A new Australian technology promises to reduce the computing power required by generative AI (genAI) systems by two-thirds, providing a potential breakthrough for a rapidly expanding industry grappling with power shortages. Experts predict that by 2027, up to 40% of AI data centers could face limitations due to power constraints, underscoring the urgency of such innovations.
With over 200 data centres consuming an estimated 5% of Australia’s electricity, the Information and Communications Technology (ICT) sector is intensifying pressure on the nation’s energy resources. This figure is expected to rise, with AI-focused data centres likely pushing consumption to 8% by the end of the decade. Globally, AI has intensified the load on power grids, demanding a new approach to maintain sustainable growth while meeting energy requirements.
GenAI’s complex decoding processes necessitate significantly more power than conventional data queries, with genAI queries requiring around 3 watts each, in contrast to the 0.3 watts used by a typical Google search. Increasingly complex genAI models are expected to drive power demands even higher. The World Economic Forum notes that computational requirements for AI are doubling every 100 days. For instance, training OpenAI’s GPT-3 model required approximately 1,300 megawatt-hours (MWh) of electricity, while its successor GPT-4 consumed 50 times that amount.
These escalating demands create substantial challenges for power infrastructure, especially as consultancy TM Advisory projects Australian companies will invest $26 billion to expand local data centre capacity by 2030. Experts anticipate that this investment will create 9,600 new jobs and encourage approximately 45% of operators to switch to renewable energy sources. However, with 80% of data centres struggling to secure essential equipment to meet rising AI demands, consultants from Turner and Townsend caution that genAI systems will need to prioritize efficiency to keep up with demand sustainably.
New approaches are emerging to address these challenges. While companies like Nvidia continue developing high-powered, energy-intensive GPUs, scientists are exploring power-efficient technologies such as neuromorphic computing, which could offer cooler, more energy-efficient alternatives for AI. However, Australia’s Trellis Data is taking a different route, addressing power consumption at the software level.
Trellis Data’s recent innovation, Dynamic Depth Decoding (D3), optimizes the decoding process through speculative analysis, boosting genAI speed by 44% over popular models like EAGLE-2. This improvement translates to what CEO Michael Gately describes as a 68.4% reduction in power consumption for highly computationally intensive AI models. “The right model for the right purpose delivers significant value,” Gately told Information Age. “What better way to mitigate AI’s environmental impact than by making them faster and more efficient?”
Gately noted that Trellis Data’s clients, many of whom are confidential government agencies, require AI solutions that operate independently from the internet yet still meet the operational standards of AI technology on a global scale. Reducing power requirements also allows for smaller or fewer servers, expanding AI’s applicability to areas previously limited by infrastructure constraints.
Despite advancements like Trellis Data’s, the genAI industry continues to face mounting pressure. The industry’s rapid pace of development has surpassed Moore’s Law, the long-standing observation that computing power doubles every 18 months. Historically, this doubling effect sustained technological progress even as cloud-based data centers amassed vast computing resources. However, the rise of genAI has exposed the limits of Moore’s Law as it pertains to energy efficiency.
The cryptocurrency boom provided an early warning, driving global power consumption to previously unseen levels. However, AI’s demand for power is proving to be on an entirely different scale. Gartner vice president and analyst Jorge Lopez remarked that few could have anticipated the massive energy requirements of supercomputing facilities dedicated to AI. He explained that although cryptocurrency mining was initially seen as highly energy-intensive, genAI demands have “shifted the entire conversation around energy consumption.”
This growing demand is expected to test existing power grids, many of which have operated using largely unchanged technologies for decades. Gartner predicts that global data center power consumption will reach 500 terawatt-hours (TWh) by 2027, a 2.6-fold increase from 2023. Despite infrastructure upgrades and technological advancements, Gartner warns that 40% of AI data centers will reach power capacity limits by 2027, restricting further growth.
To address these concerns, some tech giants are exploring nuclear power as a long-term solution. For example, Microsoft recently signed a 20-year contract to restart the Three Mile Island nuclear facility, site of the United States’ most infamous nuclear accident in the 1970s. This agreement reflects an increasing acceptance of nuclear energy as a reliable source to meet AI’s growing energy demands.
Lopez emphasized that while innovations will likely help reduce consumption in the future, substantial advances may still be years away. “To reach our goals, significant infrastructure investments and technological breakthroughs are essential,” he said. For the ICT industry and AI developers, the challenge is clear: balancing technological progress with sustainable power usage to enable continued growth without overwhelming global energy resources.
