When Jensen Huang unveiled NVIDIA's Vera Rubin platform at GTC, the AI infrastructure world quietly pivoted. The headline numbers - 50 PFLOPS of FP4 inference per GPU, 288 GB of HBM4 memory, 3.6 ExaFLOPS in a single NVL72 rack - are stunning enough on their own.

But the real story isn't the silicon. It's the operating envelope it demands.

Vera Rubin is the first NVIDIA platform where liquid cooling is no longer a recommendation. It is a hard prerequisite. No air-cooled configuration exists. None is planned. If a data center can't dissipate 150-200 kW of heat per rack through direct-to-chip (D2C) liquid loops, it simply cannot host this generation of GPUs.

That's a conversation Indian enterprises, AI labs, and government missions need to start having today - because the gap between "GPU-ready" and "Vera Rubin-ready" is wider than most colocation providers want to admit.

What Vera Rubin Actually Is

Vera Rubin, named after the astronomer whose observations confirmed the existence of dark matter, is NVIDIA's next-generation accelerated computing platform, succeeding the Blackwell (B200/B300) family. It pairs two new pieces of silicon:

• Rubin GPU - the compute accelerator, built on next-generation process technology with HBM4 memory
• Vera CPU - an 88-core NVIDIA-designed Arm-based host processor with 1.5 TB of LPDDR5x memory, linked to the GPU over NVLink-C2C at 1.8 TB/s

These come together in the NVL72 rack-scale system - a single 19-inch rack containing 72 Rubin GPUs and 36 Vera CPUs, wired together through NVLink 6 into what NVIDIA calls "one supercomputer per rack."

The specifications that matter

To put that into business terms: NVIDIA projects that a Vera Rubin NVL144 CPX deployment can generate up to $5 billion in token revenue per $100M of infrastructure investment. That is not a typo. That is the economic gravity now pulling every serious AI builder toward this generation.

Why Vera Rubin Breaks Traditional Data Centers

A conventional Indian colocation facility, even a recent one, is typically engineered for 5-10 kW per rack. Air-cooled. Hot-aisle/cold-aisle containment. CRAC and CRAH units pushing chilled air through perforated tiles.

That architecture cannot host Vera Rubin. It cannot host Blackwell GB300 NVL72 either. The physics simply do not work.

Here's why:

1. Heat density. Vera Rubin pushes 150-200 kW per rack. Air cannot move enough thermal energy fast enough at that density, you would need to blow air at speeds that vibrate the servers off the rails.

2. Power delivery. A 200 kW rack draws roughly 30× the current of a legacy rack. That demands new busbar designs, new branch-circuit metering, and N+1/2N redundancy engineered for far higher fault tolerances.

3. Coolant infrastructure. Direct-to-chip cooling requires CDUs (Coolant Distribution Units), secondary loops, manifold routing, leak detection, and flow/pressure/temperature telemetry at every rack, none of which can be bolted onto a legacy hall.

4. Floor loading. A fully loaded Vera Rubin NVL72 rack weighs significantly more than a traditional rack. Many older Indian DCs cannot structurally support it.

This is why "retrofitting" an existing data center for Vera Rubin is, in most cases, not retrofitting at all, it is demolition followed by purpose-built construction.

The Physics Dividend of Liquid Cooling

The reason liquid cooling matters isn't just heat - it's economics. The engineering numbers tell the story clearly:

• 23× more heat removal versus air, using liquid-to-liquid CDUs
• 25% of the power budget reclaimed - power that previously fed cooling fans now goes back to driving GPUs
• 2× GPU hardware lifespan - because liquid keeps junction temperatures 10-15°F lower than air, slowing electromigration and silicon aging

In a Vera Rubin deployment, that 25% reclaimed power budget alone is the difference between hosting 8 NVL72 racks and hosting 10 - a meaningful number when each rack delivers 3.6 ExaFLOPS.

A small but important note on the Indian context: a handful of domestic operators have been quietly building toward this density curve for the last couple of years. Cyfuture, for instance, has a 10 MW purpose-built liquid-cooled AI facility coming online in late 2026, engineered around the 150 kW/rack envelope and D2C loops that Vera Rubin will require - proof that the infrastructure side of the conversation is no longer hypothetical in India. That's the kind of build the rest of the market will need to catch up to.

What This Means for India's Data Center Builders

For most of the last decade, the Indian data center industry competed on relatively comfortable parameters, uptime, location, connectivity, certifications, and price per kilowatt at 5-10 kW racks. Vera Rubin invalidates that competitive frame almost entirely. The new questions are different, and harder.

• Can you deliver 200 kW to a single rack, reliably, with N+1 or 2N redundancy, and stand behind it with SLAs?
• Can your cooling loop dissipate that heat through a closed liquid path without a single drop of coolant ending up where it shouldn't?
• Can your structural floor carry it?
• Can your network fabric saturate ConnectX-9 NICs without becoming the bottleneck?
• Can your power and thermal headroom absorb the next generation after Vera Rubin; Rubin Ultra in 2027, at 600 kW per rack on the new Kyber form factor, without forcing tenants into a forklift upgrade?

For the vast majority of Indian operators, the honest answer to those questions today is no, or not yet. That isn't a criticism; it's an industry-wide reality. Liquid cooling at production scale, multi-MW phased build-outs around 150 kW racks, MeitY-empanelled SEZ facilities with the right power architecture, these are engineering bets that had to be placed two or three years ago to be ready for this moment. Most weren't.

This is where Cyfuture's role becomes interesting to watch. The company committed early to a 10 MW, purpose-built, liquid-cooled AI facility, engineered from the floor slab upward around the 200 kW/rack envelope, D2C cooling loops, 800G fabric, and the multi-OEM thermal validation regime that Vera Rubin-class deployments demand. It also chose to locate inside an SEZ, which gives Indian AI buyers a duty-free pathway for GPU imports and a meaningful CAPEX delta on multi-million-dollar build-outs. And it built modular phasing into the design, so the same hall that hosts a Vera Rubin NVL72 in 2026 can absorb a Rubin Ultra Kyber rack in 2027 without stranded investment.