What is the main challenge for India’s AI infrastructure push?
The challenge is to develop a complete system around AI chips, including storage, networking, power, cooling and software. Expanding hardware capacity without addressing these dependencies can raise costs and limit performance.
Source note, 17 October 2026: Technical background sources: NVIDIA GPUDirect Storage configuration guide; Neysa HPC architecture guide.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
