IT Hardware & Networking – The De-Commoditization of the Hardware Stack Begins

研报英文原文证据摘录

IT Hardware & Networking – The De-Commoditization of the Hardware Stack Begins

Cloud remains the natural starting point for pilots, bursty workloads, experimentation and use

cases that require immediate access to frontier models. But Lenovo’s utilization analysis

suggests the breakeven threshold can be lower than investors may appreciate. In its H200

comparison, Lenovo estimates that an on-prem system becomes more economical than renting

comparable GCP capacity at only ~4.3 hours of daily utilization over a 5-year period. That said,

as enterprises move from PoCs to always-on AI use cases across customer service, software

development, analytics, knowledge work and automation, the cost curve begins to shift from

flexibility toward efficiency. That is where DELL and HPE are most relevant in our view, as both

are positioned to sell the enterprise infrastructure required to lower cost per token at scale, while

CSCO benefits from the networking layer needed to keep those AI systems performant,

connected and utilized.

Figure 2: Ownership vs. Rental – Lenovo White Paper

Source: Lenovo, Evercore ISI Research

And it’s worth noting that we don’t believe on-prem is universally cheaper in every scenario. For

bursty workloads, uncertain demand, short-lived pilots or use cases that require constant access

to the newest proprietary frontier models, cloud-based consumption will likely remain the better

choice. However, for enterprises with sustained AI usage and repeatable inference patterns, on-

prem and private-cloud architectures can become meaningfully more attractive on a total cost of

ownership basis. That is especially true as agentic AI expands token volumes. The more often a

workflow runs, the more valuable it becomes to own or control the infrastructure layer that