In an interaction with Express Computer, Vinay Sinha, Managing Director, India Sales, AMD, discusses how enterprises can build cost-efficient, scalable and energy-conscious AI infrastructure by adopting a workload-centric approach, balancing CPUs and GPUs, and leveraging software optimisation to maximise performance and long-term value. He also examines the evolving economics of AI infrastructure in India and the priorities that will shape the next generation of AI-ready data centers.
Q. AI investments are accelerating, but enterprises are under pressure to justify returns. How is the conversation around AI infrastructure evolving from raw performance metrics to performance-per-dollar and business outcomes?
AI infrastructure decisions are increasingly moving beyond peak performance alone. Enterprises increasingly want to understand what each investment delivers in terms of business value, operating efficiency and costs over time.
This means metrics such as performance per dollar and performance per watt matter alongside raw compute performance. Hardware utilisation, software, networking, energy consumption, and overall system density all influence the economics of AI deployment.
A powerful system delivers limited value if accelerators are under utilised, workloads are difficult to deploy, or operating costs rise without a corresponding improvement in output. That is why the economics need to be considered across the complete system, from CPUs and GPUs to memory, networking and software.
In particular, software can extend the value of AI infrastructure over time. As models, frameworks and inference techniques evolve, new runtimes, broader framework support and software optimisations can improve performance and enable new capabilities without replacing the underlying hardware. That makes software an important part of the long-term economics of AI infrastructure.
At AMD, we give customers the flexibility to match the right compute and system architecture to the workload through AMD EPYC CPUs, AMD Instinct GPUs, AMD Pensando networking and AMD ROC open software. Ultimately, the return on AI infrastructure investment comes down to how quickly organisations can move new ideas into production, how effectively it improves productivity, and how it translates AI investment into business impact.
Q. As organisations deploy increasingly complex AI workloads, what are the key factors data center operators should consider when balancing compute performance, energy consumption, and total cost of ownership?
The important question is not just about how fast the chip runs, but how much useful AI work an entire infrastructure footprint can deliver within real power, cooling, and space constraints.
As AI becomes more agentic, operators are supporting much more than model inference. Orchestration, data services, APIs, caching, middleware and tool execution can all become significant parts of the workload. That means infrastructure needs to be balanced across CPUs, GPUs, memory, networking, and software so that one component does not become a bottleneck for the rest.
This is also why we believe rack-level capacity is becoming a more meaningful planning metric than isolated component performance. Customers are deploying racks, not benchmark results. The question is how much useful service capacity they can deliver within a given power envelope, while maintaining high utilisation and manageable thermal and operational complexity.
From an economics perspective, the winning architecture is therefore not necessarily the one with the highest peak performance. It is the one that delivers the right performance per watt, keeps accelerators productive, minimises infrastructure overhead, and gives customers the flexibility to adapt as workloads evolve. That is where balanced, open, and end-to-end AI infrastructure becomes increasingly important.
Q. India is witnessing rapid growth in digital services, AI adoption, and data localisation requirements. How do you see the economics of AI infrastructure evolving in the Indian market, and what unique opportunities or challenges does it present?
Indian organisations have a compelling combination of strong demand for compute and a sharp focus on cost and infrastructure efficiency. At the same time, organisations also have to consider where data is processed, and practical constraints around power, cooling and space. This makes performance for every rupee invested, performance for every watt and infrastructure utilisation particularly relevant in India.
There will not be one deployment model that works for every use case. We see a growing role for right-sized, full-stack infrastructure in India, with enterprise data centers, cloud, neo clouds, hybrid environments, the edge and AI PCs all playing a role depending on data requirements, latency, compliance and cost. Not every workload requires the same combination of CPUs, GPUs, memory, networking and software, creating opportunity to build AI infrastructure around the workload rather than around a single type of accelerator.
For AMD, the opportunity is to give customers that flexibility across AMD EPYC CPUs, AMD Instinct GPUs and open software such as AMD ROCm, supported by our partner ecosystem – helping organisations optimise performance, efficiency and overall infrastructure economics.
Q. Many enterprises are moving from AI experimentation to production-scale deployments. What lessons are emerging around optimising AI workloads to achieve maximum efficiency without significantly increasing infrastructure costs?
Production changes the economics of AI. A model may perform well during proof of concept, but at scale, issues such as latency, data movement, accelerator utilisation and software readiness become much more visible.
Understanding where different parts of the workload run most efficiently is therefore important. Large training workloads and high volume inference can benefit from accelerators, while orchestration, data preparation, data services and some inference workloads can run efficiently on CPUs. Matching the workload to the right compute helps avoid using unnecessary resources where they are not needed.
Utilisation is the other major lesson. If GPUs are waiting for data, or networking and software become bottlenecks, infrastructure costs rise without producing more useful output. Testing under production conditions and understanding utilisation and workload costs before scaling gives enterprises a much clearer view of the infrastructure they actually need.
Last but not least, software also plays an important role in extending the value of AI infrastructure over time. As models, frameworks and inference techniques evolve, new runtimes, broader framework support and software optimisations can improve performance and enable new capabilities without requiring new hardware. This makes software an important part of the long-term economics of AI infrastructure.
Q. How are advancements in CPUs, GPUs, and heterogeneous computing architectures changing the performance-per-dollar equation for data centers, particularly for AI training and inference workloads?
The performance-per-dollar equation is increasingly defined at the system level, not by a single processor. As AI workloads diversify across training, inference, and agentic workloads, the most efficient infrastructure is one that assigns each task to the compute engine best suited to it.
GPUs are critical for large training workloads and high-volume inference, while CPUs handle orchestration, data preparation, data access and other tasks needed to keep accelerators productive. For smaller models and some enterprise inference workloads, CPUs can also be an efficient option. Memory capacity and bandwidth, networking and software are equally important because bottlenecks in any of these areas can reduce accelerator utilisation and increase the cost of running the workload.
This changes how performance for every dollar invested needs to be measured. Rather than focusing on peak processor performance, the more useful question is how much productive workload capacity a system or rack can deliver for each dollar and watt consumed. Right-sizing the architecture can improve utilisation, reduce infrastructure overhead, and deliver better overall economics.
That is where AMD’s end-to-end AI portfolio across AMD EPYC CPUs, AMD Instinct GPUs, AMD Pensando networking, and AMD ROCm open software can help customers optimise infrastructure around the workloads they actually run, while retaining flexibility as those workloads evolve.
Q. With power availability and energy efficiency becoming critical concerns, how can organisations align their AI ambitions with sustainability goals while maintaining strong economic returns on infrastructure investments?
What matters is how much useful work an AI infrastructure can deliver within its available power and cooling capacity. Better performance for every watt, higher utilisation, and efficient memory use can help organisations get more from the infrastructure they already have.
This means evaluating infrastructure at the systems and rack level. Higher compute density can reduce server count, rack footprint and associated power and cooling requirements, while efficient architectures can improve both energy consumption and costs. AMD takes a holistic approach to energy efficiency across silicon, systems, and software, with the goal of delivering more useful AI capacity within real power constraints.
Q. In future, what will distinguish the most successful AI-ready data centers in terms of architecture, economics, and operational efficiency, and what should Indian enterprises start doing today to prepare for that future?
The next generation of AI data centers will not be defined simply by how many accelerators they deploy. What will matter is how efficiently the complete infrastructure performs within real power, cooling, space and cost constraints.
Flexibility also matters because AI workloads will not stand still. Models will change, inference will take on a larger role and workloads will run across different environments. CPUs, GPUs, networking, memory and software therefore need to work as a balanced system that can adapt without requiring the infrastructure to be redesigned each time when business requirements evolve.
For Indian enterprises, it is important to look at the full range of workloads across the organisation, from general-purpose enterprise applications to AI workloads, and determine where each is best deployed. Depending on the workload’s requirements for performance, latency, data governance, scalability, and cost – the right model could be a mix of cloud, AI-focused neo clouds, on-premises infrastructure, the edge or increasingly capable AI PCs. It is also important to evaluate these environments at the system level, rather than relying only on individual component benchmarks, so organisations understand how the full infrastructure performs under real workloads and operating conditions.