India AI Impact Summit 2026

building-ai-readiness-from-compute-to-capability

7 min read

TRACK:Democratizing AI Resources
SPEAKERS
  • Dr. Paneerselvam M, Ministry of Electronics & IT
  • Thomas Zacharia, AMD
  • Timothy Robson, AMD
SESSION OVERVIEW

The proliferation of Generative AI, driven by the advent of Large Language Models (LLMs), Visual Language Models (VLMs), and Agentic workflows, has fundamentally transformed the requirements for high-performance computing infrastructure. For the users of the IndiaAI compute portal and the broader ecosystem, the selection of Graphics Processing Units (GPUs) has evolved from a simple metric of peak floating-point performance (FLOPS) into a complex optimization challenge involving memory bandwidth, interconnect topology, and architectural specialization.

VIDEO RECORDING

Building AI Readiness: From Compute to Capability

Abstract

The session explored what “AI readiness” means for a nation and its ecosystem, moving from raw compute resources to end‑to‑end AI capability. Dr. Paneerselvam set the stage by congratulating the IndiaAI community and emphasizing the need for both technical and organizational preparedness. Thomas Zacharia described the U.S. Department of Energy’s “Genesis Mission” and the forthcoming American Science Cloud that will run on AMD’s MI‑355 exascale hardware, illustrating how sovereign AI programs are being built around high‑performance compute, data federation, and secure governance. Timothy Robson followed with AMD’s historical leadership in supercomputing, highlighted the Helios rack (72 GPUs delivering 2.9 exaflops @ 220 kW), and explained how open‑source software stacks (PyTorch, Triton, SG Lang) enable rapid, vendor‑agnostic AI development. Dr. Paneerselvam then shifted to the business side, stressing change‑management, AI‑readiness quotients for SMEs, and the critical role of startups in delivering value. Thomas added a language‑diversity viewpoint, describing day‑zero support for Indian‑language LLMs and the need for an open ecosystem. Finally, Gil Garcia presented the emerging “physical AI” domain—edge‑centric robotics and industrial systems that require lightweight, dedicated accelerators rather than bulk GPUs. Throughout, speakers underscored open standards, talent development, and strategic partnerships as the pillars of a nation’s AI future.

Detailed Summary

  • Dr. Paneerselvam M opened by congratulating the 30,000 AMBers worldwide (including ~10,000 in India) and thanked the audience for the opportunity to discuss AI readiness.
  • He emphasized that AI is far broader than GPUs; GPUs are a critical component, but true readiness spans talent, data, governance, and societal change.
  • He introduced the three‑speaker format: he would discuss the sovereign side, while Tim (Timothy Robson) would cover the enterprise side.

2. Sovereign AI & the U.S. DOE “Genesis Mission”

2.1. Genesis Mission Overview (Thomas Zacharia)

  • The slide shown originated from the U.S. Department of Energy (DOE) under the Trump administration’s “Genesis Mission.”
  • Funding Context: The United States spends ~$1 trillion / year on R&D, with 20‑30 % coming from the federal budget; the rest is industry‑driven.
  • Goal: Increase research output efficiency—the gap between money spent and scientific breakthroughs is widening. AI is positioned as a multiplier to accelerate discovery, reduce cost, and improve global collaboration.

2.2. DOE Structure & Sovereign Priorities

  • DOE operates 17 national labs (e.g., Oak Ridge National Laboratory, where Dr. Zacharia once led).
  • Three national priorities:
    1. Discovery Science
    2. Energy
    3. National Security (with the unique civilian control of the nuclear arsenal).
  • The mission’s hypothesis‑experiment‑analysis loop: hypothesis → experiment → data → AI‑assisted analysis → faster insight.

2.3. International Expansion & Multi‑Sector Impact

  • Anticipated multinational participation (Japan, Europe, UK) to create a global sovereign AI framework.
  • Down‑stream domains: healthcare, education, skilling—all government‑function‑driven.
  • Emphasis on federated compute & data across heterogeneous assets (light‑sources, neutron sources, large‑scale labs).

2.4. Public‑Private Partnership & “American Science Cloud”

  • The initiative was announced by Secretary Wright (U.S. DOE) alongside AMD CEO Dr. Lisa Su.
  • The American Science Cloud will initially run on an AMD MI‑355 cluster, forming the backbone for AI‑driven scientific innovation.

3. AMD’s Historical Leadership & Future Compute Roadmap

3.1. Exascale Heritage

  • Thomas recounted his 30‑year career in supercomputing, noting four systems that topped the TOP500 list, each a “first‑of‑its‑kind.”
  • Key lesson: Government risk‑taking (e.g., early NVIDIA GPU deployments) paved the way for today’s AI boom.

3.2. The Helios Rack Demonstration

  • Tim introduced the Helios rack on display:
    • 72 GPUs delivering 2.9 exaflops (FP4) for AI workloads.
    • Power envelope: 220 kW – an order‑of‑magnitude reduction versus earlier designs that would have required 3–4 GW.
    • Demonstrates exascale‑class efficiency and the feasibility of scaling to Yotta‑scale (1 yotta‑flop) within a decade by stacking many racks.

3.3. Scaling Vision

  • Current state: 10⁹ users of generative AI today; projected 5 × 10⁹ in a few years.
  • Roadmap:
    • Zetta‑scale (10³ exaflops) attainable with ≈300 racks.
    • 10 000‑fold increase in compute capacity expected within 10 years.

4. Governance, Open Ecosystem & Software Stack

4.1. Governance vs. Regulation

  • Tim distinguished governance (human‑in‑the‑loop, peer‑review, safety checks) from regulation (formal laws).
  • Analogy: Professors overseeing student publications—human oversight ensures responsible AI output.

4.2. Open‑Source Foundations

  • AMD’s commitment to open standards:
    • PyTorch, Triton, SG Lang – all vendor‑agnostic, enabling day‑zero support for new AI models.
    • Day‑zero support meaning that when a new model (e.g., DeepSeek) is released, AMD’s stack can run it out‑of‑the‑box with optimized performance.

4.3. Multilingual LLM Challenges

  • Thomas highlighted the language‑diversity problem: many Indian languages have < 5 million speakers (e.g., Dogri, Sindhi, Nepali).
  • AMD’s experience with Finland’s “Lumi” supercomputer (a Uralic language) demonstrated a workflow for low‑resource language model training.

4.4. Tooling for Startups & SMEs

  • AMD Developer Cloud: 50‑100 hours free GPU compute, pre‑packaged Docker containers to simplify setup.
  • Accelerator Cloud: Transition path from proof‑of‑concept (PoC) to production‑ready (PoP).
  • Day‑zero model support for popular frameworks (e.g., Baidu’s Paddle, Meta’s LLaMA, DeepSeek).

5. Business‑Readiness & Change Management (Dr. Paneerselvam M)

5.1. From Technology to Business Value

  • AI is not a plug‑and‑play product; it demands process redesign, skill uplift, and organizational change.
  • SMEs/SMBs (owner‑managed family businesses) need a change‑management lens to adopt AI successfully.

5.2. The “AI‑Readiness Quotient”

  • Proposed a measurement framework:
    • Intent – strategic willingness to adopt AI.
    • Investment – financial, talent, and infrastructure commitment.
    • Implementation – execution capability (consultants, pilot programs, rollout).
  • “What you don’t measure, you don’t manage.” – a call to develop a readiness scorecard for Indian enterprises.

5.3. Role of Startups & Consulting

  • Startups act as AI‑native providers that can demonstrate ROI to traditional businesses.
  • Advisory firms must supply domain expertise, talent pipelines, and implementation support (e.g., turnkey PoC‑to‑Prod services).

5.4. Data & Intelligence Layers

  • Four‑layer model for AI adoption:

    1. Data Layer – ingest, cleanse, unify siloed datasets.
    2. Intelligence Layer – model training, insight extraction.
    3. Interface Layer – agentic AI, human‑in‑the‑loop UI.
    4. Compute Layer – underlying hardware (GPUs, TPUs, accelerators).

  • Emphasis on human involvement at each layer to avoid “black‑box” pitfalls.

6. Physical AI & Edge Computing (Gil Garcia)

6‑1. From Cloud‑Centric to Edge‑Centric AI

  • Physical AI = AI that perceives, decides, and acts at the edge (robots, autonomous vehicles, industrial plants).
  • Cloud latency is unacceptable for real‑time safety‑critical tasks; therefore dedicated, low‑power accelerators are required.

6‑2. AMD’s Edge Portfolio

  • AMD offers compact inference accelerators (e.g., MI‑450x variants) suitable for robotic vision, tactile sensing, and low‑latency actuation.
  • Case study: Gene 01 – a humanoid robot built on AMD tech (Italian startup) showcasing visual, tactile, and motor integration without relying on external GPUs.

6‑3. Open‑Source Stack for Edge

  • The same open‑ecosystem tools (PyTorch, Triton, SG Lang) extend to edge devices, enabling vendor‑agnostic development and fast iteration.

6‑4. Opportunities for India

  • India’s large SME base, manufacturing sector, and medical‑device ecosystem present a fertile ground for physical AI deployments.
  • AMD’s small‑form‑factor accelerators can empower local startups to build edge‑first solutions without waiting for large‑scale cloud resources.

7. Closing & Call‑to‑Action

  • Tim invited attendees to visit the AMD booth to see the Helios rack, Neo‑cloud partners, and starter kits for AI development.
  • Dr. Paneerselvam thanked the audience, reiterated the importance of a shared path forward, and signaled upcoming government‑startup‑industry collaborations.

Key Takeaways

  • AI readiness is multidimensional – it requires compute, data, talent, governance, and change management.
  • The U.S. DOE Genesis Mission exemplifies a sovereign AI program: massive R&D spend, public‑private partnership, and a dedicated American Science Cloud built on AMD’s MI‑355 hardware.
  • AMD’s exascale heritage enables highly efficient AI compute (e.g., 2.9 exaflops @ 220 kW in a single rack), providing a path to Yotta‑ and Zetta‑scale systems within a decade.
  • Open‑source software stacks (PyTorch, Triton, SG Lang) deliver day‑zero support for new models, ensuring vendor‑agnostic, low‑cost AI development.
  • Multilingual LLM challenges are being tackled by leveraging AMD’s experience with low‑resource languages (e.g., Finland’s “Lumi” project).
  • Business adoption hinges on change‑management and a concrete AI‑Readiness Quotient (intent → investment → implementation).
  • Startups and consultancies are critical enablers, providing turnkey PoC‑to‑Prod pathways and demonstrating tangible ROI to SMEs.
  • Physical AI shifts the focus from cloud‑centric to edge‑centric, low‑latency inference, requiring compact accelerators and robust software stacks.
  • India’s ecosystem—large SME base, government labs, and a vibrant startup community—stands to benefit from open, modular AI infrastructure and public‑private collaborations.
  • Open ecosystems and open standards are the strategic foundation for scalable, inclusive, and sovereign AI across research, industry, and society.

See Also:

Graph View

Backlinks