A common response to concerns about AI data center power demand is that the problem will solve itself: next-generation chips are more efficient, so the grid will eventually catch up. This argument sounds plausible. It is almost certainly wrong — and understanding why matters for anyone making infrastructure investment decisions over a 5-year horizon.
THE EFFICIENCY ARGUMENT
The efficiency improvement in AI chips is real and significant. Nvidia's B200 delivers substantially more useful AI computation per watt than the H100. Future generations will continue this trend. If you believe the demand for AI computation is fixed, then efficiency gains reduce the power required to meet that demand.
The problem is that demand for AI computation is not fixed. It is expanding faster than efficiency is improving.
THE JEVONS PARADOX
In 1865, economist William Stanley Jevons observed that improvements in steam engine efficiency led to more coal consumption, not less. When a technology becomes more efficient — and therefore cheaper to operate — the economic incentive is to use more of it, not less. The efficiency gain is redeployed as consumption growth.
This pattern has repeated in energy technology for 160 years. More fuel-efficient cars led to more driving. More efficient refrigerators led to larger refrigerators. More efficient data centers led to more data processing.
THE AI JEVONS EFFECT IN NUMBERS
GPU deployment growth 2023–2025: estimated 8–10x increase
Net power consumption change: significant increase
IEA projection: data center power consumption triples by 2030
Source: IEA Electricity 2024 report, Nvidia product specifications
The B200 is three times more efficient than the H100. But hyperscalers are buying far more B200s than they bought H100s. The efficiency gain does not reduce power demand — it makes AI more economically attractive, which increases deployment, which increases total power consumption.
THE AGENTIC AI MULTIPLIER
Jensen Huang's description of the agentic AI era adds a further dimension. Agentic AI systems — models that operate autonomously, take actions and run continuously without human prompting — represent a qualitative shift in inference demand.
A generative AI model answers a question when a user asks it. An agentic AI system runs continuously, processing inputs, taking actions and generating outputs without waiting for human interaction. The compute and power requirements of a world where millions of AI agents run continuously dwarf the requirements of current prompt-and-response systems — regardless of per-operation efficiency gains.
THE IEA FORECAST
The International Energy Agency's 2024 Electricity report projects global data center electricity consumption to double between 2022 and 2026, and to triple by 2030. These projections incorporate the efficiency improvements from next-generation chips. The tripling of consumption occurs despite efficiency gains because deployment growth overwhelms them.
"Efficiency improvements reduce the cost of AI per operation. Lower cost per operation increases the volume of operations demanded. Higher volume overwhelms the efficiency gain. This is the Jevons Paradox applied to AI — and it means power infrastructure investment made today will not be stranded by efficiency improvements." — GridReadiness Intelligence
IMPLICATIONS FOR INFRASTRUCTURE INVESTMENT
For infrastructure investors and data center developers evaluating whether the transformer shortage and grid capacity constraint is a temporary problem:
- The IEA projects data center power demand tripling by 2030, accounting for efficiency gains
- Agentic AI deployment has not yet been modelled into most demand forecasts — the upside is material
- Transformer manufacturing capacity will improve over 2026–2028 but will not fully close the demand gap within the decade
- Infrastructure secured today — grid connections, transformer slots, power contracts — will remain valuable assets through the 2030s
The transformer shortage is not a problem that efficiency improvements will solve. It is a structural constraint that will persist for the duration of the agentic AI buildout. The time to secure grid-ready assets is now.