#What Are the Key Changes in the AI Industry?
The artificial intelligence sector is poised to experience a significant inflection point. Epoch AI, a nonprofit focused on AI trends, forecasts that the computational power used to run AI models will outpace that required for their development by the year 2030.
#What Do the Projections Say About Inference and Training Compute?
Epoch AI's findings suggest a transition in the AI landscape, where the emphasis will shift from development to deployment economics. The organization predicts that by the end of the decade, nearly half of all inference compute will transition to using Application-Specific Integrated Circuits, known as ASICs. Unlike general-purpose GPUs that dominate today’s AI operations, ASICs are specialized chips optimized for specific tasks. Historical data reveals that inference has already accounted for 60% to 80% of the compute resources in real-world applications.
Currently, while the sector sees a stable approximate share of 5% for training compute, the growth rate for training efforts for advanced AI models is remarkable, at four to five times annually. The total AI compute capacity is also growing at a comparable rate, indicating a robust expansion in the industry.
By 2030, it is expected that the performance of frontier training runs will reach about 2e29 FLOP, which parallels the evolution from models like GPT-2 to GPT-4, investments that will exceed $100 billion. This ambitious growth anticipates that training runs may demand between 4 and 16 gigawatts of computational power. Epoch AI estimates that by 2030, AI power capacity in the U.S. could exceed 50 gigawatts, with global capacity surpassing 100 gigawatts.
#Why Are ASICs Gaining Popularity for Inference Tasks?
Various tech giants have already recognized the advantages of ASICs. Companies like Google and Amazon have developed specialized chips designed to enhance inference efficiency. This trend indicates that ASICs will represent a significant portion of the inference market, potentially capturing up to half by the decade's close.
However, the growth of this market isn't without its challenges. Three primary constraints may impact expansion efforts: power requirements, chip manufacturing capabilities, and data transfer limitations. Despite these hurdles, Epoch AI believes that addressing these issues remains feasible under current projections.
#What Should Investors Consider in Light of These Changes?
As inference compute becomes the primary growth area, investors should rethink opportunities within the semiconductor field. Although GPUs continue to be essential for training, the advancing market for inference suggests a substantial long-term potential for recurring revenues. Your firm should prepare for a monumental increase in infrastructure, including data centers, power generation, and cooling systems to support an estimated 50 gigawatts of capacity in the U.S. alone.
Investors will need to monitor the reliability of the projected four to five times annual growth rate in compute resources, as factors such as energy restrictions and geopolitical influences on chip supply could pose risks to this exceptional growth trajectory.