Palo Alto Networks CEO Nikesh Arora: The Combination of Models and Computing Power is the Moat
Palo Alto Networks CEO Nikesh Arora posted, "Sam may have been ahead of his time, but he might be right - the combination of models and computing power is the moat." He reflected on multiple meetings with Sam, where the latter continuously raised his ambitions for computing power to the gigawatt level, and pointed out the misalignment in the current debate on value creation and capture: the market is eager to downplay models and debate open source, while real actions are taking place at the infrastructure level; new cloud vendors are quickly securing licenses, land, and electricity, while hyperscale cloud vendors continue to ramp up capital expenditures, and leading large model players are also working hard to aggregate computing power to meet training and deployment demands.
The difficulty of deploying computing power is rising sharply with demand, making infrastructure the biggest beneficiary of current expenditures. For leading models to succeed, they must resemble hyperscale cloud vendors more and embed themselves in the profit pools of application layers (coding, science, law, and physical AI). The competition in event-driven AI is shifting from model capabilities to a dual axis of "computing power control + application monetization."
Source: Public Information
ABAB AI Insight
As the Chairman and CEO of Palo Alto Networks, Nikesh Arora previously served as President and COO at SoftBank, where he was deeply involved in technology investments and has long observed the evolution of infrastructure and application layers in Silicon Valley. This time, he views the binding of computing power and models as a core moat, continuing his consistent judgment of enterprise-level AI focusing on "depth rather than breadth."
In terms of capital and resource pathways, hyperscale cloud vendors and new cloud companies are gaining an advantage by locking in electricity and land early, while leading model companies are forced to transition from pure models to "self-built or locked computing power + embedding high-profit applications." The motivation is to transform short-term training advantages into sustainable deployment and monetization capabilities, avoiding becoming a subsidiary of open source or cheap inference.
Analogous to the early internet's migration from content websites to infrastructure and platforms, and cloud computing's shift from software to hyperscale data centers, AI is currently in a phase of converging from "model capability demonstration" to "computing power and application closed loops."
Essentially, this is about capital concentration and technological substitution: as computing power becomes a scarce bottleneck, value shifts from replicable model weights to non-quickly replicable electricity, land, and deployment capabilities, while also forcing model companies to integrate computing power upstream and embed application profit pools downstream, reshaping the pricing power of the entire industry chain.
ABAB News · Cognitive Laws
- Models are replicable, computing power cannot be quickly formed.
- The true moat always lies at the bottleneck.
- Scale ultimately forces you to become what you once relied on.