Credo Technology Group Holding Ltd (NASDAQ: CRDO) said on October 12, 2026, that it is launching a new standardization effort inside the Open Compute Project aimed at one of the thorniest bottlenecks in AI infrastructure: how to move data between tens of thousands of accelerators without the connections themselves becoming the limiting factor. The San Jose-based company announced the Short-Reach Micro-Emitter Optics Workstream, or SRMO, which will sit inside the OCP Networking Project and work toward a published specification by 2027, according to the company announcement carried on Financial Times Markets.
The announcement lands at a moment when AI data center operators are running out of easy answers for interconnect bandwidth. As clusters grow from thousands to tens of thousands of GPUs, the copper and optical links tying them together have to carry more data, over tighter power budgets, with less room for incompatible parts. Credo’s pitch is that the industry needs one agreed-upon way to build “wide, parallel optical connectivity” for both scale-up links (the tight mesh connecting accelerators inside a rack or pod) and scale-out links (the longer-haul fabric connecting racks and pods across a data center floor).
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Credo Actually Announced on October 12
The SRMO Workstream is not a shipped product. It is the formation of a collaborative technical effort under the OCP Networking Project, the same umbrella group that has previously hosted open standards work on switch hardware and rack-scale interconnects. Credo’s announcement frames SRMO as a specification project, with a stated goal of reaching OCP approval in 2027. The scope, as described in the release, covers link power, efficiency, signaling, parallelism, physical form factors, module and connector footprint, link budget, and compliance testing – essentially the full stack of engineering decisions that determine whether an optical component from one vendor can talk to a receiver from another.
That breadth matters because optical interconnects for AI clusters have historically been built in silos. A given accelerator vendor, a given switch vendor, and a given cable or transceiver maker would often converge on a working combination through bilateral engineering rather than a published open standard. That approach can work for a single generation of hardware, but it creates friction the moment a buyer wants to mix components from multiple suppliers, or when a supplier wants to sell into more than one customer’s AI cluster design without re-engineering every time. SRMO is Credo’s attempt to replace that ad hoc process with a documented, multi-vendor specification.
The technology at the center of SRMO is what Credo calls micro-emitter optics: arrays of micron-scale light sources operating in parallel to create a wide, high-channel-count optical link. Rather than a single high-speed optical channel, the approach spreads the data across many simultaneous low-power channels, an architecture that in principle scales channel count – and therefore aggregate bandwidth – without pushing each individual channel to extreme speeds that come with higher power and cost penalties. It’s a design philosophy more familiar from parallel electrical interconnects than from traditional long-haul optical transceivers, applied here to the short, dense links inside and between AI racks. Credo says its micro-emitter optics close the 3-to-50m gap between copper and traditional optics, cutting power up to 60% at 100M-hour MTBF reliability.
Why Scale-Up and Scale-Out Links Have Become the Bottleneck
The distinction between “scale-up” and “scale-out” interconnects has become one of the defining architectural questions in AI infrastructure over the past two years. Scale-up refers to the dense, extremely high-bandwidth links that connect accelerators within a single server or a tightly coupled pod – think of the NVLink-style fabrics that let multiple GPUs act as one giant compute unit. Scale-out refers to the fabric that connects those pods to each other and to the rest of the data center, typically Ethernet or InfiniBand-based networking running over longer optical or copper runs.
Both categories face the same underlying pressure: AI training and inference clusters keep getting bigger, and every additional accelerator added to a cluster increases the amount of east-west traffic that has to move between chips to keep them synchronized during distributed training runs. When interconnect bandwidth doesn’t keep pace with compute growth, expensive accelerators sit idle waiting for data, which is the exact failure mode cluster operators are trying to avoid after spending tens of billions of dollars on GPU and custom silicon procurement. Credo’s framing in its own announcement – that the industry is moving from thousands to tens of thousands of accelerators per cluster – is a direct acknowledgment of that scaling pressure, and it is the justification the company gives for why a patchwork of incompatible point solutions is no longer tenable.
This is also why the OCP Networking Project is the right home for this kind of effort rather than a single vendor’s proprietary roadmap. OCP was built specifically so hyperscale operators and their suppliers could agree on open hardware designs rather than each cutting one-off deals with component makers. A specification that emerges from OCP carries more weight with buyers precisely because it isn’t controlled by one company, even if, as in this case, one company is driving the initial workstream.
Who Is Already in the Room
Credo’s announcement names a specific set of organizations as among the first participants in the SRMO Workstream: AFL, ams OSRAM, BizLink, Foxlink, Heraeus Covantics, SENKO, and YOFC. That list spans several distinct links in the optical supply chain – fiber and cable manufacturing, optical component and emitter technology, connector systems, and cable assembly – which is consistent with the workstream’s goal of defining an end-to-end interoperable specification rather than a single component standard. ams OSRAM joined the Credo-led effort on October 13, 2026, according to ams OSRAM.
The release also references participation from what it describes as a leading hyperscaler, without disclosing that company’s name. Hyperscale cloud operators are typically the ones setting the real-world requirements that OCP specifications end up reflecting, since they are the buyers deploying AI infrastructure at the scale where interconnect standardization actually pays off. The identity of that hyperscaler is not confirmed in Credo’s announcement, and nothing in the public record ties the project to a specific named cloud provider at this stage.
SENKO Advanced Components, one of the named participants, is a connector and interconnect specialist, and the company’s Obie Roy, Vice President of Strategic Commercial Development, described the goal of establishing a common connector interface as channel counts scale into the thousands. That statement gets at a very concrete, unglamorous problem: as optical links move from dozens of channels to thousands, the physical connector – the part a technician actually plugs in – has to be standardized too, or every vendor ends up building incompatible mechanical interfaces around otherwise similar optical technology.
L.K. Bhupathi on the Case for a Standard
Credo’s own case for the workstream came from L.K. Bhupathi, the company’s Associate Vice President of Product. Bhupathi said, “As AI workloads scale from thousands to tens of thousands of accelerators, the industry needs a standardized approach to wide, parallel optical connectivity – not a patchwork of components that do not fit together or interoperate,” according to a report republished on Investing.com.
That line is worth sitting with because it describes a failure mode that’s already visible across the AI hardware industry: different accelerator makers, different switch vendors, and different optical suppliers each optimizing their own corner of the stack, with integration left to whoever is building the finished cluster. Bhupathi’s framing suggests Credo sees the current moment – before channel counts in the thousands become the norm – as the right window to lock in a shared approach, rather than trying to retrofit compatibility after several generations of incompatible designs have already shipped.
Separately, Credo CEO Bill Brennan has already been laying rhetorical groundwork for this kind of move. On the company’s September 1 earnings call, Brennan discussed extending the reach of Credo’s micro-emitter-based active optical cables, saying, “Key difference there is we will extend the length to 30 meters,” a reference to active optical cable products built on the same micro-emitter approach now being proposed for OCP standardization, according to commentary shared on social media from the call. Brennan went further on the same call, describing the broader trajectory of the technology: “I think that is just step 1 for ALCs and the micro emitter technology. A next natural step forward on that is to apply that to what comes with the scale up opportunity.” Those September remarks predate the formal SRMO announcement by six weeks, but they telegraph the same direction – Credo moving its micro-emitter optical technology from a component-level product into something it wants the rest of the industry to build around.
Inside the Open Compute Project’s Role
The Open Compute Project has spent more than a decade building a track record of turning hyperscaler engineering requirements into shared hardware designs, starting with server and rack specifications and gradually expanding into networking, storage, and now AI-specific interconnects. Placing SRMO under the OCP Networking Project specifically signals that Credo and its initial partners want this treated as networking infrastructure – not a niche optical component spec, but a piece of the fabric that AI data centers will depend on structurally, alongside switches, NICs, and cabling standards already governed by OCP.
This isn’t the only recent OCP-adjacent move from the optical and networking supply chain tied to AI buildouts. Marvell used October’s OCP Global Summit to detail its own Teralynx T100 switch silicon, a part the company has positioned around AI data center fabric needs, as covered in our report on Marvell’s OCP 2026 switch announcement. The timing isn’t a coincidence – OCP’s annual summit has become the venue where the AI infrastructure supply chain stakes out competing and complementary technical positions for the next several years of data center buildouts, and Credo’s SRMO launch fits that same calendar.
A 2027 target for OCP approval is itself a meaningful data point. Standards efforts inside OCP typically move through working groups, draft specifications, and interoperability testing before a final ratified spec emerges, and a roughly one-year runway from launch to approval is an aggressive but not unprecedented pace for a OCP workstream, especially one with support from established component suppliers rather than starting entirely from scratch. Credo has been active on other OCP interconnect proposals this cycle as well, including a separate scale-up memory interconnect contribution tracked by StockTitan’s CRDO coverage, underscoring how active the company has become across multiple OCP workstreams simultaneously.
Market Impact: What This Means for Credo and Its Supply Chain
For Credo, leading an OCP workstream around a technology it has already commercialized is a classic first-mover standardization play: if the industry converges on an approach modeled on the technology Credo has been shipping and refining, the company gains both a reputational edge and a head start on products built to the eventual spec. That dynamic has played out before in networking history – companies that get their approach adopted as the open standard tend to retain disproportionate share of the resulting market because customers trust the originator’s implementation and because the originator’s engineering choices get baked into the spec’s defaults. The October 12 release was syndicated across multiple wire and press-release distribution channels, including a listing on FinancialContent’s press release feed, a sign of how widely business and financial outlets picked up the announcement within hours.
For the named participants – AFL, ams OSRAM, BizLink, Foxlink, Heraeus Covantics, SENKO, and YOFC – the calculus is different but complementary. These are suppliers of fiber, connectors, cable assemblies, and optical components who benefit from a larger addressable market once a specification exists that multiple hyperscalers and accelerator vendors can design against, rather than each supplier building bespoke parts for individual customers. A shared connector and link-budget specification, the kind SENKO’s Obie Roy described, lowers the cost of qualifying new designs and should, in principle, shorten the time between a new AI accelerator generation shipping and the optical ecosystem around it being ready.
The unnamed hyperscaler’s involvement, even without confirmation of who it is, is the detail institutional investors and competing optical vendors will likely scrutinize most closely in the weeks following the announcement. Hyperscaler backing is often the strongest signal of whether a proposed standard will actually get designed into next-generation data centers, as opposed to remaining a paper specification that never achieves broad adoption.
The 1.6T ZeroFlap ALC as the First Associated Product
Credo announced the 1.6T ZeroFlap Active Light Cable (ALC) as the first associated product of the SRMO effort, and described it as a wide-parallel, row-scale optical interconnect for AI scale-out and front-end networks. Credo President and CEO Bill Brennan announced the 1.6T ZeroFlap Active Light Cable and the Short-Reach Micro-Emitter Optics Workstream during Day 1 of OCP Global Summit 2026 in San Jose, which ran October 12-15 at the San Jose Convention Center.
According to Credo’s own pages, the ZeroFlap ALC is designed to draw 12W per cable-end, roughly 30% less than comparable active electrical cables (AECs), is 10x lighter than AECs and extends the reach to 50m enabling row-scale connectivity. Credo’s 1.6T ZeroFlap active light cable (ALC) solution will be shipping in 2027, and no public official price for the 1.6T ZeroFlap ALC was found. Credo also showcased a ZeroFlap blind-mate optical backplane at OCP Global Summit 2026, and said it would demonstrate OmniConnect Weaver on the AMD Versal AI Edge Series Gen 2 adaptive SoC evaluation board at Booth B2.
For buyers and suppliers, the practical reading is that the cable previews the direction Credo wants the standard to take, but it is a vendor product and not proof that the specification is settled. Anyone planning designs around wide, parallel micro-emitter links should treat the cable’s published figures as Credo’s claims for now, and wait for the workstream’s draft parameters before committing to a connector or link-budget assumption.
Competitive Landscape: Where SRMO Fits Among AI Interconnect Approaches
SRMO enters a crowded field of competing and complementary approaches to the same underlying problem: getting enough bandwidth between AI accelerators without blowing through power and cost budgets. Traditional pluggable optical transceivers, co-packaged optics efforts from several silicon vendors, and linear-drive optics are all pursuing overlapping goals using different tradeoffs between integration density, power efficiency, and manufacturing complexity. The table below lays out how the publicly described characteristics of the SRMO approach compare with the general categories of competing short-reach interconnect technology discussed across the industry, based on Credo’s own announcement and widely understood characteristics of each approach category.
| Approach | Core Mechanism | Primary Target | Standardization Status |
|---|---|---|---|
| SRMO (Credo-led) | Arrays of micron-scale micro-emitters operating in parallel | Scale-up and scale-out links in AI data centers | New OCP workstream, targeting 2027 approval |
| Traditional pluggable transceivers | Single or few high-speed optical channels per module | General data center and scale-out networking | Mature, governed by established MSAs |
| Co-packaged optics | Optical engines integrated directly next to switch/accelerator silicon | Reducing power and reach loss at the chip edge | Multiple vendor-specific and consortium efforts, not yet unified |
| Linear-drive pluggable optics | Simplified signal path removing DSP stages | Lower-power short-reach links | Industry MSA work ongoing across several groups |
| Active electrical/copper links | High-speed copper cabling, no optics | Very short in-rack scale-up connections | Mature but reach-limited as rack power and density rise |
What distinguishes SRMO in this landscape, according to Credo’s description, is the explicit goal of covering both scale-up and scale-out use cases under one specification umbrella, with a stated emphasis on parallelism – many channels operating together – rather than pushing individual channel speeds higher. That is a different design philosophy from much of the transceiver industry’s recent focus on raising per-lane speeds, and it reflects a bet that power efficiency and manufacturability at scale matter more than raw per-channel throughput for the specific problem of connecting tens of thousands of accelerators.
Historical Context: How Optical Standards Have Shaped Data Centers Before
The pattern Credo is following – a vendor with a proprietary technology pushing it toward open standardization once the technology proves itself commercially – has precedent across networking history. Ethernet itself began as a single company’s lab project before becoming the IEEE 802.3 standard that underpins virtually all of modern networking. More recently, the pluggable optics industry organized around multi-source agreements that let buyers mix transceivers from different vendors inside the same switch, a shift that lowered costs and avoided single-vendor lock-in across an entire generation of data center buildouts.
The Open Compute Project itself was founded on exactly this principle: Facebook’s original 2011 OCP contribution of its own data center and server designs was meant to prevent every hyperscaler from re-engineering the same hardware problems in isolation. Since then, OCP has hosted standardization efforts spanning rack power delivery, server form factors, and – increasingly over the past several years – the networking and interconnect hardware specific to AI infrastructure, including switch designs like the one Marvell detailed at the same October 2026 OCP summit window covered in our Teralynx T100 report. SRMO is a continuation of that same institutional pattern, applied to a new layer of the interconnect stack.
The Bigger Picture: AI Infrastructure Spending Needs Interconnects to Keep Up
SRMO’s launch doesn’t exist in isolation from the broader capital spending wave reshaping AI data centers. Microsoft, Meta, and Google alone have pushed combined AI capital expenditure guidance into the hundreds of billions of dollars, a trend detailed in our coverage of the three companies’ $725 billion AI capex commitments. Spending at that scale is almost entirely about building and filling data centers with accelerators, and every dollar spent on compute is only as useful as the interconnect fabric that lets those accelerators work together efficiently.
That same capital intensity has already forced related standardization and policy moves elsewhere in AI infrastructure. FERC’s approval of a PJM registry process for data centers, covered in our report on the PJM data center registry decision, reflects grid operators trying to get ahead of AI data center power demand the same way OCP is trying to get ahead of interconnect fragmentation – both are attempts to impose order on a buildout that has been moving faster than the supporting standards and infrastructure around it. Cooling is facing a similar reckoning, as shown by investor enthusiasm behind thermal management suppliers detailed in our coverage of Trane Technologies’ AI cooling stock run, and even energy storage providers are being pulled into the AI buildout story, as covered in our report on battery storage demand tied to AI data centers. Compute, power, cooling, and now interconnects are all converging on the same realization: ad hoc, vendor-by-vendor engineering doesn’t scale to clusters with tens of thousands of accelerators, and someone has to build the shared rulebook.
Other recent silicon-level moves in AI infrastructure reinforce the same theme. AMD’s push into dense, agentic-AI-oriented rack designs, detailed in our report on AMD’s Helion rack architecture, is built around packing tens of thousands of cores into tightly coupled systems – exactly the kind of dense compute that needs the scale-up interconnect bandwidth SRMO is targeting. The interconnect layer and the compute layer are being redesigned on parallel, interdependent tracks across the industry right now.
Specification Scope: What SRMO Still Has to Define
Credo’s announcement lists eight specific technical areas the SRMO Workstream will need to resolve before a complete specification can go to OCP for approval: link power, efficiency, signaling, parallelism, physical form factors, module and connector footprint, link budget, and compliance. Each of those represents a genuinely difficult multi-vendor negotiation, not a formality.
| Specification Area | Why It’s Contested |
|---|---|
| Link power | Every watt spent on optics is a watt not available for compute inside a fixed rack power budget |
| Efficiency | Determines total cost of ownership across a cluster running continuously for years |
| Signaling | Must work reliably across parts sourced from multiple, independent suppliers |
| Parallelism | Channel count directly trades off against connector size and manufacturing complexity |
| Physical form factors | Has to fit existing and next-generation rack and server mechanical designs |
| Module and connector footprint | Drives whether parts from different vendors are physically interchangeable |
| Link budget | Defines the maximum reach and signal loss tolerance the spec will support |
| Compliance | Needs agreed test methodology before any vendor can claim conformance |
None of these eight areas has a publicly disclosed resolution yet. Credo’s release frames them as the workstream’s agenda rather than settled outcomes, which is consistent with a specification effort that just launched rather than one nearing completion. SRMO remains an active development effort and not an approved OCP standard: OCP approval in 2027 is a target, not a completed approval, and no approved SRMO specification or numbered version has been released. Expect the next twelve months to bring incremental announcements – draft parameters, additional OCP members joining, perhaps early interoperability demonstrations – well before the 2027 approval target Credo has stated.
What Happens Between Now and 2027
OCP workstreams typically follow a fairly predictable arc: initial member recruitment, draft technical proposals circulated among participants, interoperability and compliance test development, and eventually a vote to formally ratify the specification as an OCP standard. Credo’s stated 2027 target for approval implies the company expects this full cycle to run in roughly a year to 18 months from the October 2026 launch – an ambitious but achievable timeline if the named participants are already aligned on the core technical approach, which their early involvement suggests they are.
The more interesting open question is whether other major optical and interconnect vendors – companies not named among the initial SRMO participants – choose to join the workstream, stay on the sidelines building competing approaches, or try to steer the specification toward designs closer to their own existing products. OCP’s history shows all three outcomes are possible depending on how much a given technology has already been locked in by the time a workstream reaches critical mass.
Predictions: Where SRMO Standardization Goes From Here
Several outcomes look likely based on how similar OCP standardization efforts have historically unfolded and what Credo has already disclosed about SRMO’s structure and goals.
- Additional component and cable suppliers beyond the seven named founding participants will likely join the SRMO Workstream over the next two to three quarters, following the typical OCP pattern of early-adopter cores expanding once a workstream gains visibility.
- The unnamed hyperscaler’s identity, or additional hyperscaler participants, will probably surface publicly well before the 2027 approval target, since hyperscaler backing tends to become a selling point vendors want to disclose once contractual or competitive sensitivities ease.
- Credo’s existing micro-emitter-based active optical cable products, including the extended-reach variants Bill Brennan discussed on the September earnings call, will likely be positioned as early reference implementations of whatever the eventual SRMO specification formalizes.
- Competing interconnect approaches, including co-packaged optics and linear-drive optics efforts from other vendors, will continue to be pursued in parallel rather than abandoned, since AI data center operators are unlikely to standardize on a single interconnect philosophy across every layer of their networks this early in the buildout cycle.
- Expect at least one more OCP-adjacent interconnect or networking standardization announcement tied to AI scale-up and scale-out infrastructure within the next two quarters, given how active OCP’s Networking Project has become around AI-specific hardware this year.
The Bottom Line
Credo’s SRMO Workstream launch is a bet that the AI data center industry is at the exact inflection point where standardizing short-reach optical interconnects pays off more than continuing to let every vendor build proprietary solutions. The company isn’t claiming to have solved the problem yet – the eight-area technical agenda disclosed in the announcement makes clear there’s substantial engineering work ahead before a 2027 OCP approval becomes real. But the list of initial participants, spanning fiber, connectors, cable assembly, and optical component suppliers, plus an unnamed hyperscaler, suggests Credo has lined up enough of the supply chain to make a credible run at setting the terms for how AI clusters wire themselves together for the next several hardware generations.
Frequently Asked Questions
What is Credo’s SRMO Workstream?
SRMO stands for Short-Reach Micro-Emitter Optics. It is a new standardization effort launched by Credo Technology Group Holding Ltd on October 12, 2026, operating within the Open Compute Project’s Networking Project, aimed at creating an open specification for high-speed optical interconnects used in AI data center scale-up and scale-out applications.
When will the SRMO specification be finalized?
Credo has stated a target of 2027 for OCP approval of the specification. No exact month or finalized draft has been disclosed as of this announcement.
Who is participating in the SRMO Workstream?
Credo’s announcement names AFL, ams OSRAM, BizLink, Foxlink, Heraeus Covantics, SENKO, and YOFC as among the first participants, along with an unnamed hyperscaler whose identity has not been publicly confirmed.
What is micro-emitter optics technology?
It refers to arrays of micron-scale light sources operating in parallel to create wide, high-channel-count optical links, as opposed to relying on a small number of very high-speed optical channels. Credo has already used this approach in its active optical cable products.
What is the difference between scale-up and scale-out interconnects?
Scale-up interconnects link accelerators tightly within a server or pod to behave as one large compute unit. Scale-out interconnects connect those pods to each other and to the rest of the data center network, typically over longer distances.
Why does the Open Compute Project matter for this standard?
OCP is an industry-wide, multi-vendor body that hyperscalers and suppliers use to agree on open hardware designs rather than relying on proprietary, single-vendor specifications. A standard ratified by OCP typically carries more credibility with buyers because no single company controls it.
How does SRMO compare to co-packaged optics?
Co-packaged optics integrates optical engines directly next to switch or accelerator silicon to cut power and reach losses, while SRMO focuses on parallel micro-emitter arrays for short-reach links. The two approaches target overlapping problems using different architectural philosophies, and neither has been publicly described as replacing the other.
Does this announcement affect Credo’s existing products?
Credo’s existing micro-emitter-based active optical cables, including extended-reach variants discussed by CEO Bill Brennan on the company’s September 1 earnings call, are built on the same underlying technology now proposed for OCP standardization. Credo’s October 12 release itself does not detail specific product roadmaps tied to the SRMO specification, but Credo has since announced the 1.6T ZeroFlap Active Light Cable (ALC) as the first associated product, with shipping planned for 2027. The cable should not be described as an approved SRMO-compliant product, because the specification itself is still under development.
