The AI revolution isn’t just happening on your browser screen — it’s happening underground, inside massive climate-controlled warehouses, consuming electricity at a scale most people never think about. While we debate the latest language models or the next consumer GPU, a silent war for watts, cooling capacity, and silicon is reshaping energy geopolitics and the global hardware market.
In this post, we’ll break down the real numbers: what datacenters consume today, what they’ll demand tomorrow, and how that insatiable appetite is already affecting everything from the U.S. power grid to the price of your next RAM upgrade.
| 📊 NUMBERS THAT PUT THE DEBATE IN CONTEXT |
| ~1.5% of global electricity consumption datacenters today (IEA) | 300% projected growth by 2035 Deloitte/TechCrunch forecast | 123 GW AI capacity by 2035 vs. 4 GW in 2024 |
1. The Electrical Appetite of AI Datacenters
To grasp the magnitude of the problem, we need to start with the fundamentals. A modern AI-focused datacenter is not just a glorified server room — it’s a high-density infrastructure with electrical and thermal demands comparable to a small city.
1.1 The Current State
According to the International Energy Agency (IEA), datacenters already account for roughly 1.5% of global electricity consumption. That percentage may sound modest, but in absolute terms it exceeds 460 TWh consumed in 2022 alone — more than the annual electricity usage of several mid-sized European countries.
Within that landscape, the fastest-growing segment is AI servers, whose consumption is expanding at nearly 30% per year. Training a single state-of-the-art large language model can require more than 1,200 MWh per training run — and a single query to a generative AI system consumes up to ten times more energy than a traditional Google search.
| ⚡ For context: one GPT-4o query uses roughly 0.42 Wh. Multiply that by 700 million daily queries and the aggregate impact becomes massive, even if each individual request seems negligible. |
1.2 The Projections That Are Raising Alarms
Forward-looking figures are even more striking. Recent studies from Deloitte indicate that, by 2035, energy demand from AI-dedicated datacenters could jump from 4 GW (in 2024) to 123 GW — a more than 30-fold increase in a single decade.
McKinsey projects that between 2025 and 2030, 125 GW of new datacenter capacity for AI will be added — a volume comparable to Spain’s entire installed electricity generation capacity. For the United States specifically, datacenters could consume up to 12% of all national electricity by 2028.
Regions like Northern Virginia, Georgia, Ohio, and Texas are already feeling the strain on the power grid, with rising electricity prices for both industrial and residential users. The IEA estimates that datacenters alone will generate demand of approximately 945 TWh by 2030.
2. Inside the Consumption: From Training to Inference
From a technical standpoint, the energy lifecycle of AI breaks down into two major phases with completely different dynamics:
Phase 1 — Training
This is the most energy-intensive process. A single pre-training run for a large-scale model can last weeks, requiring thousands of GPUs operating in parallel. Consumption is predictable but extremely high and temporally concentrated. Models like GPT-4 and Gemini Ultra represent the state of the art in this demand tier.
Phase 2 — Inference
Unlike training, inference has lower consumption per request — but it runs billions of times per day. Research published on arXiv points to token generation as the primary driver of energy use in production environments, and so-called ‘reasoning models’ — which generate hundreds of intermediate tokens before delivering a final answer — amplify this consumption significantly.
Models like OpenAI’s o3 and DeepSeek-R1 can consume more than 33 Wh per extended query — over 70 times the cost of compact models like GPT-4.1 nano. At the other end of the spectrum, Claude 3.7 Sonnet stood out in studies as the most efficient model in the performance-to-energy ratio.
| 🔍 Technical perspective: DeepSeek-R1, despite consuming 50–75% less energy during training (by not using the most advanced GPUs available), demands roughly 41% more energy than Meta’s model to respond to the same prompt in production. Architecture matters — but the bottleneck shifts between phases. |
3. The Scarcity Chain: From AI to Your Hardware
The energy impact of datacenters has a less-discussed but equally relevant sibling for infrastructure professionals: the pressure on the hardware supply chain. And it’s directly affecting what you’ll pay for your next server or workstation upgrade.
The GPU War
Nvidia continues to dominate the datacenter GPU market with over 90% market share, generating datacenter revenue exceeding $62 billion in the last reported quarter — 75% year-over-year growth. This staggering demand for AI accelerators is creating a direct side effect: shortage of components for the consumer market.
According to representatives from manufacturers like ASUS, models such as the RTX 5070 Ti entered EOL (End of Life) status not due to lack of public interest, but because of critical chip shortages. Nvidia is reportedly shipping board partners only half the component availability compared to the same period the previous year.
The practical result? If you were planning to build or upgrade a high-performance server or workstation in 2026, prepare for higher prices and tighter availability.
Pressure on Memory and CPUs
The problem isn’t limited to GPUs. Demand for HBM (High Bandwidth Memory) — the memory type used in AI accelerators like the H100 and Blackwell — is creating a shortage across the entire DRAM supply chain. Analysts at Counterpoint Research noted that a 3% supply-demand imbalance in DRAM is enough to trigger sharp price spikes. RAM prices jumped as much as 30% in a single quarter of 2025.
In the datacenter CPU segment, the situation is equally strained. Intel stated it expects to reach its lowest inventory levels in Q1 2026, while silicon wafer shortages are identified as a structural bottleneck. ‘Wafers don’t grow on trees’ — a line from an industry analyst that neatly captures the problem.
Goldman Sachs estimates that AI infrastructure investment will reach between $600 and $650 billion in 2026. That capital, concentrated in a few segments of the production chain, naturally drives up costs for everyone else.
Brazil on the Global Radar
Far from being a passive bystander, Brazil is emerging as a relevant player in this market — and the primary reason is energy. The country’s predominantly renewable electricity matrix (hydro + solar + wind) becomes a significant competitive advantage for datacenter operators concerned about carbon footprint and operational costs.
The CBRE Global Data Center Trends 2025 report positions São Paulo as the Latin American market leader, with 493 MW of capacity in operation in Q1 2025 — nearly 13% growth compared to the same period in 2024. For reference, mature markets like London and Frankfurt operate with over 1 GW of capacity.
There is enormous room for growth — and with it, increasing demand for IT professionals skilled in high-density infrastructure design, operations, and maintenance. That’s exactly the profile the market needs most right now.
Sustainability: Real Problem or Marketing Spin?
The tech industry has been actively building its sustainability narrative — and part of it is genuine. Major players like Google, Microsoft, and Amazon have committed to operating on 100% renewable energy. Microsoft even reopened discussions about nuclear power to feed its datacenters.
An analysis from Schneider Electric argues that the energy ‘cost’ of scaling AI globally would represent just 1% of total electricity consumption in a developed country — while the potential return, through AI-driven optimization of buildings and industries, could yield up to a 15% reduction in overall consumption. The argument is that AI is the solution, not the problem.
However, the picture is more complex. Accelerating technology cycles force server replacements every 3–5 years, generating growing volumes of e-waste containing critical materials like rare earths, gallium, cobalt, and tantalum. Water consumption for cooling is another emerging bottleneck, particularly in water-stressed regions.
| 💡 Trend to watch: The SMR (Small Modular Reactor) market — modular nuclear reactors ranging from 15 to 75 MW — is being actively explored by major tech companies to provide stable, carbon-free baseload power directly on datacenter campuses. First commercial deployments are expected around 2027. |
Conclusion: What This Means for You
If you work in IT — whether in infrastructure, development, or management — understanding this energy dynamic isn’t just intellectual curiosity. It’s career and business strategy.
Demand for professionals with expertise in datacenter energy efficiency, liquid cooling, high-density architecture, and AI infrastructure is growing at the same pace as server consumption. Those who can navigate the intersection of energy, hardware, and software will have a meaningful competitive advantage in the years ahead.
On the immediate practical level: if you’re planning hardware investments — for your homelab, your company, or clients — consider that we’re in a cycle of price pressure and supply constraints with no clear end date. Anticipating critical purchases and diversifying suppliers are valid strategies in this environment.
| Key Takeaways Energy scale: datacenters already consume 1.5% of global electricity, with AI’s share growing ~30% per year 2035 projection: AI capacity may grow from 4 GW to 123 GW, with total datacenter demand rising 300%Chip shortage: demand for GPUs and HBM is causing shortages in the consumer market and driving up RAM and SSD pricesBrazil on the map: Sao Paulo leads Latin America in datacenter capacity with 493 MW and 13% growth in Q1 2025Sustainability: SMRs and renewable energy are promising bets, but e-waste and water consumption remain real challenges |
Tags: #Datacenter #ArtificialIntelligence #Energy #GPU #Infrastructure #IT #NVIDIA #Hardware
Leave a Reply