'By 2025-2026, the cost of training AI models will exceed 1 trillion yen, and AI that poses a threat to humanity will appear,' predicts CEO of AI company Anthropic

AI development requires large-scale computing resources consisting of GPUs and AI accelerators. AI company Anthropic CEO Dario Amodei estimates that the cost of training AI models will reach about $10 billion (about 1.54 trillion yen) in the next few years.

Transcript: Ezra Klein Interviews Dario Amodei - The New York Times


Training large-scale language models and multimodal AI models requires a large number of GPUs and AI accelerators. Semiconductor companies such as NVIDIA, AMD, and Intel are developing chips that can efficiently perform AI training, but these AI-specialized chips are characterized by their high performance but also their high cost. For example, NVIDIA's AI-specialized GPU 'H100' is reported to cost $25,000 (about 3.8 million yen) per unit.

What exactly is NVIDIA's AI-specialized GPU? Is it different from a gaming GPU? - GIGAZINE

Large companies such as Google and Amazon have purchased tens of thousands of H100s to build computing facilities for AI training, while Meta and Microsoft have reportedly purchased an extraordinary number of H100s, totaling 150,000. As such, the current situation is that high-performance AI training is extremely costly.

Meta and Microsoft buy 150,000 NVIDIA H100 GPUs, three times more than Google, Amazon and Oracle - GIGAZINE

Dario Amodei, who is involved in the development of large-scale language models such as GPT-2 and GPT-3 and is the CEO of AI company 'Anthropic' at the time of writing, said on a New York Times radio program, 'The current training cost of a high-performance AI model is about $100 million (about 15.4 billion yen), but the training cost of an AI model announced in late 2024 to early 2025 will rise to about $1 billion (about 154 billion yen).' In addition, he said, 'The training cost of a high-performance AI model will reach about $10 billion (about 1.54 trillion yen) in 2025-2026,' indicating that the training cost of AI will continue to rise rapidly.

Anthropic classifies the threats to safety that come with improved AI performance using an index called 'AI Safety Levels (ASL)'. At the time of writing, the AI had already shown signs of being misused in the development of biological weapons, and was classified as ASL-2. Anthropic classifies 'AI that poses devastating risks compared to search engines and textbooks' as ASL-3, and 'AI that has performance far removed from current AI and whose level of danger cannot be defined' as ASL-4. Mr. Amodei predicts that AI equivalent to ASL-4 will appear somewhere between 2025 and 2028.

in Software,   Security, Posted by log1o_hf