AI efficiency gains collide with rising energy costs

AI efficiency is improving. Inference costs are falling. Model performance per watt is climbing. And yet the aggregate energy draw from data centers continues to rise, according to the Financial Times. The efficiency gains are real, but they're being overwhelmed by the sheer volume of new compute being deployed. Every optimization buys headroom for the next wave of workloads.

The paradox is structural. When a technology gets cheaper to run, demand scales faster than the cost curve can flatten. The result is more total energy consumption even as unit economics improve. The same dynamic played out in cloud computing a decade ago, and it's playing out again now in the AI stack.

The energy question is no longer a footnote in the AI buildout—it's a binding constraint. Utilities are fielding requests for power allocations that rival small industrial complexes. Data center operators are locking in long-term contracts with renewables providers and reviving conversations about nuclear. The infrastructure layer is repricing around a world where compute is the load, not the office park.

Elsewhere in the tech stack, India's Electronics Mart is targeting new markets as job-loss fears ripple through established hubs, Reuters reports. The distributor is expanding beyond its traditional retail and B2B channels, sensing opportunity as enterprise buyers diversify supply chains and rethink where tech spending flows. The move reflects a broader recalibration: as some nodes in the global tech network tighten, others position to absorb the shift.

The AI energy story and the India distributor story are different tempos, but they share a theme. Infrastructure is being re-rated in real time. What worked when compute was cheap and centralized is being tested now that compute is expensive, distributed, and power-constrained. The next twelve months will clarify which bets were early and which were late.