FS D5130 Optical Switch Expands, but Scale Is the Real Test

FS expanded its D5130 optical switch family on October 8, adding multimode and customizable models to a lineup capped at 32×32 ports. The FS D5130 optical switch targets AI clusters that need physical connections to change as training, inference, and testing workloads move.

That expansion places FS inside a consequential shift in data center networking. Optical circuit switches route light directly between fibers, avoiding an optical-electrical-optical conversion inside the switching layer. FS says this design reduces adaptation complexity while supporting traffic from 100G through 800G and beyond.

The tension lies in what the release does not establish. The D5130 reaches more deployment types, but FS has not published independent performance tests, production references, or large-cluster results. Competitors already promote switches with hundreds of ports, while Google has operated a custom optical switching layer at data center scale for years.

This is therefore not a simple product refresh. It is a test of whether smaller, commercially accessible optical switches can move from laboratories and specialized links into operational AI infrastructure.

What FS Added to the D5130 Lineup

FS broadened the D5130 family across fiber types and configurations, but it did not increase the platform beyond 32×32 ports.

The company’s October 8 product announcement renamed the former DCS-W series and introduced an expanded set of D5130 configurations. The family now includes single-mode, multimode, and customizable versions.

Standard matrix sizes include 8×8, 16×16, and 32×32 configurations. An 8×8 switch can connect eight optical inputs with eight outputs, while preserving any-to-any path selection within that matrix.

FS lists three single-mode products and three corresponding multimode products. It also offers a configurable D5130-W-C model, although custom systems remain limited to 32×32 in the published specifications.

That choice gives the FS optical switch lineup several possible roles. Small matrices can support optical laboratories, component testing, equipment validation, and limited production interconnects. Larger models can sit between network devices in an AI or high-performance computing environment.

The switch uses a non-blocking architecture. In this context, non-blocking means any available input can connect to any available output without another active circuit preventing the requested path.

FS also describes the D5130 as protocol and rate transparent. The switch steers an optical signal rather than reading Ethernet, InfiniBand, or another packet format. That approach allows the same optical path to carry different protocols and data rates.

Published specifications cover 100G through 800G and higher speeds. However, this wording describes optical transparency, not a validated benchmark at every possible rate. The transceivers, fiber type, link budget, and connected systems still determine end-to-end performance.

The product includes dual redundant AC or DC power supplies. A browser interface handles configuration, optical-path monitoring, and alarms. Selected configurations add optical power monitoring, which helps operators detect signal degradation across an active path.

FS also claims support for dark-fiber switching. Dark fiber means an optical path can be configured even when no light is present at the input. That capability matters for provisioning standby links, testing inactive circuits, and preparing failover routes.

Bidirectional path control expands the possible uses further. An operator could assign optical routes for testing, equipment sharing, network restoration, or changing cluster topologies without manually moving fibers.

Yet the release provides no customer deployment, switching-latency measurement, insertion-loss figure, or reliability result for the renamed family. Those omissions do not invalidate the specifications. They define the limits of what buyers can conclude from the announcement.

The resulting event is specific: FS expanded the available configurations of an existing optical circuit switch. It has not shown that the D5130 can replace the electrical switching fabric inside a large AI cluster.

Why AI Data Centers Want Reconfigurable Optical Paths

AI infrastructure makes optical circuit switching attractive because expensive accelerators often communicate in structured, bandwidth-heavy patterns.

A conventional electrical packet switch examines and forwards packets through electronic switching silicon. It can react quickly to changing traffic, buffer contention, and direct individual packets toward different destinations.

An optical circuit switch performs a different job. It establishes a physical light path between ports, then leaves the transmitted data untouched. The circuit behaves more like a temporarily assigned fiber than a packet-processing device.

This distinction can remove an electrical switching stage from selected parts of a network. It can also reduce conversions, avoid dependence on a specific protocol, and preserve an optical path across changing interface generations.

The OCS white paper published by the Open Compute Project outlines several AI data center uses. These include replacing electrical spine switches, reconfiguring links after failures, and changing topology around workload demands.

Its spine-replacement example illustrates the energy argument. The paper models 64 electrical spine chassis consuming about 30 kilowatts each, producing a combined load near 1.9 megawatts.

That figure is a scenario, not a universal savings estimate. Actual consumption depends on switch utilization, optics, redundancy, cooling, topology, and the equipment replaced. Still, it shows why data center operators investigate optical alternatives.

AI workloads strengthen the case because their communication is often coordinated. Distributed training systems repeatedly exchange gradients, parameters, activations, and expert-routing traffic according to software-controlled schedules.

Operators can sometimes predict those exchanges more effectively than ordinary internet traffic. A scheduler can establish an optical path before a sustained transfer, then reconfigure it for the next communication phase.

Google provides the strongest production precedent. Its published Jupiter network history says the company used optical circuit switches and software-defined networking across most of its data center networks.

Google’s Apollo switch uses two arrays of tiny movable mirrors. Software rotates those mirrors to map an input fiber onto a selected output fiber. The network then reallocates bandwidth according to application demand.

This precedent proves that optical circuit switching can operate at data center scale. It does not prove that every commercial switch can reproduce Google’s result.

Google co-designed the switch, transceivers, circulators, control systems, and network architecture. Its engineers spent several generations balancing port count, optical performance, switching time, cost, and software behavior.

That integration burden explains why an off-the-shelf device is only one piece of the system. AI operators also need topology control, workload awareness, failure recovery, telemetry, and orchestration.

The FS optical circuit switch addresses the physical path layer. The planned AmpCon integration is intended to connect that layer with a broader management system. Until that integration ships, customers must evaluate how much orchestration they must build themselves.

How the FS D5130 Optical Switch Routes Light

The FS D5130 optical switch uses MEMS mirrors to redirect entire optical channels without processing their packets.

MEMS means microelectromechanical systems, which are microscopic mechanical structures controlled through electrical signals. Inside an optical switch, small mirrors tilt to guide a beam toward a chosen output.

A common three-dimensional MEMS design uses one mirror for an input and another for an output. Coordinated movement aligns the incoming beam with the selected destination fiber.

The signal remains optical while crossing the switch. There is no internal receiver converting the light into bits, followed by another transmitter recreating the optical signal.

That absence of optical-electrical-optical conversion gives the architecture its rate transparency. If the optical path and connected components support a new interface, the mirror fabric does not need to understand its encoding.

This feature can extend the useful life of a physical switching layer. A switch deployed around 400G links could potentially carry faster interfaces later, subject to its optical bandwidth and loss characteristics.

The same property creates operational freedom. One D5130 could carry Ethernet on one circuit and InfiniBand on another, provided the surrounding optics and endpoints meet the applicable requirements.

MEMS also provides a mechanical form of path selection. Unlike an electrical switch, it does not forward every packet independently. The mirrors hold a circuit until the control system requests another mapping.

Lumentum’s explanation of MEMS architecture shows why the method has attracted AI infrastructure suppliers. Its design uses two mirrors per optical connection and scales mirror count linearly with the number of ports.

Lumentum says its own MEMS switches typically keep insertion loss below 1.5 decibels without amplification. That figure belongs to Lumentum’s platform, not the D5130, because FS has not published an equivalent number in its announcement.

Insertion loss measures how much optical power disappears as light crosses a component. Higher loss reduces the remaining link budget and can constrain fiber reach, connector count, or transceiver choice.

FS describes the D5130 as a low-loss system, but that description needs a numerical specification for engineering decisions. Buyers should request guaranteed loss across every path, not only a typical or best-case route.

Switching time requires the same scrutiny. Mechanical mirrors need time to move and settle. A millisecond-scale transition may work for long-lived circuits but perform poorly when traffic changes at packet timescales.

The D5130 release calls switching fast without publishing a measured interval. That prevents a direct comparison with other MEMS products, silicon photonic switches, or electrical packet switches.

The right comparison also depends on the workload. A training job that holds one topology for minutes can tolerate more reconfiguration delay than interactive inference with rapidly shifting communication.

This is why the FS D5130 optical switch should not be described as a universal replacement for electrical networking. It offers a configurable optical layer that can complement electrical switches in selected architectures.

A hybrid network can keep electrical packet switching for bursty or unpredictable traffic. It can assign sustained, high-volume flows to optical circuits where avoiding another electrical stage provides a useful benefit.

That mechanism is the product’s strongest argument. Its success depends less on headline bandwidth than on whether software can identify the right flows and configure paths at the right moment.

The 32-Port Ceiling Defines FS’s Competitive Position

FS enters a market moving toward high-radix systems, so the D5130 competes on accessibility and deployment fit rather than maximum scale.

Radix is the number of ports available on a switch. Higher radix lets one device connect more endpoints or network groups without adding another switching layer.

The D5130 tops out at 32×32. That capacity can serve laboratories, validation systems, specialized optical paths, and smaller network domains. It is not comparable with platforms designed to connect hundreds of optical ports directly.

DiCon Fiberoptics announced 64×64 and 300×300 competing switches in March 2026. The company also outlined a 1024×1024 matrix planned for 2027.

DiCon says its new switches use the same three-dimensional MEMS technology as its existing 384×384 platform. It lists end-to-end loss below 3.0 decibels, although that remains a company-supplied specification.

Molex has separately introduced a high-radix optical circuit switching platform for AI infrastructure. Lumentum, Coherent, Calient, and other optical specialists also bring established switching or photonics experience.

These companies do not all target the same deployment. Some sell components to hyperscalers, while others address laboratory automation, telecom networks, or complete AI fabrics.

FS approaches the opportunity from a broader networking catalog. It already sells switches, transceivers, cabling, data center interconnect equipment, and management software. That portfolio can simplify purchasing for customers that do not design custom optical hardware.

The smaller matrix may also lower the organizational threshold for experimentation. A customer can test reconfigurable optical paths around a limited set of systems before committing to a fabric with hundreds of ports.

Multimode support strengthens that position. Multimode fiber commonly serves shorter data center links, while single-mode fiber supports longer reach and different optical designs. Offering both allows FS to address more existing environments.

Customization adds another route into specialized deployments. Laboratories may need unusual connectors, path behavior, monitoring, or fiber configurations that do not fit a fixed hyperscale product.

However, smaller matrices create scaling costs. Multiple switches require additional fibers, control coordination, redundancy, and topology planning. Those additions can erode the simplicity promised by optical switching.

A 32-port platform also cannot demonstrate the same consolidation potential as a switch with several hundred ports. Operators must calculate the total system, including management, cabling, spares, and failure domains.

FS has not disclosed D5130 customers, shipment volume, production capacity, or qualification results. It has also not identified an AI cluster size that the expanded family currently supports.

Those gaps matter because optical switching is not purchased through specifications alone. Buyers need confidence that mirror alignment remains stable, switching behavior stays predictable, and failures do not interrupt large accelerator pools.

The competitive question is therefore not whether FS has the largest switch. It clearly does not. The question is whether FS can make optical switching easier to test, acquire, and integrate for customers below the hyperscale tier.

Software Integration and Workload Fit Remain Unproven

The D5130’s main uncertainty is not whether its mirrors can connect fibers, but whether software can use those paths productively under real workloads.

FS says it plans to integrate the D5130 with AmpCon, its network management platform. The announcement does not provide a release date, feature list, supported topology, or migration process.

Today’s browser interface can configure paths and expose alarms. That is useful for manual operation and testing, but AI fabrics need faster coordination across compute jobs, network state, and failures.

A production control plane must know which devices require bandwidth and when they require it. It must prevent conflicting paths, detect degradation, and restore connectivity after a component fails.

It also needs to coordinate with an electrical network. Optical circuits cannot efficiently handle every short burst or unpredictable flow, so most practical designs preserve packet switching somewhere in the fabric.

Reconfiguration delay becomes decisive in that hybrid design. Recent research on inference thresholds found that an OCS may require sub-700-nanosecond reconfiguration to outperform electrical packet switching without substantial link fan-out.

That result applies to the modeled inference conditions, not every AI workload. It still demonstrates why the word “fast” is insufficient when suppliers discuss switching.

Many MEMS devices reconfigure on a millisecond timescale. Such products can still serve topology changes, scheduled training phases, protection paths, or test automation. They are less suited to chasing individual inference exchanges.

FS does not identify the D5130’s switching interval. Without that number, readers cannot determine which workload timescales the product can follow.

The same issue affects optical power monitoring. FS says monitoring appears on selected configurations, but the announcement does not describe measurement accuracy, sampling frequency, or controller access.

Operations teams will also want failure data. A mirror may support many switching cycles in controlled testing, yet production reliability includes power supplies, controllers, connectors, fibers, and calibration behavior.

The release lists redundant power options but provides no mean-time-between-failure figure. It also offers no independent shock, vibration, or long-duration stability result.

Protocol transparency presents another subtle tradeoff. The D5130 does not inspect packets, so it cannot provide packet-level buffering, congestion control, or routing decisions inside the optical layer.

Those functions remain with endpoints, electrical switches, or orchestration software. An optical path can carry traffic efficiently while still becoming underused, congested, or poorly scheduled.

AI communication patterns are structured, but they are not automatically static. Mixture-of-experts models, shared clusters, inference services, and job preemption can change traffic in ways that complicate circuit allocation.

Operators therefore need evidence beyond a successful optical connection. Useful validation should measure accelerator utilization, job completion time, network recovery, power consumption, and operational effort.

The absence of public benchmarks is the strongest reason for caution. FS has announced capabilities and configurations, not a quantified AI cluster outcome.

That distinction should guide procurement. The D5130 deserves evaluation where reconfigurable optical paths solve a defined problem. It should not be assumed to improve every AI network merely because it avoids one conversion layer.

Three Signals Will Show Whether the Expansion Matters

The next test is execution: management integration, disclosed deployments, and measured comparisons will determine the D5130’s relevance.

First, watch for the promised AmpCon integration. A credible release should expose topology discovery, programmable path control, monitoring, alarms, and documented interfaces for external orchestration.

The strongest version would connect optical reconfiguration with workload scheduling or data center automation. A basic dashboard update would provide convenience but leave the central coordination problem unresolved.

Shipping that integration would strengthen FS’s case as a systems supplier. A prolonged delay would keep the D5130 closer to a manually managed optical component.

Second, watch for a named deployment or technical case study. The important evidence would identify matrix size, workload, link rate, network role, and measurable operational result.

A laboratory demonstration can validate interoperability. A production reference can show whether the system handles failures, maintenance, changing jobs, and sustained use.

Buyers should look for results such as reduced electrical switching stages, lower network power, improved accelerator utilization, or faster recovery. Each metric needs a clear baseline and test method.

Third, watch how FS answers the scale gap. It could introduce matrices above 32×32, explain a modular design using several D5130 units, or focus firmly on smaller environments.

Any of those strategies can be coherent. The present announcement leaves the target boundary too broad by grouping laboratories, HPC systems, and AI factories under one product story.

Competitor activity will sharpen that choice. High-radix suppliers are increasing port counts, while silicon photonics developers pursue faster switching at different loss and manufacturing tradeoffs.

Electrical networking will continue advancing as well. Faster switch chips, co-packaged optics, and improved congestion control can preserve packet switching while changing where optical conversion occurs.

The FS optical switch lineup therefore enters a moving market, not an empty one. Optical circuits, electrical fabrics, and co-packaged optics can coexist because they solve different parts of the data-movement problem.

For infrastructure teams, the practical next step is to define the traffic pattern before evaluating the hardware. Identify which flows persist long enough to justify a circuit, then request switching-time and insertion-loss guarantees.

Ask how failures return traffic to the electrical fabric. Confirm which monitoring data can enter existing operations systems. Test the exact optics, fiber type, and protocol combinations planned for production.

The FS D5130 optical switch has made reconfigurable optical networking available in more configurations. Its real value will emerge only when measured deployments show where those 32 ports improve an AI system, and where they do not.