The photonic infrastructure market is making a familiar mistake. It is treating the optical switch as if it is where durable value will accrue. It is not.
The switch will matter. It may even be scarce, but scarcity is not a moat, and essential does not mean differentiated. The enduring value in photonic infrastructure will sit above the glass and silicon: in the software that understands resources, composes them into systems, and decides what should connect to what.
The optical layer moves bits. The control plane determines whether those bits create a useful computer.
The Wrong Bet
Too much of the market is capitalizing the physical layer as though photonics will produce a new category of vertically integrated, hardware-defined winners. That is pattern-matching to the wrong precedent.
The better question is not, "Who makes the best optical switch?" It is, "Who can make a heterogeneous fleet of optical switches, accelerators, memory, and storage behave like one adaptable system?"
Those are different businesses. One sells a component. The other owns the operating logic of the data center. Hardware will continue to improve, but once interfaces stabilize and multiple credible suppliers exist, differentiation at the device layer narrows. That is what infrastructure markets do. They reward the layer that absorbs complexity for everyone else.
The Familiar Pattern
We have seen this movie across every major fabric wave. Ethernet began as a collection of proprietary implementations, distinctive boxes, and fiercely defended hardware positions. Over time, interfaces standardized, merchant silicon advanced, and the switch became broadly interchangeable. The value did not disappear. It moved into network operating systems, automation, policy, observability, and the software that made large fleets manageable.
InfiniBand followed a related path. Its performance mattered, and still matters, but the strategic question for operators was never just bandwidth or hop count. It was how to provision a fabric, schedule jobs across it, isolate faults, and keep an application from experiencing topology as an operational constraint.
Fibre Channel provides an even starker example. The fabric was vital to enterprise storage, yet the durable control point was not simply the switch. It was the software and operational model that mapped workloads to storage, enforced policy, and made capacity consumable without exposing the fabric's internals to every application team.
In each case, the physical network became a platform feature. The control layer became the leverage point. Photonics will not escape this pattern. It will compress it.
Faster This Time
The physical layer of photonics is unusually susceptible to standardization. Optical modules already live in an ecosystem shaped by common form factors, multi-source agreements, and interoperable supply chains. Demanding engineering does not guarantee enduring pricing power; well-specified challenges attract capable suppliers.
The same dynamic is taking shape around optical switching. There are multiple approaches to switch fabrics, multiple packaging paths, and multiple co-packaged optics roadmaps. That diversity is the real innovation. It is also a signal that no single implementation can assume it will define the market forever.
The differentiation surface is thin. A switch can improve latency, density, power, reliability, or manufacturability. Those are meaningful advantages, but increasingly specified requirements. Once buyers can compare those requirements, the switch becomes negotiable.
Software does not become negotiable in the same way. The system that knows the state of every resource, the intent of every workload, and the consequences of every topology decision accumulates operational context. That context compounds.
Value Above Optics
The hard problem is not opening an optical path. The hard problem is deciding which path should exist, for how long, with which performance guarantees, and what should happen when a component or workload changes state.
Consider what a modern AI environment actually asks of infrastructure. A training run may need a particular group of GPUs with high-bandwidth access to a memory pool. An inference fleet may favor locality, fault containment, and rapid reconfiguration. A batch job may accept a different topology if it can start sooner. The right answer is not static.
It depends on workload intent, resource availability, topology, congestion, maintenance state, policy, and economics. It depends on whether an operator wants to optimize time to completion, utilization, energy, isolation, or recovery. No optical switch can answer those questions by itself. That is a control-plane problem.
At Corespan, this is the premise behind Corespan Composer: infrastructure should be defined as software, then composed from available resources without forcing the workload to understand the physical arrangement beneath it. The control plane must be vendor-neutral, generation-neutral, and capable of treating the fabric as a programmable substrate rather than a fixed wiring diagram.
This is where value accrues because this is where choices are made. If the control plane can make dissimilar components behave as a coherent system, it becomes more valuable as the underlying hardware ecosystem becomes more diverse.
PCIe Is the Test
PCIe remoting is where the distinction becomes impossible to ignore. Moving a packet through an optical fabric is one thing. Extending PCIe semantics across physical distance, while keeping the experience acceptable to the operating system and workload, is another. The system has to respect latency budgets, preserve expected device behavior, manage ordering and coherence expectations where they apply, and expose failures in a way that does not turn a transient fabric event into an application disaster.
That is not a switch feature. It is systems software backed by deep hardware knowledge.
In a composable environment, a remote device cannot merely be reachable. It must be discoverable, attachable, isolatable, observable, recoverable, and governable. Software must understand failure domains and decide whether to reconnect, fail over, drain, or preserve workload state.
This needs to be done transparently. The best composable infrastructure does not ask every framework, operating system, or application to become fabric-aware. It makes remote resources feel native enough that operators gain flexibility without handing developers a new class of infrastructure exceptions.
That is why Corespan's Photonic Resource Unit (PRU) 2500 matters only in the broader system context. The strategic value rises when the control plane can place its GPU or SSD capacity where it is needed, connect the appropriate topology, and manage its lifecycle without binding the customer to one vendor or one generation.
The IP compounds in that translation layer: from workload intent to resource composition, from physical events to software policy, and from heterogeneous hardware to a predictable operating model.
A useful analogy comes from IP networking. A single physical network carries many VLANs, and endpoints attach to the LAN that has the resources they need. In a multi-domain PCIe fabric the same idea applies below the operating system: the CPU is the VLAN, and it attaches to different pools of accelerators and SSDs today, and to pools of memory and Gen6 PCIe devices tomorrow — all over the same physical interconnect. The optical layer supplies the any-to-any reach. The control plane decides which CPU sees which pool, for how long, and under what policy.
This is why multi-domain PCIe matters. In the Corespan fabric, the CPU behaves like a VLAN on a shared physical network: several hosts share one PRU pool at the same time, each isolated in its own logical PCIe domain, each seeing only the accelerators, SSDs, and — on Gen6 — memory devices Composer has bound to it. Isolation, routing, and composition are properties of the control plane, not properties of the switch.
The Investor Read
Investors should be wary of confusing a major capital-spending cycle with a durable control point. Optical switching will attract investment because it addresses a real scaling problem. The relevant question is where margin and strategic leverage remain after multiple suppliers qualify and customers treat the device as an input rather than an architecture.
That answer is unlikely to be the bare switch.
The companies with durable economics will own the abstraction layer that reduces integration risk, increases utilization, and keeps customers free to adopt the next hardware generation. They will make the data center more modular without making it more operationally fragile.
That is a more defensible position than owning a particular optical implementation. It benefits from hardware progress instead of needing to outrun it.
Above The Pipe
The temptation is to pattern-match photonic infrastructure to Nvidia: a new hardware primitive, a constrained supply chain, and a critical component of the AI buildout.
The more useful pattern match is VMware, Kubernetes, and the software-defined networking wave. Those categories created enduring value by abstracting physical resources, enforcing policy, and turning heterogeneous infrastructure into an operating environment. Photonics will be foundational, but that foundation is not the same as moat.
The winning companies will not merely move light faster, better, cheaper — they will decide how light, compute, memory, and storage assemble into a machine, a Host, the server. The pipe matters. The value lives above it.