NVIDIA's India Footprint: $4–5 Bn in Commitments Face a Post-Huang Test

Indian operators have committed more than $4–5 Bn to NVIDIA-based AI data centres, according to Counterpoint Research figures quoted by Moneycontrol. Jensen Huang's exit from NVIDIA therefore lands on an ecosystem that has already paid, or promised to pay, a serious amount of money for one company's silicon and software.

The hardware ledger is concrete. Yotta Data Services has committed to deploying more than 20,000 NVIDIA Blackwell Ultra GPUs. AM Intelligence, backed by Greenko's promoters, has announced orders for 29,000 next-generation Rubin GPUs. L&T's Vyoma is building a 10,000-GPU AI Factory on NVIDIA's B300 architecture. Neysa, ESDS and E2E Networks sit in the same procurement camp, though the arrangements differ — direct purchase in some cases, OEMs and systems integrators in others.

The dependence is now policy infrastructure, not just private capex. The IndiaAI Mission's common compute programme had onboarded over 45,000 GPUs by June 2026. By August it had approved 237 projects for subsidised compute, with 93.18 lakh GPU hours sanctioned. A disruption in advanced GPU supply would touch startups, researchers and government-backed projects alike.

How deep does the lock-in run?

The IndiaAI Compute Portal's empanelment table makes the concentration plain. It lists 44,077 compute units from 14 bidders:

• 35,985 NVIDIA units
• 5,548 Google Trillium TPUs
• 1,200 AWS Trainium and Inferentia units
• 1,096 AMD MI300X and MI325X units
• 248 Intel Gaudi units

Roughly four of every five sanctioned GPUs carry the NVIDIA mark. Hardware, however, is the shallower half of the moat. CUDA, with its surrounding stack of cuDNN, TensorRT, NeMo, NIM and the Nemotron open-model family, creates switching costs that a GPU swap cannot clear. Firms may need to rewrite software, re-optimise workloads and retrain engineers to leave. Sarvam AI, Gnani.ai, Chariot, CoRover.ai and Tech Mahindra already build on NeMo and Nemotron; TCS, Infosys and Wipro construct enterprise AI on the same stack. This reaches companies that own no GPU clusters at all.

Sanchit Vir Gogoi, founder of Greyhound Research, argues rivals must replicate more than hardware: they must reproduce the software ecosystem and the industry relationships. On diversification, his verdict is measured. "Yes, by workload, and with evidence," he said. "Inference is where portability is closest to ready, because serving software increasingly hides the accelerator from the application."

What about export controls?

Huang's departure has also exposed a geopolitical fault line. The Biden administration's 2025 AI Diffusion Rule placed India in a middle tier with import caps; the Trump administration suspended enforcement, but a May 2026 Government Accountability Office assessment found the formal rescission incomplete, notes Vipul Agrawal, partner at Clavius Legal.

"On the US Government's published position, India currently enjoys comparatively favourable treatment for buying advanced AI chips, but that is not a guarantee of supply," Agrawal said. "Suppliers may ask more about where chips are deployed, who the customers are and where the chips may go next, which could slow procurement."

The US Commerce Department is reportedly weighing tighter licensing for advanced chip shipments to seven countries, including Singapore, Malaysia and Thailand. India is not on that list. The compliance burden still grows. "Indian neoclouds and data centre operators must assess their obligations under the US Export Administration Regulations (EAR)," said Shashi Mathews, partner at CMS INDUSLAW. "This includes verifying the export classification of the GPUs, applicable licensing requirements, and the identity of the supplier, consignee, and end-user." He adds that giving foreign nationals or restricted entities access to controlled compute can legally function like exporting the physical item.

Can anyone compete?

Alternatives exist, at least at the margins. Bengaluru-based Turiyam is building a full-stack platform around inference workloads, CUDA-free middleware and RISC-V hardware. Ziroh Labs claims its Kompact AI runtime delivers NVIDIA A100-class inference on Intel Xeon processors. Neither is a substitute for the GPU ecosystem, and nobody claims otherwise.

The open-source question matters here too. NVIDIA has agreed to acquire Hugging Face, one of the world's largest open-model repositories, for approximately $13 Bn. Rahul Sai Poruri, CEO of the FOSS United Foundation, does not expect the strategy to change under new leadership. "Open models are likely to be optimised for whatever GPU is most widely available to developers. Currently, that is NVIDIA," he said. He estimates it could take half a decade or more for the ecosystem to meaningfully reduce its dependence.

What does the money say?

NVIDIA has moved beyond vendor status. Its Inception programme, free and equity-free, supports more than 4,000 Indian AI startups with training, tools, pricing and investor introductions. In August 2026, NVIDIA joined Sarvam AI's $74 Mn funding tranche, becoming the startup's first strategic chip-industry investor. In October 2026 it participated in Simplismart's $14 Mn Series B; the startup integrates with NVIDIA's Dynamo inference software.

Huang's personal diplomacy built much of this. His September 2023 visit produced collaborations with Reliance and Tata around AI infrastructure; a year later came partnerships with TCS, Tech Mahindra, Tata Communications and Zoho. Those relationships are institutional assets, and they do not evaporate with a change of corner office.

The open question for Indian operators is therefore not succession but supply. Anyone whose business assumes uninterrupted GPU access, on current terms, is making a wager that export policy has already tested elsewhere.