Lisa Su: 60% of Global Computing Power for Inference
Lisa Su stated that approximately 60% of global computing power will be used for inference this year.
She also mentioned that 2026 will be the first year when AI computing power used for running models exceeds that used for training.
This assessment clearly shifts the demand focus towards inference, benefiting the infrastructure supply chain aimed at deployment, invocation, and agentic AI. amd+1
Source: Public information
ABAB AI Insight
This is not merely a demand assessment, but a narrative shift that AMD has been promoting over the past two years: AI is moving from training large models to large-scale invocation and deployment, with the focus of computing power shifting from a few super-large customers to more industry clients.
On the capital path, the rising share of inference means that procurement logic will shift from "buying the strongest training clusters" to "buying systems with better cost-performance ratios, lower latency, and more suitable for continuous operation," which will directly benefit chips, memory, networks, and rack-level integrated solutions.
Similar cases can be compared to the mobile internet shifting from "building content platforms" to "doing distribution and invocation," or cloud computing moving from "opening services" to "long-term hosting"; in terms of industry positioning, AI infrastructure is entering a second phase focused on deployment efficiency and cost per token.
Essentially, this is a shift in pricing power, because when inference becomes dominant, competition will no longer focus solely on peak training performance, but on unit cost, energy consumption, bandwidth, and sustainable supply capacity; those who are more suitable for large-scale operations will find it easier to secure orders.
ABAB News · Cognitive Laws
Training determines the starting point, inference determines the outcome.
In the second half of computing power, the competition is about the cost of each invocation.
When models are used, infrastructure starts to make money.