Credo Trades Speed For Uptime

One hundred million hours is about eleven thousand four hundred years. That is the mean time between failures Credo is targeting for the 1.6T ZeroFlap Active Light Cable it unveiled on 12 October at the Open Compute Project Global Summit in San Jose, and the number is the whole argument. It is not a bandwidth figure, not an energy-per-bit figure, not a reach figure. It is a reliability figure, quoted first, in a product category that has spent two decades quoting line rate first. The title of the talk Credo's president and chief executive, Bill Brennan, gave that afternoon put it about as plainly as a conference session title can: "Wider. Slower. Better: Optics That Can Beat Copper."

The ALC is a pluggable optical cable with a dense array of micro-emitters and photodetectors at each end, feeding flexible fibre bundles rather than a pair of polished single-mode fibres. Credo's own product page is careful about what the emitters are: micro-VCSELs or microLEDs. That distinction matters more than it looks, because it means the architecture is defined by the array, not by the light source — whichever emitter the supply chain delivers at the right cost and yield can be dropped into the same arrangement.

The signalling is deliberately unambitious. Each 200G electrical lane is split into many lower-rate optical channels carried with NRZ modulation — on-off keying, two levels, the oldest trick in the book — and because the per-channel rate stays low, the design avoids what Credo calls the power and latency costs of complex DSP-based signal processing. A conventional 1.6T pluggable transceiver does the opposite: it takes PAM4 at 100 Gbaud, equalises it, forward-error-corrects it and spends both power and latency doing so. Credo's cable is designed to draw 12 W per cable end, roughly 30 per cent less than a comparable active electrical cable, and the company puts it at about ten times lighter than an AEC with, on its product page, around 75 per cent less cable bulk. Targeted reach is up to 50 metres.

Fifty metres is the number to hold onto. Credo's copper ZeroFlap AECs reach 0.5 to 7 metres at 56G and 112G per lane, which covers a rack and not much more. Fifty metres covers a row. The gap between those two distances is exactly where an AI fabric now has to put its densest links, and it is currently served by pluggable optics that cost more, draw more and — this is the claim under test — fail more often than the people buying them would like.

A link flap is a connection that goes down and comes back up. On a general-purpose network that is an annoyance; traffic reroutes, a retransmission happens, nobody notices. On the back-end fabric of a GPU cluster it is something closer to a stall, because collective operations are synchronous: a reduction across thousands of accelerators proceeds at the speed of its slowest participant, and a participant whose link has just dropped is not slow, it is absent. The cost is not the lost packets. It is the idle time on every other device in the collective while the fabric works out what happened.

That is the economics Credo is pricing against, and it explains a design that otherwise looks like a step backwards. Running many channels slowly instead of a few channels fast buys two things. It buys margin — a low-rate NRZ channel sits far enough from its error floor that a degrading laser or a contaminated facet does not immediately put it into FEC — and it buys granularity, because with hundreds of channels there is such a thing as a partial failure, and a partial failure can be engineered around.

So the cable carries spares. ZeroFlap link monitoring watches each channel, logs events, and when one weakens it moves that traffic onto a built-in spare. Credo calls the transition hitless and claims zero data loss across it. The link, from the switch's point of view, never goes down at all; it just quietly has one fewer spare than it did yesterday. This is redundancy of a kind that a four-lane PAM4 transceiver structurally cannot offer, because with four lanes there is nothing to fail over to.

On the same day, Credo announced a Short-Reach Micro-Emitter Optics workstream inside the OCP Networking Project, aimed at an approved specification in 2027. The scope is the dull, decisive stuff: link power and efficiency, signalling and parallelism, physical form factors including module and connector footprint, link budget and compliance. The initial participants are AFL, ams OSRAM, BizLink, Foxlink, Heraeus Covantics, SENKO and YOFC, alongside an unnamed hyperscaler.

Read the participant list and the shape of the problem appears. These are connector houses, fibre makers, emitter makers and packaging companies — not switch vendors. Wide-parallel optics does not fail at the physics; it fails at the mechanical interface, where a few hundred channels have to land on a few hundred detectors with both lateral and rotational alignment held, repeatedly, by technicians in a data hall. SENKO's contribution to the release is about precisely that: getting the physical connector interface exactly right, as Obie Roy, the company's vice president of strategic commercial development, put it, "especially as channel counts scale into the thousands."

A specification matters here for a reason specific to this architecture. A buyer can second-source a 400G-DR4 module from four vendors tomorrow, because the standard says what the module must do; there is no equivalent for micro-emitter optics, so every deployment today is a single-supplier commitment on a component sitting in the most critical links in the building. No hyperscaler signs up for that at volume.

No wavelength. No channel count. No per-channel rate. No insertion loss, no link budget, no measured bit error rate — the release says only that BER is better than traditional optics, which is a comparison without a number attached to either side. The 100-million-hour MTBF is described as a target, as are the 50-metre reach and the 12 W per end, and MTBF targets for optical assemblies are conventionally derived from component-level models rather than from fleets of cables run to failure. The product ships in 2027, so none of these gaps is unreasonable. They are simply the gaps.

The wide-and-slow argument is not new; it has been made on conference slides for several years, mostly by startups. What changed on 12 October is that a public company with an installed base in copper AECs put a 1.6T product behind it, attached a reliability specification to it rather than a bandwidth one, and moved the same day to standardise the connector and form factor through OCP with seven component suppliers in tow. That is a supply-chain announcement wearing a product's clothes.

For anyone making optics rather than buying them, the consequence is where the difficulty moves. A link built from hundreds of low-rate channels asks almost nothing of any individual emitter and almost everything of the array: uniformity across it, alignment into the bundle, and a connector that survives being unplugged. The hard components stop being the fast ones and start being the many ones — a different manufacturing problem, with different tolerances and a different set of suppliers, and OCP is about to write down what it has to meet.