By Stuart Dee
The future of data centres is often framed through dramatic and attention-grabbing concepts such as cryogenic cooling, orbital compute platforms, or even data centres in space. These ideas are technically fascinating, yet they share a common flaw. They relocate today’s architectural constraints into environments that introduce far greater complexity, cost, and operational risk. They attempt to solve symptoms such as power draw, heat, and inefficiency without addressing the underlying cause.
The real pressure on data centres does not come from their location. It comes from the long-standing assumption that compute, memory, and storage must be tightly located together. For decades, this assumption has shaped everything from server design to facility layout. That assumption is now dissolving.
The End of the Locality Assumption
For more than thirty years, data centre architecture has been governed by a simple model. Compute is local, memory is tightly attached, and the network is a slow and unreliable boundary. This model survived the rise of virtualisation, cloud computing, and hyperscale growth. However, artificial intelligence, real time analytics, and data intensive workloads have exposed a fundamental truth. Memory, not compute, is the dominant constraint.
At the same time, advances in optical networking are redefining what the word remote really means. Latency budgets that once made memory disaggregation impossible are being rewritten by photonics. The result is a shift as significant as the move from mainframes to distributed computing.
This shift is embodied in the Innovative Optical and Wireless Network, known as IOWN. Developed through the IOWN Global Forum, it is not simply a faster network. It is a rethinking of the entire relationship between compute and memory, built on the premise that photonics can dissolve the traditional boundaries that define system architecture.
When combined with software defined memory, IOWN enables a radical outcome. Data centres can operate primarily as network addressable memory arrays, with compute located elsewhere.
From Electrical Boundaries to Optical Continuity
Although modern networks rely heavily on fibre, they remain fundamentally electronic systems. Every switch, router, and server interface converts signals between optical and electrical domains. At today’s data rates, these conversions introduce significant latency, jitter, heat, and power consumption. They are no longer minor inefficiencies. They are architectural bottlenecks.
IOWN proposes an all optical approach in which optical paths are maintained from end to end wherever possible. Instead of relying on packet switched, best effort networking, optical paths are provisioned explicitly, delivering predictable performance with minimal conversion overhead.
The implications are profound.
- Latency becomes predictable rather than variable.
- Energy consumption per bit drops dramatically.
- Network distance becomes far less relevant to workload design.
Crucially, the IOWN roadmap extends photonics beyond transport networks and into boards, packages, and eventually communication between chips. As optics move closer to the processor, the distinction between inside the server and across the network begins to disappear.
Data Centric Infrastructure and the Decoupling of Memory
IOWN’s architectural expression is often described as data centric infrastructure. Instead of treating servers as fixed bundles of compute, memory, and storage, resources are separated and composed dynamically over an optical fabric. Compute, accelerators, storage, and memory become independent pools, allocated according to workload needs rather than physical proximity.
This aligns directly with the long-standing ambition of software defined memory. Today, memory is routinely over provisioned because it is statically bound to servers. Meanwhile, memory intensive workloads are constrained by local capacity limits, even when unused memory exists elsewhere in the estate.
Historically, memory disaggregation has failed because of latency. Even microseconds of additional delay can break memory semantics. Optical interconnects change that equation. Photonic links offer high bandwidth, low energy per bit, and far better scaling characteristics than electrical alternatives. This makes it feasible to place memory at rack, data centre, or even regional distances without prohibitive penalties.
The Emergence of the Optical Memory Plane
IOWN enables what can be described as an optical memory plane, a fabric in which memory pools are accessed over predictable optical paths rather than best effort networks. In this model, the network becomes the memory backplane.
A data centre can therefore be optimised almost entirely around dense and energy efficient memory arrays. These memory centric facilities can service multiple compute centric data centres, accelerator clusters, or edge locations. To software, remote memory appears as an extension of system memory, orchestrated and tiered by software defined control planes.
As photonics moves closer to the processor package, the latency gap between local and remote memory continues to shrink. What was once a hard architectural boundary becomes a policy decision. Where should memory reside to optimise performance, cost, energy efficiency, and resilience?
Why Software Defined Memory Benefits from IOWN
Software defined memory has always been constrained less by software maturity than by physics. Control planes can allocate and migrate memory with remarkable sophistication, but they cannot compensate for unpredictable latency or congestion in traditional networks.
IOWN addresses these constraints directly.
- Latency becomes predictable through provisioned optical paths that avoid contention and jitter.
- Distance becomes energy efficient, making remote memory economically viable at scale.
- Resource sharing becomes practical across sites, enabling metropolitan scale memory fabrics rather than rack scale compromises.
With these capabilities, software defined memory becomes not a workaround but a first class architectural principle. Instead of working around the limitations of electrical networks, it can fully exploit the continuity and determinism of photonic fabrics.
The Memory Centric Data Centre as a Service
The logical end state is a federated model in which some data centres are optimised for compute density, while others are optimised for memory density, all interconnected by IOWN style optical fabrics.
Artificial intelligence training clusters may expand into remote memory pools for model parameters. Edge inference sites may access centralised memory without local replication. Memory can be located where power, cooling, and carbon intensity are optimal, while compute sits closer to users or data sources.
This inversion of the traditional model transforms data centres from monolithic units into specialised components within a larger optical system.
Rethinking Infrastructure Design
IOWN does not merely promise faster networks. It enables a shift from server centric to memory centric infrastructure, where distance is abstracted by optics and resources are defined by software.
In that world, a data centre does not need to resemble a compute factory. It can exist primarily as a shared memory service, tightly coupled to remote compute through optical continuity rather than electrical compromise. As software defined memory matures and photonics continues its progression into the compute stack, this architecture moves from speculative to inevitable.