India’s Stake in the US-China GPU Fight

Foreign Policy, Defence & Geopolitics technology-geopolitics semiconductors

India does not want China to dominate all five layers of AI infrastructure. But India also does not benefit from a world in which access to advanced chips is contingent on the policy preferences of a US administration that reversed its own H20 policy four times in three years.

Origin

Developed by Pranay Kotasthane in Anticipating the Unintended #342 (May 2026), drawing on Jensen Huang’s five-layer AI stack framework and the Nvidia-China export control debate. Accompanied by a Takshashila discussion document.

What it says

Jensen Huang describes AI as a five-layer stack: energy, chips, infrastructure, models, and applications. American export controls sacrificed the chips layer to protect the models layer, but China adapted across the other four layers — abundant energy, algorithmic innovation, and a large engineering workforce. The controls bought two to three years on manufacturing but no meaningful time on AI capability.

India’s position in this stack is distinct from both the US and China:

• Applications layer: strong ingredients. IT services sector could become the global delivery platform for enterprise AI.
• Chip design layer: every major fabless company runs large Global Capability Centres in India.
• Below those layers: dependent. Data centre capacity an order of magnitude smaller than the US or China. Energy constraints limit brute-force compute scaling. India imports advanced chips and will continue importing frontier models.

This consumer-integrator position creates a strategic calculus that neither the Jensen Huang position (sell to everyone) nor the Dario Amodei position (restrict everything) represents:

• Want: Chinese AI labs to keep releasing open-source models — they are inputs, not threats.
• Want: a competitive global model ecosystem that reduces any single country’s leverage over India’s AI access.
• Do not want: Huawei hardware in sensitive Indian infrastructure — firmware risks are genuine.
• Do not want: chip access contingent on a single administration’s shifting preferences.

Open-source models are separable from Chinese hardware. Indian firms can optimise Qwen’s weights without deploying Huawei Ascend chips.

Applied

Three tasks for India:

1. Articulate a position through Pax Silica — the US-led technology initiative India joined in early 2026 — rather than deferring to Washington’s framing.
2. Fund GPU-agnostic middleware — public investment in open alternatives to reduce CUDA lock-in, as a priority for the India AI Mission.
3. Maintain model pluralism — use American proprietary models where best, Chinese open-source models where useful, and sovereign models like Sarvam for applications where dependency is unacceptable.

When it falls short

The framework assumes the hardware-model distinction remains clean. If future AI models require tight hardware-software co-optimisation (custom silicon for specific architectures), the separability argument weakens. The framework also does not account for the possibility that China might restrict its own open-source models — a scenario analysed separately.

Related frameworks

• [When Will China Block Its Open-Weight Models] — when and why China might restrict the open-source models India currently benefits from.
• [India’s Leverage Assessment Framework] — how India can deploy leverage in technology negotiations.
• [Instruments of Technology Geopolitics] — the full toolkit of technology statecraft.
• [Decoupling Dynamics] — the broader US-China technology confrontation.

Originally explored in Anticipating the Unintended #342 on Anticipating the Unintended.