The next bottleneck for your AI ambitions might not be talent or chips. It might be the power grid.
Across Silicon Valley and other tech-dense regions, electricity prices are climbing as data centres race to feed power-hungry AI workloads. Local utilities are scrambling to secure new energy supplies, and some areas are already hitting capacity limits. What was once a facilities management concern is now landing squarely on the desks of CTOs and CFOs.
The Numbers Behind the Surge
Training a single large language model can consume as much electricity as a small town uses in a year. Running inference — the process of actually using that model to answer questions or generate content — multiplies that demand across millions of daily queries.
In parts of Northern California, industrial electricity rates have jumped by double-digit percentages over the past 18 months. Data centre operators report that securing new power connections now takes 18 to 24 months in some regions, up from six months just three years ago. The constraint is not just generation capacity but transmission infrastructure — the physical wires and substations that deliver power to facilities.
This is not a localised American problem. Any market where AI compute clusters are concentrated will face similar pressure. For Indian enterprises running workloads on global cloud infrastructure, these costs will eventually flow through to your monthly bills.
Cloud Providers Are Already Adjusting
The major cloud providers — Amazon Web Services, Microsoft Azure, and Google Cloud — are responding in ways that will reshape how enterprises think about AI deployment. All three have signed multi-billion dollar agreements with energy companies. Microsoft recently announced a deal to restart a reactor at Three Mile Island to power its data centres. Google has committed to operating on carbon-free energy around the clock by 2030 and is investing heavily in next-generation geothermal and nuclear.
These investments are defensive moves. Cloud providers know that customers will eventually ask hard questions about availability and pricing stability. They also know that regulatory pressure on energy-intensive computing is coming, particularly in Europe and increasingly in the United States.
For now, the hyperscalers are absorbing much of this cost pressure. But that will not last forever. Expect to see energy-related surcharges, premium pricing for guaranteed capacity, or incentives to run workloads in regions with cheaper power.
What This Means for Build Versus Buy Decisions
If your organisation is considering on-premises AI infrastructure — GPU clusters for training proprietary models or running sensitive inference locally — energy economics deserve serious analysis. A rack of high-end GPUs can draw 40 to 50 kilowatts. Scale that to a meaningful training cluster, and you are looking at megawatts of continuous demand.
In India, industrial power costs vary dramatically by state. Locations with reliable supply and competitive rates — parts of Gujarat, Maharashtra, and Karnataka — will become more attractive for compute-heavy operations. States with power deficits or high tariffs will struggle to attract AI infrastructure investment.
This creates an opportunity for enterprises willing to think ahead. Long-term power purchase agreements, particularly with renewable energy producers, can lock in predictable costs while satisfying sustainability commitments. Some Indian data centre operators are already signing 10 to 15 year deals with solar and wind developers to secure capacity before demand spikes further.
The Regional Diversification Play
Smart infrastructure planning now includes geographic arbitrage. Running training workloads in regions with surplus renewable energy — Scandinavia, parts of Canada, or even specific Indian states — can meaningfully reduce costs while improving your carbon footprint.
This is not theoretical. Several global technology companies have begun shifting batch processing and model training to data centres in Iceland and Norway, where geothermal and hydroelectric power is abundant and cheap. The latency penalty is irrelevant for training jobs that run for days or weeks.
For Indian enterprises, this might mean reconsidering the assumption that all workloads should run in Mumbai or Singapore availability zones. Hyderabad and Chennai are emerging as data centre hubs with improving power infrastructure. The calculus is changing.
What This Means for You
Power is no longer a background assumption for AI strategy. Here is what deserves attention in your next planning cycle:
First, ask your cloud provider about energy cost trends and what protections exist in your contract. Second, if building on-premises capacity, model electricity costs over five years, not just at today’s rates. Third, evaluate whether workload placement flexibility — running jobs where power is cheapest — can become a competitive advantage.
The companies that treat energy as a strategic input, not just an operational expense, will have more room to scale AI investments profitably. The ones that ignore it will find their margins squeezed by a cost they never saw coming.