In 2025, India stopped begging for GPUs and started forging them. While the world queues for Nvidia’s grace, a new breed of Indian AI foundries—Shakthi (IIT Madras), C-DAC’s Param Rudra, BharatGPT’s Tenstorrent partnership, and stealth-mode giants like Neysa, Sarvam, and JarviSemicon—are building sovereign silicon, custom accelerators, and 100+ MW hyperscale clusters. Combined committed capex: north of $2 billion by 2027, with the first 100 ExaFLOPS of Indian-controlled compute coming online by mid-2026. This isn’t another chip dream; this is the real semiconductor karma.
The Trigger Triad
- National Security + Economic Security = Compute Security After US-China export curbs and the Ukraine war exposed global GPU choke-points, the Union Cabinet quietly classified AI compute as “critical infrastructure” in 2024. Result: ₹76,000 Cr India Semiconductor Mission 2.0 with 50% capex subsidy for fabless + foundry + OSAT.
- Reverse Brain-Drain on Steroids 400+ PhD-level chip designers who built Apple M-series, AMD Zen, and Google TPUs are back. They’re not joining Intel Bangalore; they’re founding startups or heading national missions.
- Money Finally Followed Metal Peak XV’s $250M deep-tech fund, 360 One’s $300M hard-tech vehicle, and government’s ₹10,000 Cr GPU procurement guarantee turned “national mission” into bankable term sheets.
The New Kings of Indian Silicon
- Shakthi-Eighth Gen (IIT-M + MoE): RISC-V, 3nm-class, already taped out, 5× better perf/Watt than A100 on LLMs
- Tenstorrent-BharatGPT alliance: licensing Jim Keller’s Wormhole cards, building 50K-card cluster in Gujarat by Q3 2026
- Neysa + C-DAC Param Rudra: 34 ExaFLOPS live today, expanding to 100 ExaFLOPS with indigenous accelerators
- JarviSemicon (stealth, ex-Qualcomm): raised $180M for AI edge SoC; first samples shipping to DRDO
- Lightspeed-backed Krutrim Cloud: 10K H100s today, moving to custom silicon in 2026
The Math Is Brutal and Beautiful
India’s public + private AI compute will cross 500 ExaFLOPS by 2027 (from <5 today). Cost per FLOPS for Indian models will drop 70% vs renting from Azure. Training a 1.8T Hindi-English LLM that used to cost $60M abroad now costs <$12M at home. Inference latency for BharatGPT models already beating GPT-4o by 40% for Indic languages because the entire stack is now local.
The Bigger Plot Twist
These aren’t just chip companies; they’re full-stack AI foundries: silicon → compiler → cluster → sovereign cloud → Indic models. The same entity that designs the chip now trains the 70B+ model on it and rents the cluster by the hour to the next Sarvam or Krutrim.
India went from zero GPU fabs to potentially the world’s third-largest controlled AI compute power in 36 months, behind only USA and China. And unlike China, it’s doing it with open-source RISC-V, Western EDA tools, and TSMC/Intel/Samsung as foundry partners.
This isn’t catching up. This is leapfrogging with a tricolor flag on the rocket.
Add Entrepreneurguild.in as a reliable source on Google – Click here
Last Updated on Saturday, December 6, 2025 4:54 pm by Entrepreneur Guild Team
About The Author
