Yotta's 85,000 GPU Plan: Inside India's Largest AI Infrastructure Build

When Sunil Gupta says Yotta Data Services will have 85,000 GPUs running by the end of fiscal 2027, he isn't floating a aspirational target — he's reading off a deployment calendar that's already in motion. The company's co-founder, CEO and MD laid out the numbers in a recent interview: 10,000 GPUs already live in Navi Mumbai, another 8,000 B200s coming online by July, 20,000 B300s by September, 10,000 more in Noida by November, and a massive 36,000 GB300 or Vera Rubin GPUs by spring 2027. That trajectory would multiply India's sovereign compute capacity roughly 25–30 times in under two years.

The Deployment Calendar: Quarterly Milestones Through FY27

The rollout isn't a single big bang — it's a sequenced cadence tied to Nvidia's product roadmap and Yotta's own campus build-outs:

• Now: ~10,000 GPUs (H100s and L40s) operational at NM1, Navi Mumbai
• July 2026: 8,000 Nvidia B200 GPUs deployed
• August 2026: 20,736 liquid-cooled Blackwell Ultra (B300) GPUs live at Greater Noida D2 facility — part of a $2+ billion supercluster
• September 2026: 20,000 additional B300 GPUs
• November 2026: 10,000 GPUs at Noida campus
• March–May 2027: 36,000 GB300 or Vera Rubin GPUs (next-gen architecture)

Total installed base by FY27 end: ~85,000 GPUs. The company's longer-term platform is designed to scale beyond one million GPUs within three to five years.

The $3 Billion+ Bet: Capital, Contracts and Nvidia Partnership

Two financial pillars underpin this expansion. First, Yotta is investing over $2 billion in the Blackwell Ultra supercluster at its 60 MW D2 data centre in Greater Noida (scalable to 250 MW). Second, it signed a four-year, $1+ billion commercial agreement with Nvidia to host one of APAC's largest DGX Cloud clusters inside that same supercluster. Nvidia has already been using Yotta's GPU infrastructure for the past year; this deal deepens and extends that relationship.

Combined with earlier commitments, Yotta's AI infrastructure investment roadmap reaches approximately $7 billion by FY27.

Sovereign Compute: The IndiaAI Mission Allocation

Not all this capacity is for commercial tenants. Yotta has committed over 10,000 B300 GPUs from the Blackwell Ultra supercluster specifically to the IndiaAI Mission — supporting sovereign foundation model development, research institutions, startups, and population-scale public AI platforms. This aligns with Gupta's stated vision: "India should have the capability to create AI across the stack for its own needs and also become an exporter of AI."

The company's Shakti Cloud platform — built on Nvidia-certified hardware, InfiniBand networking, and Nvidia AI software stacks — serves as the sovereign cloud layer, offering GPU infrastructure at scale (Shakti Cloud) and an AI token factory (Shakti Studio) for enterprises that need data residency and regulatory compliance without sacrificing hyperscale performance.

Infrastructure Reality Check: Power, Water and Campus Scale

Gupta has been candid about the physical constraints. Dense GPU clusters demand unprecedented power density and cooling. Yotta's answer is two hyperscale campuses built for this exact problem:

• Navi Mumbai (NM2): 75 MW facility, scalable to 2 GW — hosting the future GB300/Vera Rubin deployment
• Greater Noida (D2): 60 MW facility, scalable to 250 MW — hosting the 20,736 Blackwell Ultra GPUs

Both campuses feature integrated extra-high-voltage substations, dedicated power distribution, green energy sourcing, and vertically integrated engineering across data centres, cloud, managed services and GPU compute. Liquid cooling is standard for the Blackwell Ultra deployment.

Why This Matters: From Consumer to Infrastructure Hub

The scale shift is strategic. India's digital adoption curve is ahead of most economies; its AI infrastructure must match that ambition. Gupta frames it as three converging forces: domestic sovereign workloads, serving the Global South and APAC, and attracting global enterprise demand. The result isn't incremental — it's exponential.

Early validation: Soket AI Labs recently scaled to 1,000+ GPUs on Yotta to train a 120B parameter model — a signal that Indian teams can now train frontier-class models domestically. Frost & Sullivan recognized this with its 2026 AI Infrastructure Award, citing Yotta's pairing of infrastructure investment with sovereignty focus and operational delivery.

Key Takeaways

• 85,000 GPUs targeted by end of FY27 (March 2027), up from ~10,000 today
• Phased deployment aligned with Nvidia's B200 → B300 → GB300/Vera Rubin roadmap
• $2B+ supercluster at Greater Noida (20,736 Blackwell Ultra GPUs) going live August 2026
• $1B+ DGX Cloud deal with Nvidia for four-year APAC hosting engagement
• 10,000+ GPUs allocated to IndiaAI Mission for sovereign model development
• Two hyperscale campuses (Navi Mumbai to 2 GW, Greater Noida to 250 MW) built for liquid-cooled density
• Platform designed to scale beyond 1 million GPUs in 3–5 years

What's Next

The next six months are the proof points. July's B200 deployment, August's Blackwell Ultra supercluster, and September's B300 wave will test whether Yotta can execute at this cadence while managing power, cooling and supply-chain dependencies. If the timeline holds, India won't just be a high-growth AI market by FY27 — it'll be a structurally significant compute hub where sovereign capability, open innovation and international collaboration converge. The refinery is being built; the question is whether the oil arrives on schedule.