Habr Author: GPU Clusters and Data Centers, Not Models, Are Now AI's Bottleneck

The International Energy Agency estimates that data centers consumed about 415 TWh of electricity in 2024, roughly 1.5% of global demand. Its forecast nearly doubles by 2030, reaching 945 TWh. Against this backdrop, a Habr author, in a piece on AI infrastructure, offers a simple thesis: the industry's main constraints now lie outside the models themselves. They are GPU clusters, data centers, procurement mechanisms, regulation and technological independence.

A caveat: only the announcement is public. The main part of the article is hidden behind a cut, and we have not seen any specific data in it. Below, we therefore test the author's thesis against open sources.

Electricity and Cooling Instead of Parameters

Training and serving large models runs into physics. An H100 accelerator draws up to 700 W. A GB200 NVL72 rack with 72 Blackwell accelerators delivers around 120 kW, while a rack in a typical enterprise data center is designed for 5–10 kW. Air cooling can no longer handle that density. It takes liquid cooling loops, redesigned machine halls and dedicated substations.

That leads to the central problem: you cannot buy a site with tens of megawatts of connected capacity in a quarter. Permits, grid connections and transformer deliveries take years, while a model can be fine-tuned in weeks.

Procurement as a Strategic Function

Since October 2022, the US has been restricting exports of advanced accelerators to China and a number of other jurisdictions in stages. Buying GPUs has gone from routine to a matter of compliance, logistics and politics. The Stargate project shows the scale of investment: in January 2025, participants announced up to $500 billion over four years.

For markets outside the main supply chains, the picture is tougher. Access comes through cloud rentals, parallel imports or local accelerators. Local solutions run into the software stack: the CUDA ecosystem, its libraries and optimization tools have accumulated years of debugging, and hardware alone will not close that gap.

Key Metrics of the Infrastructure Barrier

• Data center consumption: about 415 TWh in 2024, with an IEA forecast of about 945 TWh by 2030.
• H100: up to 700 W per accelerator.
• GB200 NVL72: around 120 kW per rack, 72 accelerators.
• Conventional data center rack: 5–10 kW.
• Stargate: up to $500 billion over four years, according to the project's participants.

How to Plan Capacity

The Habr author's thesis looks convincing, but you should test it against your own workloads. For most teams, the question is not whether to build their own cluster. What matters is calculating the cost of a token or a GPU-hour at real utilization in advance, and comparing renting, reserving and buying.

The practical minimum: write guaranteed accelerator quotas into the cloud contract, not just the price. Check whether your code is portable to another stack, so you do not depend on a single vendor. For your own sites, start with a power request and a cooling design, and choose models after that. If the full Habr text includes calculations on procurement and regulation, those will show how far this scenario applies to the Russian market.