Musk: 99% of Computing Power Will Be Used for Inference in the Long Term
Elon Musk replied that in the long term, 99% of computing power will be used for inference.
The event-driven AI infrastructure narrative is shifting from training to inference, benefiting inference chips and deployment funding, while training-dominated investments are relatively adjusting.
Source: Public Information
ABAB AI Insight
Musk has previously discussed AI computing power allocation multiple times, including server-side and space computing; this emphasis on the proportion of inference reflects the sustained demand for model deployment far exceeding one-time training.
On the capital path, public judgments guide industry focus on inference efficiency, motivated by matching actual deployment scenarios like xAI and Tesla; resources are shifting from large-scale training clusters to high-throughput inference infrastructure.
Similar cases can be seen with other AI leaders predicting the ratio of training to inference; the industry is currently transitioning from a peak in training to scaling inference.
This essentially represents a technological substitution: the focus of computing power is shifting from model training to large-scale inference services, driven by continuous demand generated from user interactions and real-time applications, far exceeding consumption during the training phase.
ABAB News · Law of Cognition
Inference demand will ultimately far exceed training demand
Long-term computing power allocation is determined by usage scenarios
Public judgments can accelerate capital reallocation