OCP Global Summit 2026 | Micas Networks | Tim Zhou | From LPO to NPO to CPO: The Open Optical Roadmap

1. Why Micas Is the One Saying This

Micas Networks sits in an unusual spot: it does its own hardware design, software and manufacturing, takes input upstream from connector and optical engine (OE) vendors, and downstream from hyperscalers and neoclouds. Tim Zhou opened by admitting the industry has no consensus yet on LPO, NPO and CPO, but Micas, sitting in the middle, has "figured out one thing or two."

More importantly, Micas describes itself as the first volume shipper of CPO switches. When a system vendor that has already shipped CPO says "CPO did not go hockey stick," it carries more weight than any analyst note.

The question was never whether optics move closer to silicon, but when, in what form, and who fixes it when it breaks.

2. Breaking Down Pluggables: 100 Million Units, and a Victim of Its Own Success

Zhou cited a forecast of 100 million pluggable transceivers shipping in 2026, a number the networking industry rarely hears. But at 800G and 1.6T, three problems surface at once:

• Power: on a fully populated Tomahawk 5 or Tomahawk 6, the optics already consume more power than the switch itself.

• Reliability: pluggables are sensitive to heat and contamination and are the leading contributor to network failures.

• Cost: in the same fully populated case, the optics also cost more than the switch.

So the industry went looking for something better, and the first thing it tried was CPO.

3. CPO's Report Card Is Strong, So Why Hasn't It Taken Off?

The CPO field data Micas shared is impressive:

• About 18 months ago it shipped hundreds of CPO switches to a hyperscaler, accumulating tens of millions of 400G-equivalent hours

• Link failures were 10x lower than pluggables

• At the switch level, power was 40% lower than pluggables

Yet it did not take off. Zhou splits the market in two: closed systems (he named NVIDIA) judge the whole solution by tokens per watt per dollar, so CPO will ramp faster there; open systems are held back by two things. First, vendor lock-in: the OE must sit in the same package as the ASIC, and sometimes even the fiber array unit (FAU) must be tightly matched. Second, mean time to repair (MTTR): MTBF is excellent, but once an OE fails, the only fix is to replace the entire switch. On top of that, packaging and box-assembly yields are still climbing.

▲ Micas CPO field data: link failures are 10x better and power is 40% lower, but when an OE fails the service boundary is the entire switch (Image source: Simple Tech Trend original, data from Micas Networks' 2026 OCP Global Summit presentation)

Put simply: CPO is not losing on physics, it is losing on the repair boundary. That matches our view in NPO Is Not a Stepping Stone to CPO.

4. Four Paths Are Not Replacements, They Are a Portfolio

Micas lines up pluggables, LPO, NPO and CPO by how close the optics sit to the ASIC, not by time:

Architecture	Primary value	Main trade-off	Best-fit window
Pluggable 800G/1.6T	Operational maturity, broad ecosystem	Higher electrical loss and power at high lane rates	Now and ongoing
LPO	Removes the DSP, lower power and latency	Tighter host-channel and link-margin requirements	Near term
NPO	Shorter electrical reach, more serviceable than CPO	More integration complexity than pluggables	2026-2028
CPO	Best density and power efficiency	System yield, thermal, repair, field service	Post-qualification

▲ Micas ranks four architectures by their distance from the ASIC, marking each one's main value, cost and best window (Image source: Micas Networks | 2026 OCP Global Summit presentation (06:37))

LPO is the path of least resistance: it looks like a pluggable and is managed like a pluggable, so network operations teams love it, but it only works where reach is short and ASIC signal and PCB quality are good enough. Micas' framing: it is not bandwidth alone, it is bandwidth per watt, per rack unit, per dollar and per unit of operational risk. Nobody will change their back-end operating model to fit a new optical technology.

5. Breaking Down NPO: The Goldilocks Distance, and Where to Put the Laser

Zhou calls NPO the Goldilocks option: close enough to the ASIC while doing two things. It opens the ecosystem, since many optical module makers are now building OEs as a natural evolution from pluggables; and an MSA makes it socketable, so you can replace the OE without replacing the whole box. He also conceded NPO is not field-replaceable for now, since it is hard to imagine opening the cover in a data hall.

Micas splits NPO into two paths: Open CPX-based (ASIC to CPX socket to serviceable OE, with risk in the socket, power, cooling, telemetry, firmware and OE qualification) and OBO/OCI (a traditional BGA ASIC with a very short PCB path to a modular OBO optical engine, with risk in connector loss, external laser distribution, CMIS management and production test). The slide carries a heavy caveat: NPO reduces CPO adoption risk only when the system vendor owns the full architecture.

Laser placement is the other fork. CPO sits too close to the ASIC and needs an external laser; NPO sits about an inch away, so a low-power source such as a VCSEL can be built in. An internal laser is not exposed to fiber contamination burn-out and removes one pluggable weak point; the cost is that laser heat, aging and failure impact all stay inside the OE.

▲ Pros and cons of external lasers versus NPO integrated lasers; hyperscaler demand signals point to integrated lasers for simpler servicing (Image source: Micas Networks | 2026 OCP Global Summit presentation (12:37))

The "hyperscaler requirement signal" on the slide is blunt: built-in laser NPO can simplify service procedures and reduce dependence on external light-source plumbing. That lines up with what we observed in CPO Lasers Don't Have to Chase 400 mW: NPO With Integrated Lasers.

6. Three New MSAs and the 2026-2029 Timeline

Since OFC, NPO has quasi-standards that everyone can build to. Micas lists three MSAs: OCI MSA (optical PHY: AMD, Broadcom, Meta, Microsoft, NVIDIA, OpenAI), Open CPX MSA (optical engine socket: Ciena, Coherent, Marvell, Molex, Samtec, TeraHop) and XPO MSA (pluggable form factor, led by Arista). For OCI background see Meta, Broadcom and AMD Define the OCI 200G Line-Side Spec; for Open CPX see Ciena Turns 6.4T Optical Engines Back Into Shelf Parts With Open CPX.

▲ Three new MSAs: OCI, Open CPX and XPO define the optical PHY, the optical engine socket and the pluggable form factor, respectively (Image source: Micas Networks | 2026 OCP Global Summit presentation (14:57))

Micas' three-year view:

Technology	2026-2029 outlook
Pluggable	Continues as the volume baseline through 2026-2029
LPO	2026 validation and selective deployment; broader in 2027 only where SI margin is proven
XPO	2027 engineering exploration; may coexist with NPO
OCI/OBO	2026-2028 early systems around TH6-class platforms
Open CPX	2027 prototype and ecosystem buildout; 2028 early deployment if interoperability improves
CPO	2026 validation; switch CPO grows first in 2027-2029, XPU-CPO follows later

▲ Micas 2026–2029 optical interconnect timeline: pluggables are the volume baseline, switch CPO ramps first from 2027, and Open CPX reaches early deployment only in 2028 (Image source: Simple Tech Trend original, data from Micas Networks' 2026 OCP Global Summit presentation)

Another signal: switches are moving into scale-up. Hyperscalers with their own XPUs cannot use NVLink and need an alternative, so NPO and CPO will be used to connect XPUs, mainly for inference rather than training.

7. Production Engineering Sets the Real Ramp

Micas lists five common bottlenecks:

• Electrical channel: margins at 200G/lane and future 400G/lane depend on connector transitions, return loss, crosstalk, skew, package escape and low-loss materials.

• Thermal and power: CPO and NPO put temperature-sensitive optics next to 1-2 kW class ASIC packages, so cold plates, TIMs and leak detection must be reliable; one OE equals 8 or 16 transceiver modules, so power rails multiply.

• Yield and serviceability: OE, fiber coupling, package interface and laser failures must not turn into an ASIC or full-switch replacement.

• Fiber and connectors: high fiber counts bring bend radius, routing, contamination control and automated assembly challenges.

• Software and operations: SONiC must handle OE bank mapping, CMIS, shared lasers and fault isolation.

The slide's closing line is the one to remember: the ramp will be won through production engineering, not headline architecture labels. Micas is working all three lines: CPO deployed, LPO deployed, NPO in design, unified under SONiC, and its four asks to OCP are open interface definitions, common validation methods, open software control and an interoperable supply chain.

8. Risks and Counterarguments

• The CPO data comes from one customer: the 10x and 40% figures come from a single hyperscaler deployment, and the talk did not say which pluggable generation was the baseline or whether 40% includes cooling.

• NPO serviceability is only half there: the socket lets an OE be swapped at the factory or a repair depot, but Micas itself admits it is not field-replaceable for now, so the MTTR advantage should be discounted.

• 2027 is a free-for-all: LPO, XPO, NPO and CPO are all in prototype at once, and a delay on any path will squeeze resources for the others.

• The closed camp may define the answer first: if NVIDIA's CPO ramps first, open-ecosystem MSAs may be forced to converge toward its interfaces.

9. What It Means for Taiwan's Supply Chain

• NPO gives optical module makers a way forward. Micas said outright that building OEs is a natural extension for module makers, who have the volume and the know-how; once Open CPX makes OEs socketable and multi-source, module assemblers get a window to re-enter.

• None of the three MSAs list a Taiwanese member. The companies shown for OCI, Open CPX and XPO include no Taiwanese firm, so spec-setting influence currently sits outside Taiwan.

• Built-in lasers are a new spec for light-source makers. If the hyperscaler signal holds, built-in VCSEL or CW sources in NPO become a separate component market, with the spec focus shifting from single-device high power to heat tolerance and reliability.

• Production engineering is Taiwan's strength. Among the five bottlenecks, connectors, cold plates, fiber assembly and production test are manufacturing problems; whoever gets yield up first captures switch CPO in 2027-2029 and NPO in 2028.

10. Conclusion

The value of this talk is that it comes from a system vendor that has already shipped CPO. Micas' conclusion is not "CPO doesn't work" but that in open systems CPO lacks a workable repair boundary and multi-sourcing, not performance. NPO fills those two gaps with sockets and MSAs, which is why it sits in the 2026-2028 window; yet Open CPX only reaches early deployment in 2028, while switch CPO starts growing in 2027.

Three things to watch over the next 12 months: Open CPX prototype interoperability results, the first TH6-class OCI/OBO systems, and whether hyperscalers formally require built-in laser NPO.

CPO has proven it can run; the next battle is over who can make it fail without replacing the whole switch.

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