For two decades, data center design assumed that resources were abundant enough to package into fixed servers and scale by adding boxes. CPUs, GPUs, memory, SSDs, NICs, and power were bundled in preset ratios, racked, and treated as the unit of deployment. That model worked in an era of relative abundance; it does not work in the scarcity era AI is creating.
AI workloads have changed both demand shape and supply constraints. Training, inference, retrieval, simulation, and data pipelines all ask for different mixes of compute, memory, storage, and network, while supply chains and power availability tighten around all of those components. In this environment, overprovisioned, underutilized servers are no longer just inefficient; they are structurally misaligned with the problem.
From Corespan Systems' perspective, this is an architectural inflection point: infrastructure must move from static, server-centric allocation to pooled resources, high utilization, and dynamic, workload-driven composition.
Corespan's resource-pooled architecture: CPU hosts with FIC 2500s connect through an optical circuit switch to shared pools of GPUs (PRU 2500s) and NVMe storage, using PCIe and RDMA over optical to compose resources dynamically across workloads.
The Server Model's Core Failure: Stranded Scarcity
Traditional servers lock resources into fixed proportions. A node comes with a specific mix of CPUs, GPUs, memory, SSDs, and NICs, and once deployed those ratios are essentially frozen. Any mismatch between workload needs and that fixed shape creates stranded capacity: idle GPUs when the job is memory-bound, idle memory and SSDs when the job is accelerator-bound, capital trapped inside boxes instead of serving the broader pool.
In the past, this was tolerable because components were easier to source, power was available, and overprovisioning could be written off as the price of reliability and growth. Today, that assumption breaks down. AI workloads make the utilization problem visible: a training job, an inference service, a vector database, and a data preprocessing pipeline all consume infrastructure in different proportions, but static servers force them into the same hardware shape. The result is not optimized infrastructure; it is a compromise that wastes the very resources operators can least afford to waste.
Scarcity is Now System-Wide
Shortages are no longer just about GPUs. Memory, SSDs, CPUs, power delivery, cooling, and floor space are all becoming strategic constraints in AI data centers. The key point is not which part is scarce today; it is that the entire stack is scarcity-constrained and the specific bottleneck will keep shifting over time.
The design objective therefore changes. The question is no longer "How do we build the biggest possible server?" but "How do we keep every scarce resource doing useful work as much of the time as possible?" That is the question resource pooling is built to answer.
What Resource Pooling Actually Does
In a pooled architecture, CPUs, GPUs, memory, SSDs, and other components are no longer treated primarily as server-local assets. They become shared infrastructure elements that can be composed, reassigned, and reconfigured based on workload needs. Instead of overbuying capacity in every server to cover every possible scenario, operators deploy valuable resources into pools and allocate them where they create the most value at a given moment.
With PCIe over optical, this goes beyond incremental flexibility. Breaking the traditional 8-GPU-per-server building block becomes possible: host systems can be built with roughly three times that GPU capacity, creating large single GPU memory domains composed of 24–32 GPUs per CPU. Larger, shared building blocks reduce fragmentation and deliver better throughput and utilization than equivalent capacity scattered across three or four isolated servers.
Why Optical Switching and PCIe over Optical are the Unlock
Resource pooling depends on connecting workloads to pools with the same low-level semantics they would have if the resources were local, but with much greater topological freedom. That is why optical switching and PCIe over optical are central to the new architecture.
Electrical switching and fixed copper cabling were designed for relatively static topologies. They work well when clusters rarely change shape, but they are far less effective when operators need to shift GPUs, SSDs, and memory between workloads, tenants, and stages of the AI lifecycle without recabling or rebuilding clusters.
Optical switching changes that equation by creating a high-performance interconnect between resource pools. GPUs can be dynamically connected to the workloads that need them, storage can live in shared high-performance pools, and memory-rich or accelerator-rich configurations can be composed per job instead of being frozen in a bill of materials. For Corespan, the strategic value is not a better silo, but architectural freedom: infrastructure behaves like a dynamic system, not a warehouse of isolated servers.
At Corespan Systems, we are not building a network-based solution for computing resources, we are building an optical extension of the CPU and PCIe bus architecture. We are mapping PCIe serdes over optical and using the front-end ethernet network for scale up of bigger, dynamically changeable resource blocks.
Resources can be deployed once and then reused across changing workload profiles. Capacity can be shifted as demand evolves instead of remaining marooned in yesterday's configuration. Capital can be concentrated into high-value shared pools rather than scattered into overbuilt nodes.
AI has Compressed the Timeline
Even before the current AI surge, workloads were becoming more heterogeneous, hardware more specialized, and clusters more fragmented. Teams were already wrestling with stranded capacity, partial clusters, and faster refresh cycles. AI accelerated all of those trends and compressed the timeline for architectural change.
AI workloads are not uniform. Training and inference behave differently. Large language models, multimodal models, recommendation systems, scientific computing, and enterprise AI all stress different combinations of compute, memory, storage, and networking. Within a single workflow, the resource mix shifts: data preparation leans on storage and CPU, training leans on GPUs and interconnect, fine-tuning uses a smaller but still specialized slice, and inference demands scale-out deployment, latency control, and efficient accelerator sharing. Static servers force operators to size for too many worst cases at once; pooled infrastructure lets them match resources to the actual demands of each phase.
Capital Efficiency as a First-Class Requirement
In the era of abundance, the fastest way to solve problems was often to buy more servers. Overprovisioning could hide inside growth, and capital was cheap enough that low utilization did not always trigger architectural change. That era is ending.
When GPUs are scarce, every idle accelerator is a financial problem. When memory and SSDs are constrained, every stranded DIMM, HBM stack, or enterprise drive matters. When power is the bottleneck, underutilized equipment consumes capacity that could serve productive work instead.
Resource pooling directly addresses this capital efficiency problem by improving utilization before adding hardware. It reduces the need to duplicate expensive components across many fixed server configurations and allows capacity to be deployed in larger, more flexible pools that can be assigned dynamically. This is not just a technical gain; it is financial discipline embedded into the architecture. The next generation of AI infrastructure will be judged on performance per dollar, performance per watt, utilization per rack, and time to useful deployment—and pooling helps on all four dimensions by making infrastructure less wasteful and more adaptable.
Dynamic Reconfiguration as a Competitive Advantage
The organizations that win in AI infrastructure will not simply buy the most hardware; they will put scarce hardware to work fastest, keep it utilized, and reconfigure it as workloads change. Achieving that requires a different operating model.
Infrastructure must be composable, with resource allocation becoming dynamic rather than static. Optical switching should sit at the center of architectures that can shift between workload profiles without physical recabling, cluster rebuilds, or months-long procurement cycles. In practical terms, AI platforms should be able to create the right system shape for each workload—more GPUs for one job, more SSDs for another, different connectivity patterns as demand evolves—using the same underlying pools of hardware.
Pooling does not eliminate scarcity; it makes scarcity manageable. It gives operators a way to extract more productive work from every scarce asset they acquire and creates a more resilient foundation for future hardware cycles, where tomorrow's bottleneck may not match today's.
The Future: Pooled, Optical and Workload Driven
The server-centric model assumed fixed ratios were acceptable, stranded capacity was manageable, and scaling meant adding more boxes. AI infrastructure is forcing a different answer. The next architecture must be resource-centric, not server-centric; it must organize around pools of valuable components connected by high-performance optical switching and support dynamic reconfiguration, high utilization, and efficient capital deployment.
Corespan Systems sees this as the next major phase of infrastructure design. As CPU, GPU, memory, SSD, power, and capital constraints intensify, the industry will have to stop treating overprovisioning as normal. Scarce resources cannot be overallocated to individual servers and left underused; they need to be pooled, shared, and assigned with precision. The shift from plenty to scarcity is already underway. The infrastructure response must be just as significant: pooled resources, optical switching, dynamic composition, and architectures built to make every critical component count.