Fifteen states accounted for an estimated 80% of the national data center load in 2023. They are, in descending order from the highest: Virginia, Texas, California, Illinois, Oregon, Arizona, Iowa, Georgia, Washington, Pennsylvania, New York, New Jersey, Nebraska, North Dakota, and Nevada. The concentration of demand is also evident globally, with the IEA projecting that data centers in Ireland could account for nearly one-third of Ireland’s total electricity demand by 2026. According to the U.S. Department of Energy (DoE), the surge in connection requests from hyperscale facilities, particularly those requiring between 300 and 1,000 megawatts or more, is straining local electrical grids. These facilities often have lead times ranging from one to three years, further complicating the ability of local infrastructures to deliver the necessary power quickly. A leading factor influencing this demand, which is expected to escalate through 2030, is the energy requirements of AI applications. Recent advancements in computing hardware and software have paved the way for the emergence of large language models (LLMs), which are beginning to rival human performance across a variety of tasks. As the prominence of these models continues to grow, so do concerns for substantial increases in the energy needed to operate them. The situation reflects the broader trend of AI tools becoming increasingly integrated into the fabric of society, raising questions about sustainability and energy consumption. The DoE’s Secretary of Energy Advisory Board (SEAB) working group of key stakeholders includes: • Hyperscalers • Data center developers/innovators • Technology providers • Energy/electricity companies • Independent system operators and regional transmission operators • Environmental NGOs • Technical associations and academic researchers The DoE working group thoroughly examined various strategies to meet the rising power demands of data centers while ensuring that existing customers
remain unaffected and greenhouse gas emissions are minimized. They identified that the most significant obstacles to the expansion of data centers are often local in nature, stemming primarily from the sheer scale of these facilities and the discrepancies in the timing of infrastructure development. In light of these challenges, the group put forward several key recommendations focused on enhancing energy support. Included were measures to expedite the interconnection process for data centers, providing comprehensive assistance to operators in streamlining their business operations, and facilitating a seamless transition from smaller data centers to more efficient cloud-based facilities. Public-private collaboration is crucial to creating a balanced approach that accommodates data center growth while prioritizing sustainability and efficiency. PHOTONICS TO THE RESCUE IN 2025? With the escalating demands of AI and data center operations, there is a critical need for new, innovative strategies that leverage advances in hardware, system monitoring, computational algorithms, clean electricity procurement, and operational flexibility. Will 2025 be the year that photonics provides the breakthrough in addressing the massive power consumption problem in data centers? Here are ten photonics companies that have been working on solutions to reduce data center energy consumption. 1. Ayar Labs is developing technology to put optical I/O in the chip fabric. By moving data using light, optical I/O makes it possible to continue scaling computing system performance, enabling the next system design breakthroughs for the growth of generative AI, disaggregated data centers, and more. The technology allows faster communications inside chips and is intended to replace copper wires, which are slower conduits. The company is measuring cost and power efficiency in a throughput-per-dollar-per-watt metric, where optical technology becomes more economical. The company estimates the total photonic IC market to be more than US$38 billion in 2025 with the sales of chiplets to exceed $100 million by 2030.
I
January/February/March 2025
27
Made with FlippingBook - Online catalogs