India AI Brief Issue #8 31 May 2026 Deep Cut + Briefing

India built the cheapest AI compute
on Earth. Who gets to use it?

The IndiaAI Mission has 38,000 GPUs running at ₹67/hour — a third of global cloud rates. Four startups got a 100% subsidy. Over 500 applied. The infrastructure is real. The access problem nobody is talking about is equally real.

₹67
GPU-hour on
IndiaAI Mission
38,000
GPUs now
available
4
Startups got
100% subsidy
500+
Proposals
submitted
The Deep Cut · India AI Infrastructure

India just built the cheapest AI compute on Earth.
Here’s the access problem nobody is talking about.

The number that matters most this week is not a valuation or a funding round. It is ₹67.

That is the price per GPU-hour on IndiaAI Mission’s compute cluster — approximately one-third of what AWS and Azure charge, and a fraction of the $2–4 US rate. India now has over 38,000 GPUs including Nvidia H100 and H200 units running at subsidised rates of roughly ₹67 per hour. For context: training a mid-sized language model on US cloud infrastructure costs tens of lakhs of rupees. On IndiaAI compute, the same job costs a fraction of that.

This is genuinely significant. Not significant in the way government press releases are significant. Significant in the way that removing one foundational barrier to entry actually changes who gets to build things.

The question nobody is asking publicly: who is actually inside this system?

The four who got in — and what they received

MeitY is backing Sarvam AI, SoketAI, Gan AI, and Gnani AI with a 100% compute subsidy plus GPU access at roughly ₹67 per GPU-hour. Sarvam pulled the largest single allocation: 4,096 NVIDIA H100 SXM GPUs from Yotta Data Services, against a ₹246.71 crore project.

Read that again. 4,096 H100 GPUs. Free. At US cloud rates, that allocation would cost approximately ₹150 crore per month. Sarvam gets it for ₹246.71 crore over the entire project duration — a multi-year sovereign subsidy of extraordinary scale.

This is not a criticism. Sarvam has shipped real products: Sarvam-30B, Sarvam-105B (open-source, Apache 2.0), Bulbul voice, Saaras transcription. The allocation is earned. The point is narrower: four companies got the meaningful subsidy. India has over two lakh startups. The math does not complete.

ℹ The infrastructure by the numbers

₹10,371.92 crore — total IndiaAI Mission budget over five years. ₹4,563.36 crore earmarked specifically for compute. 38,000+ GPUs across 14 empanelled service providers including Yotta, E2E Networks, and Tata Communications. Target: 100,000 GPUs by end of 2026. All accessible via the IndiaAI Compute Portal at indiaai.gov.in.

The access problem — the part the press releases omit

The IndiaAI Mission received over 500 proposals in initial phases. Twelve startups were selected across two rounds. The subsidised rates require registration and validation — you cannot simply spin up instances like AWS. For time-sensitive projects with investor deadlines, the onboarding friction might be a dealbreaker.

Route Who can access Speed Cost
IndiaAI Mission (100% subsidy) 4 selected startups Months ₹0
IndiaAI Compute Portal (subsidised) Approved startups, researchers, MSMEs Weeks–months ₹67–116/hr
Commercial cloud (AWS/Azure/GCP) Anyone with a credit card Minutes ₹300–600/hr
US startup cloud credits Startups (AWS/Google programmes) 48 hours Free ($100K+)

The US solved this with venture capital and cloud credits. AWS, Google, and Azure all run startup credit programmes that hand out $100,000–500,000 in cloud credits in 48 hours. India is solving it with a government mission that takes months. Both approaches work. Only one works at startup speed.

“India built the cheapest AI infrastructure in the world. Four companies are fully inside it. The rest of India is reading a form.”

The talent gravity well — what this means for your hiring

The 100% compute subsidy and ₹67 per GPU-hour pricing will pull more senior ML talent into Sarvam AI, SoketAI, Gan AI, and Gnani AI. As a result, expect tighter senior-role markets in Bengaluru, Hyderabad, and Pune through 2026 and 2027. Compensation bands will climb.

NASSCOM projects India’s AI talent pool to grow from 6–6.5 lakh professionals today to over 12.5 lakh by 2027 at 15% CAGR. The supply is coming. But the IndiaAI Mission subsidy means the senior end of that pool is about to concentrate at four specific addresses. Every company that has been deferring its AI hiring strategy is now competing against Sarvam, SoketAI, Gnani, and Gan AI for the same pool of ML engineers. The salary benchmarks built six months ago are already stale.

One action per reader type

If you’re a founder
Apply to the IndiaAI Compute Portal now at indiaai.gov.in — even if you’re not ready. The queue is real and the backlog builds. The process is slow but not closed.
If you’re in HR or L&D
The four IndiaAI Mission startups will hire aggressively in 2026. Map which AI roles in your org overlap with the talent they want before salary benchmarks reset.
If you’re a CA, lawyer or consultant
Knowing Sarvam, Gnani, SoketAI by name before your clients do is your edge. These are the tools your clients will ask about in 2027 — trained on this compute, right now.
If you’re a freelancer or creator
The Hindi, Tamil and Telugu AI tools that are free today are being trained on this government compute. Sarvam’s startup programme and Bhashini’s API are the doors into this ecosystem — both free.
⚡ Affiliate partner

Make.com — Automate the work the AI freed you from

You now know AI can write, research and draft. The next step is automating the workflows around it — WhatsApp reminders, invoice follow-ups, client onboarding, GST filing triggers. Make.com connects 300+ apps without code. Free tier: 1,000 operations/month — enough for a full SME workflow.

Start Free →
Affiliate link — we earn a small commission if you upgrade, at no cost to you.
The Briefing · 5 things Indian professionals need to know

Five items. Each one has a specific India angle. Each one affects your work.

01
DPDP Act
The DPDP compliance clock is running. Most Indian businesses haven’t started.
India’s Digital Personal Data Protection Rules were finalised in November 2025. The hard enforcement date — when all substantive obligations become mandatory — is May 13, 2027. That is 12 months from this issue. Consent frameworks become operational in November 2026 — six months from now. The Data Protection Board is being constituted. Penalties go up to ₹250 crore per violation.

The DPDP Act is not GDPR-lite. It has its own consent, breach notification and data rights framework. The compliance gap between “we use some software” and “we are DPDP-compliant” is significant. Every CA, lawyer and HR professional whose clients process Indian personal data has exactly 12 months before that gap becomes a liability.

→ India angle: The professional advisory opportunity here is large. DPDP compliance for SMEs, clinics, HR departments is a billable service. The window for early-mover positioning closes as enforcement approaches.
02
AI Models
GPT-5.5 is now the default in ChatGPT — most Indian users didn’t notice.
OpenAI made GPT-5.5 Instant the default model in ChatGPT this week, quietly, across free and paid tiers. No announcement, no major press coverage. The implication: the free ChatGPT that Indian students, CAs and freelancers use every day just got meaningfully more capable without any action required.

If ChatGPT has felt faster and more accurate this week, it is. Prompts that previously failed or produced inconsistent outputs are now more reliable. Use cases you abandoned because ChatGPT “wasn’t good enough” six months ago are worth revisiting this week.

→ India angle: ChatGPT penetration in India is among the highest in Asia. Free-tier improvements matter more here than almost anywhere. No upgrade, no credit card, no action required — the tool you already use is better today.
03
Healthcare AI
ABDM adoption is accelerating — clinics not registered are being bypassed in digital referral workflows.
Ayushman Bharat Digital Mission registrations are accelerating across India. The Health Facility Registry (HFR), Health Professional Registry (HPR) and ABHA patient IDs are increasingly the prerequisite for electronic referral workflows. Clinics and hospitals that remain on paper-based or non-ABDM systems are finding themselves outside the referral networks of ABDM-compliant institutions.

The pattern: a referring hospital that has moved to ABHA-linked records cannot easily route patients to a clinic whose prescriptions and records don’t integrate. This is not a future risk — it is happening now across Tier 1 cities and spreading to Tier 2.

→ India angle: If you are a doctor or clinic owner — register on HFR at facility.ndhm.gov.in. Takes 20 minutes. It’s free. The window to do it before it affects referrals is narrowing. See our full Doctors guide for the complete ABDM walkthrough.
04
Talent & Hiring
AI talent in India is about to get expensive — fast. The salary benchmarks you have are stale.
NASSCOM projects India’s AI talent pool to grow from 6–6.5 lakh professionals to over 12.5 lakh by 2027 at 15% CAGR. Supply is growing. But the IndiaAI Mission’s 100% compute subsidy to four foundation-model startups means the senior ML engineer pool is about to concentrate at Sarvam, SoketAI, Gnani, and Gan AI — companies now operating with effectively zero infrastructure costs and a research environment that rivals Google DeepMind India.

Every company hiring AI roles in 2026 and 2027 will compete against these four for the same senior pool. Compensation bands in Bengaluru, Hyderabad, and Pune will rise faster than the overall market.

→ India angle: If your company has AI hiring plans for the next 18 months, the benchmarks you built in 2025 are wrong. Build a salary revision buffer of 20–30% for senior AI roles into your 2026–27 headcount planning now.
05
India-built AI
Sarvam’s startup programme is live — most Indian founders don’t know it exists.
In March 2026, Sarvam launched a programme offering API credits and developer tools to Indian startups building on its models. Sarvam-30B and Sarvam-105B — both Apache 2.0, both trained entirely in India — are available through a startup-tier API without the IndiaAI Mission registration process. Hindi, Tamil, Telugu, Kannada, and seven more Indian languages, with voice, text and translation capabilities.

If you are building anything that requires Indian language support and have been defaulting to Google Translate and ElevenLabs, Sarvam’s startup programme is worth 15 minutes of your time. The Bulbul voice model currently produces the most natural-sounding Hindi, Tamil, and Telugu TTS available from any provider globally.

→ India angle: Apply at sarvam.ai — free credits, no government registration required. Sarvam’s Vision OCR scored 84.3% on olmOCR-Bench, beating Gemini 3 Pro (80.2%) and ChatGPT (69.8%) on Indian language document understanding.
The Ask

One forward.
That’s all we ask.

If one item in this issue was useful — forward it to one person it would help. A founder who needs to know about IndiaAI compute. An HR colleague building a 2027 headcount plan. A doctor who hasn’t registered on ABDM yet. No referral points. No scheme. That single forward is how India AI Brief reaches the people who need it.

Pushpanjali Sharma
Founder, India AI Brief · LLMTools.in
India AI Brief · Every Saturday · Free
The AI briefing
India deserves.

Honest, India-specific AI intelligence every Saturday. No Silicon Valley hype. Real tools, INR pricing, what works for Indian professionals and what doesn’t.

✓ You’re in. Next issue lands this Saturday.
Free forever  ·  No spam  ·  Unsubscribe anytime
India AI Brief  ·  Free every Saturday
Subscribe →