SUBSCRIBE FOR ACCESS

The New Problem: Your Buildings Now Age Slower than Your GPUs

You just spent the last 12–18 months planning, funding, and deploying Blackwell class systems. Now NVIDIA’s Vera Rubin NVL72 promises to:

• Train mixture-of-experts (MoE) models with roughly one-fourth the GPUs required on comparable Blackwell NVL72 configurations.
• Deliver up to 10x lower inference cost at the token level compared to Blackwell for MoE and similar large scale models.
• Achieve multi-exaflop rack scale performance with 72 Rubin GPUs, 36 Vera CPUs, HBM4, NVLink 6, and 100% direct-to-chip liquid cooling in a cable-free, blind-mate tray architecture.

For most enterprises, the immediate risk is not “we bought the wrong GPUs.” The real exposure is that server and GPU refresh cycles have historically been planned on 3-5 year timelines, while data center facilities are designed to last a decade or more. NVIDIA is now effectively updating its AI platforms on a roughly annual rhythm (Hopper, Blackwell, Blackwell Ultra, Rubin). Many existing facilities were never engineered for Rubin-class, 100% liquid cooled, high-density racks without significant changes to MEP systems.

Colovore was built specifically for this new cadence so you can fully monetize Blackwell while being ready to land Rubin-class racks, and whatever comes after, without turning every GPU generation into a construction project.

What Rubin Signals for Blackwell Owners and Their Facilities

Vera Rubin doesn’t just move a performance bar, it rewrites the economics of AI infrastructure design.

• Rubin NVL72 is targeted to deliver up to 5x better end-to-end performance and as much as 10x lower cost per token than Blackwell for key training and inference workloads, while using far fewer accelerators for MoE models.
• NVIDIA is pushing toward significantly higher performance per watt, leveraging HBM4, improved NVLink bandwidth, and rack-scale system integration to increase compute delivered per unit of power and space.

At the same time, NVIDIA has shifted to an annual or near annual cadence: Hopper → Blackwell → Blackwell Ultra → Vera Rubin, with Rubin family systems scaling beyond NVL72 and framed as liquid cooled “AI factories in a rack.”

For CIOs and CFOs, this creates three uncomfortable facts:

• The useful economic life of each GPU generation is compressing toward 12–18 months, even as facilities are still capitalized on 10–20 year timelines.
• Power density, cooling topology, and rack-scale integration assumptions change materially from one generation to the next, especially as Rubin standardizes on 100% warm-water, direct-to-chip liquid cooling around roughly 45°C supply temperatures.
• Facilities built tightly around a single generation’s assumptions face expensive, disruptive upgrades just to remain eligible for the next wave of hardware.

Rubin makes one thing clear: the first place you’ll strand serious capital isn’t in GPUs, it’s in data centers that can’t adapt their power, cooling, and layouts to each new generation.

How Colovore Future-Proofs Your Facility for NVIDIA’s Roadmap

Colovore’s core design allows multiple generations of AI hardware to coexist and turn over inside the same halls without rebuilding the shell, the backbone power, or the Facility Water Supply cooling loop.

Power Architectures that Evolve with Each Generation

Instead of rigid busway-based layouts tuned to a single density, Colovore data halls are wired like a programmable power fabric.

• 415/240V three-phase, five-wire power is available anywhere across the hall, with the ability to deliver on the order of 8 MW of fully protected, conditioned critical power to a given location or cluster of locations.
• Both AC and DC architectures are supported: 48VDC through 800VDC, via in rack rectification or rectifier sidecars, alongside 415-480VAC feeds, allowing Blackwell, Rubin, and future platforms with different voltage requirements to live in the same white space.

For technical leaders, this means power can be re-terminated and re-balanced as NVIDIA changes rack-scale designs, without having to re engineer busways or re permit major electrical work. For CFOs, it means a single backbone investment can support several GPU generations, instead of locking capital into a configuration that becomes obsolete midway through depreciation.

Cooling Designed for 100% Liquid Rubin Racks and What Comes After

Vera Rubin NVL72 is fully direct-to-chip liquid cooled, with a fanless, cable-free tray design and warm-water operating targets that support efficient, high-density deployments and reduced reliance on traditional chiller plants.

Colovore’s mechanical design assumes that future AI platforms will look more like Rubin than like legacy air cooled racks:

• Facility water loops can support multi-megawatt loads per row, with the ability to combine high-capacity CDUs (for example, multiple ~1.5 MW units) and 50 kW class active rear door heat exchangers within the same row—without modifying primary plant or mechanical yards.
• Within a single hall, Colovore can simultaneously support air cooling, rear door heat exchangers, single-phase direct liquid cooling, immersion, and hybrid architectures, enabling a gradual migration path from A100/H100 class air cooled systems through Blackwell NVL72 into fully liquid Rubin racks.

This multi-architecture capability is critical as NVIDIA’s roadmap pushes toward larger Rubin-family “AI factory” systems such as NVL144 and NVL576, which are effectively liquid cooled supercomputers in a box with tight power and thermal envelopes. Colovore’s design allows you to introduce those form factors into existing halls rather than having to commission separate, bespoke facilities.

One Facility. Multiple GPU Generations. Zero Rebuilds.

Beyond individual rows, Colovore designs the entire facility (i.e., shell, data halls, mechanical yards, and electrical rooms) to be upgradable without disrupting customers. Within each hall, we can run air cooling, rear-door heat exchangers, single-phase direct liquid cooling, immersion, and hybrid architectures side-by-side, and then scale those patterns up as entire halls transition toward Rubin-class, AI-factory layouts over time.

• As efficiency standards tighten and Rubin-class warm-water cooling becomes the norm, decoupled cold-water and warm-water cooling loops will allow chiller based plants to be swapped for adiabatic or closed loop dry coolers while keeping existing halls in service.
• Entire halls can be re-configured from rear door centric cooling to end row CDUs, high-density fan walls, or full “AI factory” layouts as the tenant mix shifts from legacy workloads toward AI-heavy Rubin-class deployments.

The result is a facility whose useful life is measured in decades, even as the GPUs inside it turn over on annual cycles. For finance teams, this means data center capex can be modeled like long-lived infrastructure again, without pretending that the hardware roadmap has slowed.

Colovore’s Next-Generation Design: A Practical Roadmap for Blackwell, Rubin, and Beyond

Colovore’s next-generation design work extends this philosophy. It preserves the same electrical and mechanical flexibility but emphasizes more capital-efficient and modular deployment patterns tailored to today’s AI buildouts.

Rather than over provisioning entire halls for worst case Rubin-class densities years in advance, the next-generation Colovore approach uses modular high-density “pods” that can be upgraded in place:

• Phase 1 – Land Blackwell safely: Deploy high-density Blackwell and Blackwell Ultra clusters into pods engineered for their power and cooling envelopes, without paying for Rubin-class capacity everywhere on day one.
• Phase 2 – Introduce Rubin without a rebuild: As Rubin NVL72 racks enter the mix, Colovore can convert specific pods from rear door plus DLC to full 100% direct-to-chip liquid service, leveraging pre-planned power and piping stubs instead of starting from bare concrete.
• Phase 3 – Scale into AI factories: When NVL144/NVL576 class “AI factory” systems or future post-Rubin architectures arrive, pod-level changes can be escalated to hall-level topology shifts, while the base shell, electrical backbone, and Facility water loops remain in place.

Economically, this approach trades a modest premium in initial flexibility for a significant reduction in the probability of stranded capital. Competing facilities tightly optimized around one generation and targeting the lowest possible build cost per MW, which is often on the order of $11M per MW or below, face higher upgrade friction as soon as the next platform arrives.

In a world where NVIDIA’s annual accelerator cadence is widely expected to continue, and where the largest cloud and consumer platforms are projected to spend hundreds of billions of dollars annually on infrastructure, the risk of narrow, single-generation designs compounds rapidly over an asset’s lifetime.

What Rubin Means for Your Blackwell Roadmap and How Colovore Can Help

Rubin changes how you should think about Blackwell and how you plan for the future. For enterprises with existing or planned Blackwell deployments, Colovore’s model offers three concrete advantages:

• Maximize Blackwell’s useful life: Deploy into facilities that comfortably support today’s Blackwell densities and are already wired and plumbed for Rubin-class racks. That way, you can run Blackwell as long as it makes economic sense instead of retiring it early because the building can’t keep up.
• Avoid surprise capex events: Treat new NVIDIA generations as workload transitions inside a flexible plant, not as triggers for new construction, emergency retrofits, or accelerated write-downs of recently built halls.
• Protect asset value and investor capital: Position AI infrastructure as a durable platform that can host several GPU roadmaps, which supports higher asset valuations and more predictable cash flows than narrowly optimized, single-generation builds.

Colovore was purpose-built for multi-generation AI infrastructure in a world where the GPUs change every year, but your buildings need to last for decades. If your team is deploying Blackwell now and wrestling with what Rubin and post-Rubin systems mean for your next five-year plan, Colovore can help map your current footprint against Rubin-class requirements and design a migration path that avoids stranded capital.