AI's billion dollar barrier: Training costs hit $1bn and beyond
Artificial intelligence (AI) training costs are skyrocketing, with models costing up to $1 billion currently in development, according to Anthropic CEO Dario Amodei. During the In Good Company podcast, Amodei predicted that these costs could rise to $10 billion or even $100 billion by 2025, 2026, or 2027. He also discussed the evolution of AI from generative artificial intelligence to artificial general intelligence (AGI), emphasizing that this transition would be a gradual process.
Hardware: Primary cost driver in AI training
Amodei suggested that if AI models become 10 times more powerful each year, the hardware required for their training would need to be at least 10 times more powerful as well. In 2023, ChatGPT reportedly required over 30,000 GPUs for its training, with a confirmed cost of $100 million. Given the price range of NVIDIA's newest B200 AI chip and the current rate of advancements in model research, hardware requirements are expected to keep pace until more efficient technologies emerge.
Rising demand for AI chips and power supply concerns
The demand for AI chips is increasing, with Elon Musk expressing interest in purchasing 300,000 B200 AI chips. OpenAI and Microsoft are also planning a $100 billion AI data center. However, power supply and related infrastructure issues also need to be addressed by AI firms. The total estimated power consumption of all data center GPUs sold last year, could power 1.3 million homes, raising concerns about potential shortages of economically-priced electricity in the future.
Future of AI: Innovations and societal impact
As AI technologies continue to improve exponentially, hardware upgradations seem to be keeping pace. Anthropic's $100 billion estimate appears feasible if brands like NVIDIA, AMD, and Intel can deliver. However, the question remains about how these advancements in AI technology will impact our society in the future. Tech companies like Microsoft are mulling modular nuclear power for their data centers, as a solution to potential power supply issues.
