Author: Baron Fung, Dell’Oro GroupThe development of hyperscale AI infrastructure is about to reach a more developed stage. Hyperscalers are currently focused on effectively scaling AI computation and supporting infrastructure after years of fast regional development motivated by resilience, redundancy, and data sovereignty.
Despite historically high absolute investment levels, the cycle is becoming more and more characterized by execution risk and capital expenditure discipline as we into 2026.
Accelerated Servers Drive Most Spending
High-end accelerated server purchases increased significantly in 2025 and will be a major source of funding for AI infrastructure throughout 2026. Large AI clusters rely on these systems to meet their demands for GPUs, specialized accelerators, HBM, high-capacity SSDs, high-speed NICs, and networks. As hyperscalers expand AI services to millions of users worldwide, inference workloads are increasingly driving an increasing portion of installations, even while frontier model training is still crucial.
Because inference workloads need more availability, regional spread, and tighter latency guarantees than centralized training clusters, this change significantly increases infrastructure needs.
GPUs Remain Leading Component in Revenue
Even as hyperscalers use more bespoke accelerators to improve cost, power efficiency, and workload-specific performance at scale, high-end GPUs will continue to be the primary driver of component market revenue growth in 2026. The Vera Rubin platform, which enhances system complexity via increased computing and networking density and optional Rubin CPX inference GPU configurations, is anticipated to start shipping from NVIDIA in 2H26. This will significantly improve component attach rates.
With the help of its MI400 rack-scale platform and the recently reported victories at OpenAI and Oracle, AMD is positioned to increase its market share. GPUs continue to fetch a disproportionate amount of money despite increased competition because of their higher ASPs and wider ecosystem support.
Inference Depends on Near-Edge Infrastructure
In order to fulfill latency, reliability, and regulatory requirements, hyperscalers will need to spend more in near-edge data centers as the need for AI inference increases. These facilities are crucial for real-time, user-facing AI services like copilots, search, recommendation engines, and corporate applications. They are situated closer to population centers than centralized hyperscale areas.
Smaller yet very dense accelerated clusters are usually preferred for near-edge deployments, which have strict requirements for redundancy, local storage, and high-speed networking. Despite not being as powerful as centralized AI campuses, these locations’ sheer quantity and geographic dispersion indicate a significant increase in capital expenditures needed until 2026. On the other hand, unless ecosystems and application demand further develop, far-edge installations are unlikely to experience significant development and are still more use-case dependant.
CPU and Networking Transitions Unevenly
After short-term inventory digestion, the x86 CPU and NIC markets associated with general-purpose servers are anticipated to slow down in 2026. On the other hand, increased compute expansion is still closely associated with the need for high-speed networking. Scale-out fabrics are still used by inference accelerators to enable utilization, redundancy, and ultra-low latency, even if inference workloads are increasing faster than training.
As 2026 approaches, supply chains for AI infrastructure are become more and more limited. The manufacturing of higher-margin HBM is being prioritized by memory suppliers, which is reducing the capacity of traditional DRAM and NAND utilized in AI servers. Consequently, system-level costs for accelerated platforms are growing because to the fast increase in memory and storage prices.
In addition to memory, supply chain instability is being increased by longer lead times for sophisticated substrates, optics, and high-speed networking components. The supply chain is also exposed to extra risk due to tariff uncertainty and changing trade policies, which might eventually raise component prices.
High Capex Meets Rising ROI Scrutiny
The multi-year AI investment cycle is expected to continue beyond 2026 as the US hyperscale cloud service companies continue to increase their capex projections. Competitive pressures, near-edge growth, greenfield data center construction, and accelerated computing continue to be significant tailwinds. Modifications to depreciation treatment provide levers to sustain short-term investment levels and maximize cash flow. But as infrastructure spending has surpassed revenue growth, questions about capital intensity, depreciation, and long-term returns are becoming more pressing. Cash flow timing is controllable, but the underlying ROI is dependent on successful AI monetization, which raises the possibility of margin pressure if infrastructure investment is delayed by revenue growth.
About The Author
Since joining Dell’Oro Group in 2017, Baron Fung has been in charge of market research for Ethernet adapters and smart NICs, data center IT semiconductors and components, and data center IT capital expenditures. With an emphasis on server technology, cloud service provider capital expenditures, and vendor trends, Mr. Fung has greatly broadened our coverage of data center infrastructure. Prominent trade and business magazines have frequently quoted his research and analysis. Mr. Fung has also received invitations to speak at conferences, industry events, and investor gatherings.
