OCP Global Summit 2026 | Credo | Bill Brennan | Wider, Slower, Better: Optics That Can Beat Copper

1. Why Credo Is the One Saying This

For 20 years, data center connectivity has followed an unwritten rule: copper where you can, optics if you must. The reasoning was simple. Copper was more reliable, used less power, and cost less. It won on all three.

Who said it this time matters. Credo built its business on AECs (active electrical cables) and is the industry's best-known "copper company." Its ZeroFlap AECs were pulled out of it by customers: xAI, moving into a liquid-cooled facility and fighting link flaps, asked for 7-meter AECs so it could replace optics; Oracle, running air-cooled racks that needed links beyond 50 meters, had to go optical. Credo built a custom DSP for that and last year launched ZeroFlap optics, turning link stability into a real-time "yellow check engine light" so a link can be taken down in an orderly way before it actually flaps.

Optics is already Credo's second growth engine, as we broke down in Earnings Digest: Credo (CRDO) | Q1 FY2027. This OCP talk goes a step further: not "optics can work too," but "optics has to win on all three of copper's home fields."

The question is no longer whether to use optics, but whether optics can win on all three things copper does best.

2. Reliability: From a Component Problem to an Architecture Problem

Copper's reliability edge comes from being passive. Credo's baseline is 100 million hours MTBF, and that is exactly why NVIDIA ultimately pushed the industry to liquid cooling, so NVL72 could stay on copper.

The old problem with optics is active components in the data path. When one fails, the only fix is retries higher up the stack, which wreaks havoc in an AI cluster. Credo's answer is Layer Zero Redundancy: a µEmitter link has hundreds of lanes, so you simply add spare lanes. When the system sees a lane degrading, it moves the data to a spare in a hitless way, all below the OSI stack.

Put simply: copper earns reliability by not failing; µEmitters earn it by not stopping when something fails. Credo says this delivers the same 100 million hours of MTBF and MTBI (mean time between interruptions) as copper.

3. Power: Copper Is Passive, SerDes Aren't

The bundle of copper behind the rack looks like it draws no power, but there is a high-power SerDes at each end. Credo puts it at about 5 pJ/bit for 400G bidirectional, per Tx/Rx pair.

Apply that to a VR200 NVL72 copper backplane:

• Total scale-up bandwidth of 260 TB/s works out to about 10 kW, roughly 5% of rack power.

• At 2 pJ/bit for µEmitters (Credo says it is actually below 2), the saving is 60%: 6 kW per rack, 80 W per GPU.

• Reach goes from copper's 1.5 meters to 50 meters.

▲ VR200 NVL72 copper-backplane SerDes at ~5 pJ/bit draw ~10 kW; µEmitters at 2 pJ/bit cut that to ~4 kW, a 60% saving, ~80 W per GPU (Source: Simple Tech Trend, based on Credo's 2026 OCP Global Summit slides)

One thing to be clear about: the baseline is passive copper plus SerDes, not laser-based optical modules. Credo deliberately picked the industry's lowest-power option as its opponent, which is the boldest part of the whole talk.

4. Cost: A Watt Is Worth $100

This is the most useful slide of the session. Credo laid out three ways to value a watt:

• OpEx: lifetime electricity of about $5 per watt.

• CapEx: with liquid cooling now standard, power plus cooling infrastructure runs about $20 per watt.

• Opportunity cost: backed out of the compute contracts being signed today, $100 per watt or more.

The three contracts on the slide all land roughly in that range:

Contract	Value	Power	Per watt
Nscale	$45B	460 MW	~$98
Lambda	$35B	350 MW	~$100
SpaceX	$45B	300 MW	~$150

The logic: in a power-capped data center, every watt spent on networking is a watt not spent generating revenue-producing tokens. A $100 DAC whose SerDes burn 8 W at each end, 16 W in total, carries $1,600 of opportunity cost; cutting VR200 backplane power by 60% is worth $600,000.

▲ Credo backs out ~$100 of opportunity cost per watt from power contracts: a $100 DAC costs $1,600 and a VR200 backplane $600K (Source: Credo, 2026 OCP Global Summit)

There is a precondition: the rack has to actually be power-limited. If the bottleneck is somewhere else, the argument falls back to the old fight over purchase price. Credo's own claim is that purchase cost is "near copper," and there are no public numbers yet to verify that.

5. Product and Roadmap: Only One Item Is Firm for 2027

The 1.6T ZeroFlap ALC (Active Light Cable) is the first product, and Credo's first product built on micro-VCSELs:

• 12 W per end, 50 meters, 100 million hours MTBF/MTBI, integrated with Credo's PILOT telemetry and diagnostics platform

• About 200 fibers per cable; the slide says "10x longer, 10x lighter, and lower power," and Brennan added 30% less power on stage

• Volume production in 2027

Credo's slide splits µEmitters into three types, each matched to a reach:

Metric	Red µLED	Blue µLED	µVCSEL
Link power	< 1 pJ/bit	< 2 pJ/bit	< 2 pJ/bit
Fiber count	10,000s	1,000s	100s
Usable reach	≤ 2 m	≤ 10 m	≤ 50 m
Applications	Scale-in	Scale-in / Scale-up	Scale-in / Scale-up / Scale-out

Reading from the table, a 50-meter ALC with a few hundred fibers falls in the µVCSEL column, which Credo calls micro-pixels on stage.

The roadmap is framed as "faster emitters + tighter pitch = n² scaling":

• 2027: 1.6T ALC (scale-out), 6.4T NPO (scale-up, demoed on the show floor this week)

• 2028/29: 3.2T ALC and CPO at > 5 Tb/mm

• 2030+: Petabit-class scale-in

▲ Credo's µEmitter decade roadmap: 1.6T ALC / 6.4T NPO in 2027, 3.2T ALC and >5 Tb/mm CPO in 2028/29, petabit scale-in in 2030+ (Source: Credo, 2026 OCP Global Summit)

Wide-and-slow has one underrated advantage: the optics never have to clear the jump from 200G to 400G per lane; the speed-up challenge stays on the electrical side.

6. The Real Bottleneck Is Fiber, and Credo Answers with a Backplane and an Ecosystem

Slow-and-wide is not new. We covered it in It Wasn't Physics That Killed Slow-and-Wide, It Was 60,000 Fibers and ECOC 2026 | VCSELs Hit ≤1 pJ/bit Long Ago; Fiber Is the Real Bottleneck: NVIDIA paused its slow-and-wide project not because of the emitters, but because of fiber count and connectivity. A single 1.6T ALC needs about 200 fibers; scale that to a full rack and assembly and cleaning become the real challenge.

Credo addressed this head-on:

• ZeroFlap blind-mate optical backplane: 30 Pb/s and 1 million fibers, with 15% spare fibers built in for hitless failover; connectors use Senko blind-mate hardware with a Credo ferrule

• Versus today's copper backplane: 30x the density (1.8 Pb per tray), 30x the reach (50 m vs 1.5 m), and 90% less weight

• Claimed zero cleaning, zero maintenance; Credo says this is the first time anyone has truly solved the dust problem

▲ A 30 Pb/s, 1M-fiber µEmitter backplane with 15% over-provisioning: 30x density, 30x reach, 90% less weight (Source: Credo, 2026 OCP Global Summit)

Credo says outright that it has no plan to sell optical backplanes; the backplane exists to prove the whole roadmap can be built. The real move is the SRMO (Short Reach µEmitter Optics) workstream at OCP:

• Emitters: ams OSRAM, plus an open "2nd Emitter" slot whose supplier has not been named

• Fiber and connectors: Corning, Senko, AFL, Heraeus, YOFC

• Integrators: Credo, BizLink, Foxlink

• End users: one unannounced hyperscaler and two large GPU makers

▲ SRMO workstream members, from emitters and fiber connectors to assembly and end users (Source: Credo, 2026 OCP Global Summit)

It is the same playbook Credo used with AECs: rather than build the whole chain, it sets the spec and the ecosystem first and keeps the DSP and system integration for itself.

7. Risks and Counterarguments

• Choice of baseline: the 60% is measured against passive copper plus SerDes. Against NPO or CPO, the gap would shrink, and Credo did not give that number.

• Gearbox power is not broken out: the architecture shows a gearbox, management and a crossbar behind the host interface; it is unclear whether that power is inside the 2 pJ/bit.

• Still short of Microsoft's spec: Microsoft's optical I/O target at ECOC is >10 Tbps/mm (see ECOC 2026 | Microsoft Writes the Spec Sheet for Optical I/O); Credo's 2028/29 target is >5 Tb/mm, about half.

• Reliability is a design target, not field data: the 100 million hours MTBF/MTBI is derived from the redundancy architecture; real numbers come only after the ALC ships in 2027.

• "Near copper" purchase cost is unverified: a 200-fiber cable and blind-mate connectors have a completely different cost structure from a DAC.

8. What It Means for Taiwan's Supply Chain

• Integrators are the most direct beneficiaries. Of the three integrators in SRMO, besides Credo itself, BizLink and Foxlink are both Taiwanese cable assemblers. A 1.6T ALC with about 200 fibers is a new process and a new revenue line for cable assembly; watch when either company starts booking ALC revenue.

• No Taiwanese company in fiber and connectors. That segment is covered by Corning, Senko, AFL, Heraeus and YOFC; Taiwan still has a gap in high-fiber-count arrays and blind-mate connectors.

• The "2nd Emitter" slot is a window for Taiwanese LED and VCSEL makers, but first Credo has to name the supplier. Taiwan's microLED players are mostly still at the exhibit and qualification stage, and none has reported revenue from this product line.

9. Summary

The significance of this talk is not the µEmitter itself but that Credo rewrote the rules of comparison: reliability shifts from "components that don't fail" to "links that don't stop," and cost shifts from purchase price to opportunity cost per watt. As long as racks are power-limited, that math works heavily against copper.

But keep the timelines separate. The only product in volume in 2027 is the 1.6T ALC, a scale-out product; scale-up CPO waits until 2028/29, and the biggest challenge for slow-and-wide has never been the emitter but the hundreds to tens of thousands of fibers. Credo used the backplane and SRMO to show it sees the problem. Over the next 12 months, watch three things: field MTBI from the first ALC shipments, who fills the "2nd Emitter" slot, and when the SRMO draft spec goes public.

When the copper company itself retires copper's golden rule, the battle has moved from physics to fiber and specs.

This article is for technology and industry trend analysis only and does not constitute investment advice.