Our Take
Old IT infrastructure vendors have a real window to sell new hardware, but they're selling capacity, not expertise—and that's a problem when the customer already knows what they need.
Why it matters
The 1990s–2000s server market is a mature, price-competitive slog. AI workloads (GPU clusters, inference infrastructure, networking) are permission to rebuild margins, but only if these vendors can convince enterprises that buying new hardware from Dell or HPE is simpler than buying direct from NVIDIA or hyperscalers.
Do this week
Infrastructure teams: inventory your current hardware refresh cycle and GPU utilization before a vendor calls with an AI-bundled pitch—know your leverage before the conversation starts.
Old server makers spot an opening in AI infrastructure
Dell Technologies, Hewlett Packard Enterprise, and Cisco Systems are repositioning their legacy hardware brands (PowerEdge, ProLiant, and networking gear) as AI infrastructure plays. The thesis is straightforward: as enterprises deploy large language models and fine-tuning workloads in-house, they will need to replace or upgrade servers, storage, and interconnects. Companies that moved into cloud and outsourced their data centers in the 2010s may now want to buy hardware again for private AI infrastructure.
This is not a new market creation. It is a recapture play. These vendors dominated enterprise infrastructure for two decades before hyperscalers consolidated the market and sold compute as a managed service. AI workloads do not inherently favor on-premises infrastructure, but they do create a near-term purchasing event, and that is enough to get sales teams active.
Hardware sales alone won't solve the margin problem
The server market is mature and highly competitive. Selling a box to replace a box is not a margin business anymore; hyperscalers and custom OEMs (Supermicro, Wiwynn, and others) have hollowed out traditional vendors on price. A brief AI-driven refresh cycle can boost revenue, but it does not create stickiness or defensibility if customers can source the same components elsewhere.
The real risk for Dell, HPE, and Cisco is that they are marketing hardware into a market where the customer (CTO, infrastructure lead) already understands GPU bottlenecks, interconnect latency, and power density. The customer does not need a sales pitch about AI readiness; they need validated configurations and delivery. That is a commodity sale, not a consulting relationship.
Hyperscalers and cloud providers have already optimized for AI workloads. Enterprises evaluating in-house deployment are often choosing between private data centers (which means buying hardware) and renting capacity from a cloud provider. The decision is economic and architectural, not a function of brand trust or vendor "AI strategy."
Evaluate private AI infrastructure on total cost, not vendor positioning
When a traditional IT vendor pitches you an AI-ready bundle, separate the hardware specification from the narrative. You care about GPU availability, interconnect bandwidth, power efficiency, and support response time. You do not care whether the server comes from a 1990s brand or a newer ODM, as long as it meets your performance and cost targets.
Request benchmarks tied to your specific workload (model size, batch size, inference or training). Ask whether the vendor can deliver GPUs at your target price per unit and whether the interconnect architecture (NVLink, PCIe, or InfiniBand) matches your parallelism strategy. Then compare to cloud pricing for the same workload run remotely. The answer will tell you whether you should buy hardware or rent.
Do not let a vendor's AI branding or strategic narrative obscure the fact that you are still buying commodity hardware in a commodity market. The only difference is that right now, demand is high enough that vendors can ask for margin again. That window closes when supply catches up.