Empowering Cloud Providers: How Server OEMs and Data Center Infrastructure Vendors Take Different Paths to Independence from Colocation Providers

At HPE Discover 2025, networking announcements underscored the company’s ambition to deliver AI-ready infrastructure through a unified approach spanning campus, data center, and edge environments. Central to this vision were new platforms designed to handle the scale and complexity of Artificial Intelligence (AI) workloads. The QFX5250 switch, built on Broadcom’s Tomahawk 6 silicon, exemplifies this shift with ultra-high bandwidth and readiness for emerging Ultra Ethernet standards. Alongside it, the MX301 multiservice edge router brings high-capacity routing to support distributed inference at the edge. Beyond raw performance, HPE emphasized sustainability and operational agility, introducing fully liquid-cooled designs that dramatically improve energy efficiency and cooling density. This innovation enables Cloud Service Providers (CSPs) to reduce dependence on traditional colocation facilities and their utility constraints, allowing faster and more flexible technology refresh cycles without being tied to legacy power and cooling infrastructure.

These announcements fit into a wider industry context of a more heterogenous CSP landscape: Diverse enterprise requirements paired with frustration with the traditional hyperscalers have given rise to smaller, more specialized CSPs with a more specialized focus on AI and High-Performance Computing (HPC) workloads (the so-called “neoclouds”) or center their value proposition around guaranteeing a certain degree of sovereignty. As these often rent data center space within colocation providers, they are, to some degree, dependent on the infrastructure supplied by colocation providers.

This becomes particularly important when thinking about technology refresh cycles. With both NVIDIA and AMD releasing new Graphics Processing Unit (GPU) accelerators at an increasingly compressed cadence (NVIDIA moving from ~18–24-month architecture cycles toward annual releases with ~12-month in-generation refreshes; AMD accelerating RDNA/CDNA updates with similar ~12–24-month generational spacing), CSPs need to be empowered to keep pace with these shortened refresh cycles. This is making higher electricity provisioning and liquid cooling a near-term requirement for cloud providers.

To facilitate technology refresh cycles, server Original Equipment Manufacturers (OEMs), silicon vendors, and system integrators aim to increase CSPs’ independence from the infrastructure that collocation providers offer. While server OEMs focus on equipping their servers and networking portfolio with direct liquid cooling capabilities, there is increasing momentum around smaller, modular data centers as well: Infrastructure vendor Prose, for example, offers compact and modular data center solutions that enable smaller CSPs to deploy localized, high-density compute infrastructure without relying on large colocation facilities. Not only do these modular data centers give CSPs greater flexibility, but it also will help overcome grid limitations and decrease deployment timelines. Both of these aspects  are becoming an increasingly important bottleneck for data center build-outs, particularly in Europe.

These approaches should be seen as complementary to drive the build-out of AI and cloud infrastructure. To account for the fact that the CSP landscape is becoming a lot more heterogenous and providers are emerging from different backgrounds, vendors and server OEMs should carefully consider the capabilities of each CSP to determine the best sales strategy.

Silicon vendors should continue building data center reference architectures (as NVIDIA is doing with its AI Factory and AMD with the Helios rack design) to move beyond GPU-centric blueprints and pivot toward a vendor-neutral orchestration layer with the capabilities to automate lifecycle upgrades across compute, networking, and cooling infrastructures in data centers. This can, for example, include a reference control plane with open Application Programming Interfaces (APIs) to ingest telemetry data from servers. From an operational perspective, these reference designs could include data center-as-a-code pipelines that will automatically stage driver/firmware updates and roll these out with certain guaranteed guardrails. From a commercial perspective, these reference architectures should be based on software subscriptions that are priced based on cluster sizes and specific outcomes. In addition, they should look at bundling offerings together with OEMs and modular data center providers (like Prose) that include orchestration licenses and installation services.

Server OEMs should look at implementing upgrade-friendly architectures (in addition to their current efforts to design servers and networking equipment with direct-to-chip cooling). This can include hot-swappable GPU trays and tool-less chassis designs, standardized PCIe/CXL interconnect layouts for easy expansion of accelerators, and firmware orchestration tools to initiate automated upgrade controls. Furthermore, they can intensify their work on reference designs for AI/HPC. To address the growing requirements for data compliance, this should include sovereignty-ready firmware packages with secure boot, encryption, and audit logs.

Meanwhile, providers of passive data centers should actively coordinate the development of reference designs (accounting for compute, networking, cooling, and power integration) to subsequently implement at scale. In addition, they should also intensify integrating digital twin technologies for data center design and operation. This can include, for example, site-specific simulation of thermal behavior or load transitions, and pre-commission simulated dry runs for upgrade windows. To provide reliability to the end-customer (be it a CSP or an enterprise), data center infrastructure providers should look at developing joint Service-Level Agreements (SLAs) together with server OEMs around guaranteed cooling capacity, thermal stability, and power quality under specified densities.