For the better part of fifteen years, the data center industry operated on a roadmap that almost everyone agreed upon. Cloud adoption set the direction, and the work of those of us who design and build these facilities was to make them larger, denser, and more efficient within a set of assumptions we rarely questioned.
Those assumptions were specific, even if we seldom articulated them. Workloads were broadly predictable. A busy rack drew 5 to 10 kilowatts. Compute lived in a handful of large, centralized regions. Traffic flowed largely between users and servers. And air, supported by sound engineering, was sufficient to keep infrastructure running efficiently.
Today, every one of those assumptions is being challenged at once.
This is the part of the AI story that rarely makes the headlines. Most discussions concentrate on models, GPUs, and applications. Yet the more consequential transformation is taking place beneath the technology stack, within the facilities, power systems, and cooling infrastructure that make AI possible.
At Techno Digital, we see this shift firsthand as we build hyperscale and distributed edge infrastructure across India. The conversations we are having today with enterprises, cloud providers, AI companies, and public-sector organizations are fundamentally different from those we were having even three years ago.
The question is no longer simply how much capacity is required. The question is whether the underlying infrastructure is capable of supporting an entirely new generation of computing.
From Cloud Density to AI Density
To appreciate the scale of the change, one must begin with a single metric: power density.
In the cloud era, a typical rack drew somewhere between 5 and 10 kilowatts. Even a general-purpose CPU rack rarely exceeded about 12 kW. The entire design language from power distribution to airflow to floor loading to aisle containment was built around that band.
Modern AI systems such as NVIDIA’s latest GPU architectures can consume more than ten times the power of traditional enterprise racks, pushing densities beyond anything the cloud era was designed for. For context, a high-density H100 air-cooled rack tops out around 40 kW. In effect, we have moved an order of magnitude, from designing around kilowatts to designing around hundreds of them.
And the trajectory does not slow. The next platforms point toward racks in the 600 kW range, and the industry is already engineering 800-volt DC distribution to support racks approaching a megawatt later this decade.
This is the essence of the reset. Infrastructure can no longer be designed around yesterday’s averages. It must be engineered around tomorrow’s peaks, and those peaks are arriving faster than most planning cycles were built to absorb.
Power Has Become the First Question, Not the Last
About two decades ago, conversations were about redundancy and uptime. Today, they begin with megawatts, cooling strategies, and whether a facility can support the next generation of AI accelerators. That shift alone tells you how dramatically infrastructure priorities have changed.
AI has inverted that sequence entirely. Power is now the first question on the table, and increasingly the one that determines whether a project is viable at all.
The numbers explain why. The International Energy Agency estimates that data centers consumed around 415 terawatt-hours of electricity globally in 2024, roughly 1.5% of the world’s total. It projects that figure will more than double to about 945 TWh by 2030, slightly more than Japan’s entire electricity consumption today, and continue climbing toward 1,200 TWh by 2035.
The implication is significant. In the United States, the IEA expects data centers to account for nearly half of all electricity demand growth between now and 2030, and to consume more power by the end of the decade than the production of aluminium, steel, cement, and chemicals combined. Digital infrastructure is no longer a quiet line item on the grid. It is becoming one of the largest new loads in the economy.
This is why power availability, rather than GPU availability, increasingly determines where projects succeed or stall. Land, utility readiness, substation capacity, interconnection timelines, and long-term scalability now shape the map of where AI infrastructure can realistically be built. In a real sense, AI has returned our industry to first principles. Before we discuss compute, we must be candid about whether the power exists to support it, and whether it can be delivered on the timeline the business requires.
At Techno Digital, this reality has shaped our infrastructure strategy from the outset.
As part of the Techno Group, our expertise lies in building and operating critical power infrastructure across India. For decades, the group has delivered transmission, distribution, and substation projects that support some of the country’s most important infrastructure initiatives. That experience fundamentally informs how we approach digital infrastructure.
Cooling has moved from the Background to the Blueprint
As density rises, the second assumption falls: that air will be enough.
Air cooling has served this industry exceptionally well, and it is not disappearing. But physics sets a ceiling, and AI workloads are pushing directly through it. Beyond roughly 30 to 50 kW per rack, conventional air cooling becomes inefficient, and then impractical. That is precisely the threshold AI racks have already crossed, which is why a 120 kW system arrives liquid-cooled by design rather than as an option.
This is driving a genuine architectural shift toward hybrid and liquid-assisted designs, from rear-door heat exchangers to direct-to-chip cooling. The economics reinforce the engineering. Cooling has historically consumed up to 40% of a data center’s electricity, so the stakes are not only thermal but financial. Direct-to-chip liquid cooling captures heat at the source, where liquid removes it far more effectively than air, allowing facilities to operate warmer and, in many climates, to reduce or eliminate mechanical chillers entirely.
When we designed our Chennai hyperscale facility, we deliberately moved beyond conventional density assumptions. The facility was engineered with high-density AI deployments in mind, incorporating power architecture and cooling flexibility that can support workloads far beyond traditional enterprise requirements.
The real question is no longer whether liquid-assisted cooling will become mainstream. It is how quickly organizations can prepare for a future in which it becomes standard.
Speed of Execution is now a Competitive Advantage
AI adoption cycles are measured in months: new models launch quickly, and customers demand rapid access to capacity. By contrast, land acquisition, permitting, interconnection, and construction commonly span years. The resulting mismatch makes speed-to-market a deciding commercial differentiator.
To shorten timelines, leading operators are adopting:
- Pre-permitted campus strategies and utility partnership agreements.
- Modular construction and factory-built electrical/cooling skids.
- Forward purchasing of critical long-lead items and staged commissioning to bring capacity online incrementally.
Organsisations that treat deployment cadence as an engineering discipline with standardized repeatable modules and integrated supplier ecosystems capture demand that slower competitors cannot.
The Future is both Centralized and Distributed
Much of today’s attention is directed toward hyperscale AI campuses, and rightly so. Training the largest models demands enormous, concentrated pools of compute, power, and networking. But training represents only one half of the picture.
As AI moves from experimentation into production, inference becomes the dominant, everyday workload, and inference behaves differently. It needs to reside close to the user, where latency, responsiveness, and data residency matter. A fraud check, a clinical decision-support prompt, a manufacturing-line inspection, or a regional-language assistant cannot tolerate a long round trip to a distant region.
This is reviving serious interest in distributed and edge infrastructure, not as a rival to hyperscale but as a complement to it. The considered view is that the future architecture of AI will not be centralized or distributed. It will be both. Large facilities will anchor training and core workloads, while a layer of distributed infrastructure brings real-time intelligence and localized service delivery closer to where it is consumed.
This conviction is reflected in Techno Digital’s infrastructure strategy. Alongside our hyperscale developments in Chennai, Noida, and Kolkata, we are building one of India’s largest distributed edge network through our partnership with RailTel Corporation of India. With 102 edge data centers planned across 23 states and supported by RailTel’s extensive national fiber network, our objective is to bring digital infrastructure closer to the enterprises, governments, and emerging AI ecosystems that depend on it.
Sovereignty Has Outgrown Compliance
Alongside all of this sits a conversation that has matured faster than almost any other: Sovereignty.
For years, sovereignty was largely a question of where data was stored. That remains important, but the conversation has expanded well beyond it. As AI becomes embedded in finance, healthcare, public services, and critical industry, organizations and governments are asking deeper questions about who owns the infrastructure, who operates it, what network dependencies it carries, and how resilient it remains under stress.
In India, this shift is concrete rather than theoretical. The Digital Personal Data Protection (DPDP) framework is drawing more workloads onshore, and national initiatives are backing sovereign AI capacity directly, including large-scale GPU infrastructure and AI factories being built on Indian soil with both domestic and global partners. The signal is unmistakable: digital infrastructure is no longer regarded purely as a technology asset. It is being treated as critical national and economic infrastructure.
The decisions organizations make now about where workloads reside and who controls the systems beneath them will carry consequences that extend far beyond any compliance checklist.
Designing for the Next Decade
The AI industry remains in its early stages, and much will change. But one conclusion is already firm: the infrastructure that supports the next generation of AI will look fundamentally different from the infrastructure that supported the cloud.
The facilities that succeed will not simply offer more capacity. They will offer greater adaptability. They will be designed around power availability first, built for cooling architectures that can evolve without disrupting operations, engineered for speed of deployment, and grounded in resilience and long-term sustainability rather than yesterday’s averages.
For years, digital infrastructure performed its work quietly, enabling progress from behind the scenes. In the AI era, it has moved to the centre of the conversation, because the constraint itself is shifting. The future of AI will not be determined solely by the intelligence of the models we build.
The cloud era taught us how to scale applications. The AI era is teaching us how to scale infrastructure.
We believe the organizations that succeed will not simply be those with access to the most advanced models. They will be those that can align power, cooling, connectivity, sovereignty, and execution into a resilient infrastructure strategy.
Because AI may be created in software. But its future will ultimately be determined by infrastructure.
