What are experts saying about how long the AI compute build-out can run — where demand for chips, networking, and data-center hardware is headed, and which layers face the most oversupply risk?

The AI compute build-out is widely viewed as a sustained, multi-year cycle, with some experts predicting the underlying supply-demand imbalance for key hardware will persist for a decade or more [4, 7, 13]. Data center capacity is already planned through 2027-2028, making a near-term slowdown unlikely , and some forecasts suggest supply will lag demand for at least the next three to five years [26, 30]. A central consensus among analysts is that the primary bottleneck has decisively shifted away from the availability of GPUs themselves [8, 12]. The new constraints are now in the foundational physical infrastructure required to operate the chips, specifically the availability of **"powered shells"**—data centers with sufficient energy, cooling, and real estate [1, 9, 23]. This infrastructure layer faces its own severe shortages in components like transformers, switchgears, and structural steel, as well as in skilled labor such as electricians, which are now gating the pace of expansion [2, 9, 16]. Energy supply, in particular, is identified as the main bottleneck for the next five years , with one expert predicting the energy shortage may begin to alleviate around 2027 or 2028 .

Demand for this infrastructure is not only unrelenting but also evolving in nature, fueling the long-term outlook . A significant driver is the enterprise shift from cloud-based testing to on-premise production deployments, motivated by needs for data security and control [4, 7, 19]. Concurrently, workloads are transitioning from a primary focus on centralized model training to the more complex and distributed problem of inference , though training still constitutes most of the current demand . Looking forward, experts point to the emergence of "Agentic AI," where AI systems execute tasks rather than just generate content, as a paradigm that will fuel a massive new demand cycle for GPUs, high-performance CPUs, and vast amounts of memory [4, 7, 19]. This sustained demand reinforces the market position of dominant players like NVIDIA, whose hardware is exclusively requested by major customers due to the strength of its CUDA software ecosystem , a moat competitors have failed to overcome .

Despite the strong demand outlook, there is a notable tension regarding the risk of oversupply, particularly in data center capacity. Several analysts warn that the massive capital expenditures by hyperscalers could lead to an overbuild, drawing parallels to the fiber-optic glut of the dot-com bubble [14, 18, 22]. Investors are beginning to question whether projected demand will materialize by the time new facilities become operational . This could create a scenario where there is a surplus of chips that cannot be powered due to energy and infrastructure bottlenecks . However, this view is contested. NVIDIA's CEO sees a near-zero probability of a compute glut in the next two to three years, as the largest cloud providers are absorbing the initial build-out for their core businesses . Furthermore, the useful life of GPUs may be significantly longer than commonly assumed, potentially **6-8 years** or more, as a diverse ecosystem of models will create value for older hardware on less demanding tasks . This extended depreciation horizon changes the risk profile for financing hardware, suggesting that while a glut of physical data center space is a risk, the underlying demand for a matrix of computing capabilities is more durable.