- AI server electricity demand is growing at roughly 30% annually, putting total data centre consumption on track for around 945 TWh by 2030, per the IEA’s base case, more than double current levels.
- GPU clusters in advanced AI deployments can exceed 100kW per rack, making liquid cooling a functional requirement rather than an option, and forcing a wholesale redesign of data centre architecture.
- Fortune Business Insights now puts the GPUaaS market at $8.66 billion in 2026, projected to reach $162.54 billion by 2034 — a trajectory that reflects how many organisations are choosing rented compute over owned infrastructure.
Training a top large language model can consume tens of thousands of GPUs, each drawing five to eight times the energy of a typical CPU chip, according to S&P Global. That single constraint is now reshaping everything downstream: how data centres are built, how compute is priced, and which organisations can afford to participate in AI development at all.
GPUs: The Engine That Doesn’t Stop
NVIDIA‘s H100, H200, L40S and B300 have become the reference hardware for serious AI workloads, not because GPUs are new, but because the scale of current model training has no precedent. A GPU’s ability to execute matrix operations in parallel is what makes it suited to deep learning; a CPU serialises those same operations and falls behind by orders of magnitude. The industry has moved past evaluating individual chips, though. What organisations now buy, or rent, are complete systems: multiple accelerators, high-speed memory, advanced networking and the software stack that ties it together. That systems-level framing is shaping how cloud providers package their offerings and it is shaping procurement conversations inside enterprises that once bought servers by the unit.
Power Density Forces a Rethink
Standard enterprise racks run at 5 to 10kW. A dense GPU cluster pushing past 30kW is already beyond what most air-cooled facilities can handle. Advanced AI deployments now routinely exceed 100kW per rack, a threshold where air cooling physically fails to remove heat fast enough. Liquid cooling solves this by pulling heat directly off the chip package, which lets operators pack more compute into less floor space and run it more efficiently. The resulting facilities look nothing like a conventional data centre: they are optimised for dense GPU clusters, continuous high-throughput inference, and the kind of sustained power draw that strains both building infrastructure and regional grids. NVIDIA’s Vera Rubin platform is already driving a new generation of facility designs built around extreme density and thermal management from the ground up, rather than retrofitting older buildings.
Cloud Compute on Demand
Building owned GPU infrastructure requires capital most organisations do not have, and for workloads that spike unpredictably, dedicated hardware sits idle for much of its life. GPUaaS addresses both problems. The market was valued at $8.66 billion in 2026 and is projected to reach $162.54 billion by 2034, per Fortune Business Insights’ current tracking. AWS Google Cloud and Microsoft Azure dominate at scale, while specialists like CoreWeave and Lambda Labs compete on GPU availability and pricing. In India, domestic providers including E2E Networks, Jio Cloud and Yotta have moved to fill gaps left by global hyperscalers, offering INR-denominated billing that removes currency exposure for local teams. For startups, research institutions and enterprises with variable workloads, the pay-as-you-go model remains the most practical route to accelerated compute, no procurement cycle, no capital commitment, no idle hardware.
Own vs. Rent: The Numbers
On-demand H100 instances from major cloud providers range from $3.50 to $8.00 per hour per GPU depending on region and commitment level. A sustained 30-day training run on eight GPUs at an average of $4.50 per hour comes to roughly $26,000 per month. For inference running continuously, those costs compound fast. A 2023 a16z analysis of cloud infrastructure economics found that cloud hosting becomes more expensive than owned infrastructure at around 50% to 70% sustained utilisation, meaning the break-even point is real, but most organisations do not hit it. Average GPU utilisation for startups typically sits well below that threshold which is why GPUaaS economics favour most smaller teams. In India specifically, E2E Networks priced H100 instances at approximately ₹200 to ₹275 per hour as of late 2024a material discount to global hyperscaler rates for teams billing in rupees. India’s cloud compute market for AI is projected to grow at a compound annual rate above 35% through 2028, according to IDC India, though independent confirmation of that figure is limited.



