Agentic AI is transforming the underlying logic of traditional hardware cycles and could turn infrastructure supply tightness from a short-term fluctuation into a structural constraint lasting several years.
Dell Technologies Chief Operating Officer Jeff Clarke said during a recent meeting with Morgan Stanley that agentic AI is clearly different from past hardware cycles. If related AI workloads continue to grow, shortages of key components could persist for more than five years, with memory and HDDs being his biggest concerns at present. Compared with past hardware supply tightness that typically lasted several quarters, this round of demand growth could bring a longer-term supply-demand gap.
Morgan Stanley reassessed the current AI infrastructure cycle after the meeting. Analysts noted that the eight hardware boom-and-bust cycles over the past 40 years were mostly driven by device refresh and capital expenditure fluctuations, while the key variable in this cycle is whether AI workloads can continue to expand. In other words, what the Dell executive emphasized is not an ordinary hardware restocking cycle, but the long-term infrastructure expansion that could be brought by growing agentic AI demand.
Agentic AI reshapes the logic of infrastructure demand
The report noted that Clarke believes this AI cycle is fundamentally different from the hardware booms and busts of past decades. The previous eight hardware cycles, covering servers, storage, PCs and other categories, were mainly driven by changes in device penetration and refresh cycles, with demand rising and falling along with refresh rhythms, but the hardware market itself did not undergo substantial expansion in scale.
Take PC demand during the pandemic as an example. The human-to-device ratio briefly increased, but as society reopened and returned to normal, PC demand also fell back, and the total market size before and after the pandemic did not fundamentally change. The logic of agentic AI is different. Clarke believes agentic AI is decoupling cognitive output from human input, allowing enterprises to complete more work without adding staff, thereby delivering productivity gains of 10x or even 100x.
In this process, AI is no longer just assisting people in completing tasks, but itself becomes a "productivity tool" that continuously consumes computing resources. Productivity gains will push more enterprises, regardless of size, to begin AI transformation, thereby continuously expanding the infrastructure TAM.
Dell expects that by 2030, data center compute will add 200 gigawatts, ZettaFLOPS compute will grow 5x to 830, and inference token growth will be exponential. Agentic AI requires far more tokens per unit of work than basic chatbots, and Clarke expects inference token generation to grow 87x by 2030.
Cloud and on-premises deployment are not a zero-sum game
One core question in the market about this AI infrastructure cycle is whether rapidly growing token demand will ultimately flow mainly to cloud computing platforms rather than on-premises infrastructure.
Clarke's judgment is that agentic AI workloads will take a hybrid deployment form. Enterprises will continuously adjust where inference workloads run between public cloud and on-premises environments based on security and cost efficiency.
Clarke used Dell itself as an example: the company will deploy content-related workloads in the public cloud, but will not move proprietary source code or telemetry data out of its own facilities. This means there is no single deployment path for AI workloads, and on-premises infrastructure will still absorb part of the continuously growing computing demand.
Under this framework, the exponential growth in token consumption does not mean that cloud and on-premises infrastructure must necessarily be in a seesaw relationship. Clarke believes the spread of open-weight models will further drive on-premises infrastructure investment, because enterprises can optimize model output for cost in on-premises environments.
Server shipments decline, but long-term demand growth logic remains intact
Traditional server shipments are still declining year over year, which appears to contradict the explosive growth in inference tokens. Clarke's explanation is that the sharp increase in server performance density is pushing down unit shipment volumes.
Dell's 17th and 18th generation servers can replace as many as 13 traditional 14th generation servers, so in the short term, even if server shipments decline, the value and capacity carried by each unit are still rising.
As data centers gradually complete the architectural restructuring centered on accelerated computing, and as agentic AI adoption further drives demand for CPU servers, server shipments are ultimately still expected to grow again. Clarke even said that if inference tokens truly achieve 87x growth within five years, traditional server shipments could also change exponentially.
Storage will likewise be continuously driven by agentic AI. Every operation performed by agentic AI, including memory retention and artifact generation, creates ongoing data storage demand; with large-scale adoption of KV Cache, storage capacity demand will further increase.
Supply shortages may become a structural constraint lasting more than five years
Compared with macroeconomics, geopolitical conflict, excessive data center construction and power shortages, Clarke ranks supply issues as his biggest concern at present, and used "We're in neverland" to describe the current supply environment.
Morgan Stanley noted that historical commodity supply cycles usually last 2 to 4 quarters, mainly affected by booms and busts in hardware refresh cycles and supply-side decisions. But Clarke believes that if one accepts the growth logic of gigawatts and inference tokens, then key components such as memory and HDDs are no longer merely in an ordinary short-term cycle.
Therefore, Dell is currently operating on the premise that key components may face a multi-year sustained shortage, with particular attention to memory and HDDs. In other words, the core risk of this round of supply tightness is not short-term insufficient supply, but that after AI demand continues to expand, the supply side may need years to catch up.
Clarke also said that Dell is superior to peers in supply management, and this advantage is translating into share gains in multiple submarkets including servers, storage and PCs.
Behind margin expansion is a change in the supply-demand structure
Persistent tightness in key components is also changing the pricing environment of the infrastructure industry.
Morgan Stanley had previously attributed part of the gross margin expansion in the server and storage businesses to "profit stacking," meaning that while costs for components such as memory rise, manufacturers further add profit margin. Clarke offered a different explanation.
He acknowledged that server and storage profit margins for similar products are indeed expanding, but believes there are two structural factors behind it: first, when component supply is limited, companies will prioritize allocating scarce resources to the highest-margin products; second, as the infrastructure market continues to expand, pressure on companies to compete for incremental customers declines, and they do not need to lower prices to win new customers.
Therefore, the improvement in margins is not entirely an opportunistic gain brought by short-term supply-demand mismatch, but also reflects a change in the industry's pricing benchmark after sustained expansion in infrastructure demand.
If the workload growth brought by agentic AI can continue to materialize, the core variable of the AI infrastructure cycle will shift from "device refresh" to "new demand." This means demand expansion for compute, servers and storage may last longer, while supply constraints for key components such as memory and HDDs may also shift from short-term cyclical fluctuations to multi-year structural problems.
Comments