India AI Impact Summit 2026

building-resilient-sustainable-ai-infrastructure-for-people-planet-and-progress

9 min read

TRACK:Resilience, Innovation, and Efficiency
SPEAKERS
  • Ashish Aggarwal, Nasscom
  • Nicole Foster, AWS
  • Rishi Bal, BharatGen
  • Ritu Mehrotra, Shunya Labs
  • Sharad Agarwal, Sify Infinit Spaces Limited
  • Vrushali Gaud, Google, USA
SESSION OVERVIEW

AI at scale requires infrastructure that converts promise into real impact. This session explores shared compute, efficient architectures, and hybrid cloud-to-edge strategies, comparing frontier model training with deployment of right-sized models for citizens, SMEs, and public systems. Anchored in People, Planet, and Progress, we examine access, affordability, sustainability, and optimization in India as core design constraints. The discussion addresses what should be built nationally, what collaboration can unlock, and where geopolitical frictions may arise globally.

VIDEO RECORDING

Building Resilient, Sustainable AI Infrastructure for People, Planet and Progress

Abstract

The panel examined the full stack of AI‑driven infrastructure needed to turn AI promise into real impact in India. Starting from the current state of data‑center capacity, the discussion moved through resilience, sustainability, security, and sovereign AI models, before looking ahead to demand forecasts, policy levers, and sector‑specific use‑cases such as voice‑enabled services for rural users. Throughout, the speakers stressed the need for coordinated capital, government support, clean‑energy integration, and trust‑by‑design to build an AI ecosystem that serves people, protects the planet, and drives progress.

Detailed Summary

Ashish Aggarwal opened the session by stating that AI infrastructure must serve two complementary goals:

  1. Enabling AI for societal good – AI should be the engine that delivers public‑benefit applications.
  2. Using AI to make the infrastructure itself more resilient and sustainable – AI can optimise power use, predict failures, and improve operational efficiency.

He highlighted that design choices (capacity, location, energy source) are inevitably shaped by the outcomes we expect—availability, security, and climate impact.

“When we think about resilience and sustainability, those topics are shaped by our constraints and the choices we make.”

He then introduced the panel, noting that it brings together viewpoints from the data‑center layer (Sify), clean‑energy operations (Google), cloud policy (AWS), voice‑AI innovation (Shunya Labs), and sovereign‑model research (BharatGen).

2. Data‑Center Landscape in India – Capacity, Growth & Policy Support

2.1 Historical Context & Current Capacity

Sharad Agarwal traced the evolution of India’s data‑center ecosystem:

MilestoneDetail
2015‑16First hyperscaler entries; most Indian startups relied on overseas infrastructure.
2015A leading hyperscaler generated >$200 M in revenue from India without any local data‑center.
End‑2024India’s data‑center power capacity ≈ 1.3–1.5 GW (≈ 3–5 % of US capacity).
In‑development3–3.5 GW under construction, expected to go live in the next 2‑3 years.

Sharad warned that even after the projected 10 GW by 2030, India would still lag behind the US, but the trajectory is steep.

2.2 Drivers of Expansion

  1. Capital Availability – No shortage of financing; data‑centers have high‑credit‑worthy customers (banks, telcos).
  2. Government Backing – Recent union‑budget measures removed a key tax hurdle, and dedicated AI‑summit events underscore policy focus.
  3. Power Infrastructure – India enjoys surplus generation capacity and a transmission grid that can route power quickly, unlike the US where utilities need 3‑5 years to build new supply.

“The ecosystem is fully in place; we just need to execute together.”

2.3 Resilience & Incident‑Response Rating

When asked to rate the Indian data‑center ecosystem on resilience, Sharad emphasized outcomes:

  • Availability – Historically zero‑downtime incidents at the national level.
  • Physical & Cybersecurity – No recorded major breaches; engineering talent ensures robust design.
  • Safety – No serious safety incidents despite high power density (“electrons per mm²”).

He illustrated that any loss of redundancy immediately triggers client anxiety, underscoring the high bar for reliability.

3. Sustainability at Scale – Google’s Decarbonisation Playbook

Vrushali Gaud (Google) positioned sustainability as a foundational pillar rather than a CSR add‑on.

3.1 Scope of Google’s Decarbonisation

  • All‑operations & value‑chain – From raw material sourcing, construction, data‑center operation, to the software layer (green‑software).
  • Design Trade‑offs – Different geographies demand different cooling strategies (air‑cool vs water‑cool).

3.2 Clean‑Energy Milestones

  • 22 GW of clean‑energy contracts secured globally (a figure that includes power‑purchase agreements and renewable‑energy‑certificates).
  • Early PPA (Power Purchase Agreement) development – Google helped pioneer PPAs before the market matured.

3.3 Innovation & Market Pull

  • “Tech companies are big off‑takers; their early‑stage pilots drive the market for renewables.”
  • Emphasis on hyper‑local design to optimise for water‑scarce vs water‑rich sites, a critical factor for Indian locations.

3.4 Future Outlook for India

  • Anticipates more regional clean‑energy projects to feed upcoming data‑centers.
  • Recognises the necessity of grid‑storage (batteries, long‑duration hydro) to smooth solar/wind intermittency.

4. Trust, Security & Sovereignty – AWS Perspective

Nicole Foster framed trust as the single most decisive factor for AI‑infrastructure adoption.

4.1 Security‑by‑Design Evolution

  • Early Amazon experience (online credit‑card security) taught the company to bake security into every service.
  • “Secure by design, sovereign by design.” Customers choose the region where their data resides (India, Canada, etc.) at account creation.

4.2 Data Access Model

  • Zero‑visibility – AWS staff cannot view customer data; this builds confidence.

4.3 Emerging Threat Landscape

  • AI‑enabled attacks and quantum‑computing risks – AWS is deploying AI to detect and respond to threats in real time.
  • Automation of identity and access – Reduces human error and improves response speed.

4.4 Areas Needing Continued Improvement

  1. Rapid evolution of threat vectors – Constantly updating detection models.
  2. Balancing speed of deployment with security controls – Especially as AI workloads demand low latency.

5. “Voice Infrastructure” – Enabling Frugal, Multilingual AI

Ritu Mehrotra introduced the concept of voice infrastructure as a three‑layer stack:

  1. Speech‑to‑Text (Understanding) – Capturing highly nuanced local dialects (e.g., Bhojpuri).
  2. Contextual Processing – Maintaining intent and domain knowledge across translations.
  3. Text‑to‑Speech (Response Generation)

5.1 Current Gap

  • Existing pipelines translate local speech to English, then back to the local language, losing context and increasing latency.

5.2 Use‑Case: Anganwadi Workers

  • 800 k Anganwadi workers serve 1.4 M centres.
  • A naïve GPU‑only solution would consume hundreds of millions of GPU‑hours.
  • Shunya Labs aims for 20× cheaper models that run on CPU‑level hardware, preserving context and enabling real‑time, on‑device interaction.

5.3 Broader Impact

  • Financial inclusion – Voice‑driven services for low‑literacy populations.
  • Healthcare – Early disease detection through localized conversational agents.

6. Sovereign AI Models & Public‑Grievance Automation

Prof. Ravi Kiran (BharatGen) clarified that technology itself is not unique, but the contextual deployment is.

6.1 Goal: Indian‑Sovereign Models

  • Build models that are owned, curated, and operated within India for Indian data and use‑cases.

6.2 Real‑World Application – Grievance Portal

  • Citizens submit grievances via audio, text, or images to a government portal.
  • Manual triage is non‑scalable; AI can automatically route, summarize, and suggest remediation.

6.3 Distinctiveness

  • While the algorithms are globally available, local language, policy, and data‑privacy constraints make Indian‑centric solutions essential.

7. Audience Q&A – Demand Forecast, Policy Wish & Capacity Outlook

Question (Sudipto Banerji, audience) – “How realistic is the 10 GW target by 2030? What share is driven by hyperscalers? If you could grant one policy wish, what would it be?”

7.1 Sharad’s Quantitative Assessment

  • Historical CAGR – 28 % YoY growth from 0.4 GW (2020) → 1.3 GW (2024).
  • If the same CAGR holds (plus a modest 7 % acceleration for cloud‑adoption), capacity could reach ≈ 6 GW by 2030.
  • Adding AI‑specific inference workloads and regional demand (Southeast Asia) can plausibly push the figure toward 10 GW.

7.2 Demand Composition

  • Domestic demand ~ 60 % (enterprise, government, telecom).
  • Hyperscaler‑driven demand ~ 40 % (AWS, Google, Azure, etc.).

7.3 Policy Wish – Single‑Window Clearance

  • Current project timeline: ~ 4 years (land → approvals → commissioning).
  • Wish: A true “single‑window” mechanism that eliminates the two‑year approval gap, cutting total time‑to‑operation to ≈ 2 years.

“If we could zero‑out the approvals stage, we could double the speed at which capacity is added.”

8. Future Sustainability Innovations – Storage, Green Materials & Hyper‑Local Inference

Vrushali Gaud revisited sustainability with a three‑pronged view:

  1. Energy Storage – Scaling battery and long‑duration hydro to complement India’s abundant solar generation.
  2. Green Construction – Adoption of green steel and green cement to reduce embodied carbon in data‑center builds.
  3. Hyper‑Local Inference – Deploying smaller, efficient models at the edge to cut energy use for billions of users.

She noted ongoing pilot projects in Bikaner (large solar‑farm + high‑frequency grid) as a template for rapid, clean‑energy rollout.

9. Trust & Efficiency in Frugal AI – Ritu’s Follow‑Up

Ritu reiterated two critical success factors for India’s “frugal AI” movement:

  1. Trust – First‑time failures erode confidence by ≈ 22 %; repeated errors cause abandonment.
  2. Efficiency – Example: a global firm spent 680 k GPU‑hours to train a model; an Indian startup achieved comparable performance in ≈ 40 GPU‑hours with a much smaller data set.

She highlighted banking/finance as the sector likely to see the quickest rollout, because voice‑driven interfaces can reach low‑literacy users who already own smartphones but cannot read English.

10. Cloud Sovereignty – Nicole’s Strategic Outlook

Nicole Foster framed sovereignty as a layered decision:

Stack LayerCore ConsiderationLikely Trade‑off
Training infrastructureWho owns the hardware & dataEasier to localise, but limited to large players
Platform services (mid‑layer)Open‑source vs vendor‑locked models; data‑residency controlsBalances innovation access vs regulatory compliance
Application layerEnd‑user experience, latency, complianceMay require hybrid architectures (edge + cloud)

She stressed that cloud providers still add unique value (pre‑built AI services, security frameworks) that cannot be replicated by merely building local data‑centers.

11. Vision of Success & Risks – Professor Kiran’s Closing Remarks

Success “tiers” for Indian AI ecosystem (as envisioned by Prof. Kiran):

  1. Broad Success – A “UPI‑style” universal platform that democratizes AI access across industries.
  2. Mass‑Adoption Success – Analogous to the Nokia era: a universally adopted, resilient product (e.g., a standard AI‑enabled voice assistant).
  3. Hyper‑Local Success – AI reaching the “last mile”, delivering context‑aware services to remote villages.

Key Risks

  • Data silos – Prevent cross‑sector innovation; governance mechanisms must enable safe sharing.
  • Talent shortage – Need for high‑quality, specialized talent; teaching and up‑skilling are essential.

Key Takeaways

  • India’s data‑center capacity is still modest (≈ 1.5 GW) but on a rapid growth trajectory; 10 GW by 2030 is ambitious but plausible if current CAGR and AI‑driven demand continue.
  • Government policy (tax incentives, single‑window clearances) is a critical accelerator; the panel’s top wish is a streamlined approval process.
  • Resilience is already strong (near‑zero downtime, solid physical security), but redundancy expectations remain high among enterprise customers.
  • Google has secured 22 GW of clean‑energy contracts and emphasizes locality‑specific cooling designs to meet Indian climate constraints.
  • AWS builds trust through “secure‑by‑design” and “sovereign‑by‑design” services, offering regional data residency and zero‑visibility of customer data.
  • Voice infrastructure must preserve context across language translation; on‑device, low‑resource models are essential for scalability (ex: Anganwadi workers).
  • Sovereign AI models are not technically unique, but local governance, language, and data‑privacy make Indian‑centric solutions indispensable (e.g., automated grievance portal).
  • Efficiency gains (orders of magnitude reduction in GPU‑hours) can democratise AI and unlock use‑cases in finance, healthcare, and public services.
  • Sustainability hinges on storage (batteries, hydro), green construction materials, and hyper‑local inference to keep energy use low as AI scales.
  • Success will be measured in three layers – universal platform (UPI‑like), mass‑adoption product (Nokia‑like), and last‑mile contextual AI; data silos and talent gaps are the biggest threats.

See Also:

Graph View

Backlinks