Credo SRMO Standardization Challenges the Fastest-Lane Race in AI Networking

Credo launched an SRMO standardization effort on October 12, challenging the assumption that AI networks must keep pushing fewer optical lanes to higher speeds. The Short-Reach Micro-Emitter Optics workstream proposes another path. It would move data across large arrays of slower, micron-scale light emitters operating in parallel.

The company is not introducing a finished standard. It has started an Open Compute Project workstream whose participants plan to develop one for approval in 2027. That distinction matters because adoption depends on more than an emitter. Modules, fibers, connectors, receivers, signaling, compliance tests, and system designs must work together.

The main opponent is the narrow-and-fast architecture used by conventional optical links. That approach concentrates bandwidth into a limited number of increasingly fast channels. SRMO spreads the traffic across hundreds or potentially thousands of channels, creating redundancy while using simpler links.

This is more than a new connector proposal. Credo SRMO standardization asks whether AI infrastructure should optimize individual lanes or redesign the entire optical link around parallelism.

Credo SRMO Standardization Starts With an Ecosystem, Not a Product

Credo has launched a standards process designed to turn wide, parallel optics into an interoperable product category.

The company announced the SRMO workstream with an unnamed hyperscaler and seven identified optical technology providers. They are AFL, ams OSRAM, BizLink, Foxlink, Heraeus Covantics, SENKO, and YOFC.

Those participants cover different portions of the optical connectivity supply chain. Their specialties include light emitters, fibers, cables, connectors, materials, manufacturing, and system integration. That breadth is important because no single component defines a working parallel optical link.

The official announcement says the group will work within the Open Compute Project Networking community. It targets OCP approval for the specification in 2027.

The planned scope includes link power, efficiency, signaling, parallelism, link budgets, physical form factors, module footprints, connector footprints, and compliance. That list exposes the real purpose of the project.

SRMO does not merely need a definition of how light carries bits. It needs common rules governing how equipment from different vendors connects, performs, and passes qualification.

An optical link has a link budget, meaning the permitted gains and losses between the transmitter and receiver. Suppliers cannot build interoperable components unless they agree on those limits. The same problem applies to mechanical tolerances, channel arrangements, connector geometry, and signal behavior.

The workstream is responding to an expected proliferation of proprietary implementations. Companies can build micro-emitter prototypes around different fibers, channel counts, layouts, and packaging methods. Those products might perform well while remaining incompatible with one another.

Credo argues that such fragmentation would slow deployment. Buyers could become dependent on vertically integrated designs or tightly controlled supplier combinations. Component replacements would also become harder if connectors and optical interfaces differ.

The OCP workstream page frames interoperability as the central objective. It describes arrays of micron-scale light sources operating concurrently instead of driving a few channels at extreme rates.

This architecture can reserve spare channels at the physical layer. If an individual emitter fails, another channel can take its place without disabling the complete link. The workstream calls this layer-zero redundancy.

Layer zero refers to the physical elements below conventional networking protocols. Redundancy at this level can address component failure before higher layers must reroute traffic or recover a connection.

Credo executive L.K. Bhupathi said the industry needs a standardized approach rather than components that do not fit together. The statement reflects an immediate commercial concern. A new link architecture has limited value if every deployment requires custom integration.

The announcement therefore changes the status of micro-emitter optics. It moves the subject from isolated research and vendor development toward an organized attempt at ecosystem formation.

That does not make SRMO an approved standard. It makes the requirements, boundaries, and interfaces open for negotiation. The resulting decisions will determine which early products fit the category and which must adapt.

Why AI Scale-Up and Scale-Out Networks Need Another Option

AI systems are turning connectivity into a design constraint, not a secondary component choice.

Scale-up networks connect accelerators so closely that they can behave like a larger computing system. These links demand high bandwidth and low latency across servers, trays, racks, or rows.

Scale-out networks connect more systems to expand the overall cluster. They support traffic between servers and racks, often across longer distances and through more network equipment.

Both architectures are growing in physical size and connection count. More accelerators create more links, and every link adds power consumption, heat, potential failures, and deployment complexity.

Copper remains attractive at short distances because it is familiar, efficient, and reliable. However, signal loss increases as data rates and cable lengths rise. Engineers must add equalization and signal-conditioning electronics, which consume power and generate heat.

Conventional optics extends reach, but it introduces its own tradeoffs. Laser-based links use a small number of high-speed channels. Each generation pushes those channels toward faster signaling and increasingly complex electronics.

Microsoft Research describes this conflict as a choice among reach, power, and reliability. Its Mosaic research says copper links are efficient but generally limited to less than two meters in demanding applications. Conventional optics reaches farther but consumes more power and carries different reliability concerns.

The narrow-and-fast model developed for environments where fiber count was expensive. Long-distance networks benefit from sending more information through each optical channel. A limited number of cables must carry traffic across difficult routes.

An AI data center presents a different optimization problem. Links remain within a controlled facility, while connection density, power, heat, serviceability, and failure rates carry greater weight. Maximizing throughput on every individual optical channel may not produce the best system.

This pressure is especially intense in scale-up designs. Accelerators exchange large volumes of data during tightly coordinated calculations. A network delay or unstable link can leave costly compute resources waiting.

Failures also become harder to ignore as clusters expand. Even a low per-link failure rate produces more incidents when multiplied across a large fabric. A link architecture that tolerates individual emitter failures could improve system availability.

Micro-emitter optics attempts to occupy the space between copper and conventional optics. It seeks longer reach than short copper links while avoiding some complex, high-speed optical electronics.

The concept uses many small emitters rather than several lasers running at very high rates. Each channel carries less data, but the channels provide high aggregate bandwidth together.

This approach is often described as wide and slow. “Slow” is relative to current high-speed optical lanes. Hundreds of channels operating at lower individual rates can still produce an extremely fast aggregate connection.

The architecture also matches the parallel nature of AI computation. Accelerators already process many operations concurrently. A parallel optical interface can carry multiple streams without forcing all traffic through a few heavily serialized lanes.

Serialization converts parallel data into a smaller number of high-speed streams. Deserialization reverses the process at the destination. Those functions require circuitry and can add power, latency, and design complexity.

Removing every serializer is not guaranteed across all SRMO implementations. The final specification has not defined the complete electrical and optical architecture. However, lower per-channel rates can reduce the signal-processing burden.

That possibility creates pressure for existing copper, pluggable-optics, silicon-photonics, and co-packaged-optics road maps. SRMO does not automatically replace them, but it adds another candidate for rack and row-scale connections.

Suppliers must now determine whether wide, parallel optics serves a distinct reach range or competes directly with their planned products. System builders must decide whether its power and reliability claims justify new physical interfaces.

Wide and Slow Optics Changes Where Complexity Lives

SRMO does not eliminate engineering complexity; it redistributes complexity from lane speed into channel count, packaging, alignment, and connectivity.

Traditional optical development often focuses on increasing the rate of each channel. That strategy requires faster transmitters, receivers, converters, signal processors, and error-management systems.

The wide-and-slow architecture moves in the opposite direction. It reduces the rate of individual channels and increases their number. That can simplify the electronics attached to each lane, but it creates a dense physical assembly.

Microsoft’s Mosaic prototype illustrates the mechanism. Researchers demonstrated 100 optical channels, with each channel carrying 2 gigabits per second. The team described a path to 800 gigabits per second and beyond, with reach up to 50 meters.

The published results reported up to 68 percent lower power consumption than the optical comparison and 100 times higher reliability. They also reported ten times the reach of the copper comparison.

Those results belong to Microsoft’s research system, not the unfinished OCP specification. They show why the architecture attracts interest, but they do not validate every future SRMO product.

The prototype used directly modulated microLEDs, multicore imaging fiber, and a low-power analog backend. A microLED is a microscopic light-emitting diode that can switch rapidly enough to transmit data.

MicroLED arrays can place numerous light sources within a small area. Their size makes high channel density possible, while their relatively low per-emitter cost supports redundant capacity.

Parallel redundancy changes the failure model. A conventional link can lose substantial capacity when one of a few critical emitters fails. A wide link can reserve additional lanes and reroute data around an isolated channel failure.

The system still needs to detect a failure and manage the spare channel. SRMO participants must decide how redundancy operates, how much reserve capacity is required, and how equipment reports degraded performance.

ams OSRAM has offered another useful data point. Internal studies cited in an Optica industry article reported 3 gigabits per second per emitter over a 10-meter link.

The company reported energy below 2 picojoules per bit for that link and a bit-error rate below 10^-15. These are company study results rather than independent qualification across production data center systems.

The comparison still helps define the proposed operating range. SRMO is not trying to make one micro-emitter compete with a 200-gigabit optical lane. Its advantage depends on deploying many lower-rate emitters together.

That decision places heavy demands on packaging. Thousands of optical channels need accurate alignment among emitters, receivers, fibers, waveguides, lenses, and connectors. Small positioning errors can reduce coupling efficiency or disable channels.

Connector design therefore becomes part of the central technical problem. SENKO executive Obie Roy said power and density benefits depend on getting the physical connector interface right as channel counts reach thousands.

Manufacturing yield presents another issue. A design may tolerate some defective emitters, but production still requires repeatable placement, bonding, inspection, and testing. Suppliers need economical ways to identify failures before final assembly.

Fiber choice also matters. Conventional silica fiber is optimized for established optical wavelengths. Some visible-light micro-emitter approaches require different fibers or waveguides whose data center durability needs further validation.

Thermal behavior creates another test. Lower energy per bit can reduce heat, but densely packed emitters and receivers still sit near high-power compute equipment. Packaging must maintain alignment and performance across temperature changes.

Cleaning and service procedures could differ from familiar optical connections. A connector with hundreds or thousands of spatial channels may react differently to contamination, scratches, vibration, or repeated insertion.

The final standard must balance density against field usability. A highly compact connector offers little operational value if technicians cannot inspect, clean, replace, or qualify it reliably.

This is the main mechanism behind Credo SRMO standardization. Shared rules can prevent each supplier from solving those physical problems through incompatible choices.

A defined footprint lets cable and connector vendors design around a common target. Agreed link budgets give emitter and receiver companies known performance limits. Compliance procedures give buyers a basis for comparing products.

The architecture therefore trades one scaling challenge for several coordination challenges. It reduces pressure on individual lane speed while increasing pressure on ecosystem alignment.

The Standard Still Has to Survive Manufacturing and Deployment

The largest uncertainty is not whether micro-emitter links can move data, but whether vendors can manufacture and operate them at data center scale.

Credo’s announcement contains ambitious claims about reliability, efficiency, bandwidth density, and deployment speed. Those claims describe the workstream’s goals. They are not results from a completed standard or broad commercial rollout.

The OCP process targets approval in 2027. Until the group publishes drafts, outsiders cannot evaluate the exact channel arrangement, optical medium, connector geometry, redundancy policy, or compliance thresholds.

The unnamed hyperscaler also limits visibility. Its participation suggests operator interest, but the announcement does not identify planned deployments, purchase commitments, qualification schedules, or supported platforms.

Standards participation can serve several purposes. Companies may want to influence specifications, monitor a technology, preserve supply options, or prepare for future procurement. Membership alone does not guarantee adoption.

The participant list is broad, but important categories remain unclear. The announcement does not name accelerator vendors, switch-chip companies, server manufacturers, or major conventional transceiver suppliers.

Their involvement would matter because an interconnect cannot succeed on optical components alone. The electrical interface, protocol mapping, management software, thermal design, and system qualification must fit accelerator and switch road maps.

The group must also decide how SRMO relates to Ethernet, UALink, and other fabrics. A protocol-agnostic physical link can support several standards, but each integration introduces timing, error handling, telemetry, and compliance requirements.

Existing alternatives will not stand still. Copper cable suppliers continue improving active electrical cables and retimers. Conventional optics is moving toward 1.6-terabit products, linear designs, and lower-power digital signal processing.

Silicon photonics companies are integrating more functions into compact optical engines. Co-packaged optics places optical components near switching silicon, reducing the length of high-speed electrical connections.

These routes solve different portions of the problem. Copper offers operational familiarity at shorter reach. Pluggable optics supports established service models. Co-packaged optics pursues high bandwidth density near chips and switches.

SRMO needs a clear operating zone where its combined benefits exceed its adoption costs. A theoretical gain in energy per bit will not suffice if packaging, cabling, connector handling, or qualification becomes too expensive.

Microsoft’s work shows promising prototype performance, yet its paper also describes adaptation and ecosystem challenges. Research demonstrations do not automatically deliver production yield, long service life, or multiple qualified suppliers.

An IEEE Computer Society overview of microLED barriers identifies unresolved questions around optical materials, thermal performance, and long-term reliability. Those are direct concerns for high-volume infrastructure.

Component redundancy also needs careful interpretation. Spare channels can protect against an emitter failure, but they do not remove every single point of failure. Driver circuits, receivers, connectors, control electronics, and complete modules can still fail.

System reliability depends on the entire assembly. Vendors will need evidence that graceful channel degradation produces better uptime than mature alternatives under real operating conditions.

Testing may become a significant cost. A transceiver with hundreds of channels requires efficient methods to verify alignment, optical power, error rates, spare capacity, and behavior across temperature.

Compliance must remain rigorous without making qualification impractical. A weak test regime could allow nominally compatible products to fail together. An overly complex regime could slow supplier entry and raise costs.

Supply-chain diversity is another stated goal. A standard can reduce dependence on one vendor, but only if multiple suppliers can build interchangeable parts without relying on one company’s protected implementation.

Intellectual property terms, reference designs, and access to manufacturing processes will shape that outcome. OCP approval does not automatically create a competitive market.

Credo also has a commercial position in the result. Its portfolio includes electrical cables, retimers, optical components, and diagnostic software. A new link category can expand its addressable role across AI infrastructure.

That interest does not invalidate the proposal. It means buyers should distinguish ecosystem benefits from vendor forecasts. Independent interoperability events and operator qualification will provide stronger evidence.

The right question is not whether SRMO sounds better than existing optics. It is whether a complete, multi-vendor link can achieve its claimed advantages after manufacturing, installation, and years of operation.

Three Signals Will Show Whether SRMO Becomes a Real Standard

The next year must produce a concrete specification, interoperable hardware, and public operator validation.

The first signal is a detailed OCP draft with measurable requirements. The workstream needs to define channel organization, reach classes, link budgets, signaling, power limits, form factors, connector behavior, and compliance methods.

A draft would reveal whether participants agree on one architecture or merely share a broad concept. Clear mandatory requirements would strengthen the case that SRMO can support interchangeable components.

Vague profiles or numerous vendor options would weaken that case. Too much flexibility can preserve innovation, but it can also recreate the fragmentation the workstream was formed to prevent.

The second signal is a multi-vendor interoperability demonstration. It should connect emitters, receivers, fibers, connectors, modules, and host equipment from different participants.

A useful demonstration would report aggregate bandwidth, reach, energy per bit, error behavior, thermal conditions, and response to channel failures. It should also show replacement or substitution across suppliers.

A single vertically integrated prototype would prove less. It could validate the underlying technology while leaving the interoperability question unanswered.

The strongest demonstration would include compliance testing against an openly described specification. That would connect laboratory performance to the standardization effort.

The third signal is hyperscaler qualification. The unnamed operator or another major buyer needs to identify an intended deployment class, testing program, or procurement timeline.

Operator validation would show that SRMO solves a pressing infrastructure problem rather than an abstract engineering challenge. It would also clarify whether the first market is scale-up, scale-out, front-end networking, or another connection type.

A production commitment is not necessary immediately. Published qualification criteria, deployment trials, or reference architectures would still provide valuable evidence.

Conversely, silence from system buyers through 2027 would weaken the commercial story. Optical suppliers can standardize interfaces, but demand must come from the organizations building AI clusters.

Credo SRMO standardization should also be judged against changing alternatives. If copper extends farther at acceptable power, the opening for SRMO narrows. If conventional optics improves reliability and efficiency faster, switching costs become harder to justify.

A successful SRMO standard would not need to replace every cable. It could win a defined range where copper becomes strained and conventional optics carries unnecessary complexity.

That outcome would give AI system designers another architectural option. They could spread accelerators across racks or rows without accepting the same power and reliability compromises.

The broader effect would be a change in how the industry measures optical progress. Aggregate bandwidth, redundancy, energy, and serviceability could matter more than the highest individual lane rate.

For infrastructure teams, this is the moment to track specifications rather than marketing labels. Ask whether vendors disclose complete link measurements, interoperable components, failure behavior, and qualification conditions.

Developers and AI users will not select these cables directly. They will still feel the outcome through cluster availability, inference latency, capacity, and the cost of moving data between accelerators.

Watch the 2027 OCP process, the first cross-vendor demonstrations, and hyperscaler qualification announcements. Together, those signals will show whether wide-and-slow optics becomes infrastructure or remains an appealing laboratory result.