Greg Osuri, founder of Akash Network, warns that accelerating AI training will push electricity demand beyond what current grids and fuel supplies can deliver. The warning matters to data center operators, infrastructure investors and regulators because it signals higher costs, steeper emissions and tighter capacity. Osuri bases the claim on the power draw of interconnected GPUs used to train models and on projections reported across industry coverage.
News reports and interviews quote Osuri saying the extra load will raise household bills and carbon output. The International Energy Agency projects that AI-focused data centers could consume 945 TWh in 2030, up from 415 TWh in 2024.
IDC expects total data center power use to more than double between 2023 and 2028, with AI-specific workloads reaching 146.2 TWh in 2027, a 44.7 percent compound annual growth rate. Goldman Sachs forecasts a 165 percent rise in data center power demand by 2030, underscoring the scale of the shift.
Rising electricity demand and collateral impacts
Coverage also flags collateral impacts, including more water needed for cooling—figures from Microsoft and Google are cited—as well as increased electronic waste as hardware turns obsolete. Together, these pressures point to higher environmental and operational costs for the ecosystem around AI infrastructure.
Osuri’s focus on the aggregated draw of interconnected GPUs highlights how training clusters can amplify grid stress even when individual devices appear efficient.
Osuri offers decentralization as a countermeasure, promoting small-scale GPU clusters in homes alongside renewable micro generation so compute load spreads away from giant facilities. Akash Network operates a marketplace where owners of idle GPUs lease capacity to buyers; the firm claims this lowers cost and eases grid stress by distributing demand.
In Q1 2025, Akash booked one million dollars in leasing revenue, which the original report treats as evidence of demand for distributed GPUs. The AKT token’s inclusion in several crypto indexes adds visibility to the model and its marketplace dynamics.
Risks remain despite decentralization. Aggregate energy costs are likely to climb, and regulators may cap consumption or impose sustainability rules. Incentives could shift toward distributed designs, potentially eroding centralized cloud dominance, yet heterogeneous hardware introduces operational headaches that can complicate execution.
The debate now turns to execution and regulation. Akash plans more GPU listings and tools for autonomous agents, and the next quarterly numbers will test whether decentralization can cut cost and relieve grid load while meeting AI training demand at scale.
