Micas expands Malaysian production for commercial CPO

Micas expands Malaysian production for commercial CPO

Micas is expanding Malaysian manufacturing for commercial co-packaged optics systems. The capacity increase covers optical integration, assembly, testing, and clean-room production.


IN Brief:

  • Micas is expanding Malaysian manufacturing capacity around its commercially available co-packaged-optics networking platform.
  • Production capability includes clean-room processes, optical integration, system assembly, and validation.
  • The company's existing 51.2Tbps CPO switch combines Broadcom switching silicon with integrated optical engines for AI networking.

Micas Networks is expanding its manufacturing operation in Malaysia to increase capacity for commercial co-packaged-optics networking equipment, adding clean-room production, optical integration, system assembly, and validation around its 51.2Tbps Ethernet switching platform.

The development is a manufacturing step rather than another laboratory demonstration of CPO. Micas already offers a co-packaged-optics switch built around Broadcom switching silicon, and the latest investment is intended to increase the amount of photonic and system integration that can be carried out within its Malaysian production operation.

Co-packaged optics moves optical engines closer to the network switching ASIC instead of relying entirely on removable transceivers connected through long, high-speed electrical channels across the PCB. As lane rates increase, those electrical paths become progressively harder to route while maintaining acceptable signal integrity, power consumption, and thermal behaviour.

Micas’ existing CPO platform operates at 51.2Tbps and uses a 128-port 400GbE configuration. The company identifies Broadcom technology as central to the product and says the switch is already in volume production, moving the architecture beyond the prototype stage that still characterises much of the CPO market.

The electrical argument for co-packaging becomes stronger as switches move towards higher aggregate bandwidth. Conventional pluggable modules require high-speed SerDes signals to travel from the switch ASIC across the PCB to the front-panel optics. Equalisation and digital signal processing are then used to compensate for losses through copper traces, connectors, and module interfaces.

Moving optical conversion closer to the ASIC shortens those electrical paths. That can reduce signal-conditioning requirements and shift more of the high-bandwidth journey into optical fibre, where distance can be increased without the same electrical loss. The trade is that optical assembly becomes part of the switch manufacturing process rather than a replaceable module added after the main electronics have been built.

That difference explains the importance of clean-room and optical-integration capacity. Photonic components introduce requirements around contamination control, fibre handling, alignment, coupling, optical inspection, and calibration that are less prominent in a conventional Ethernet switch assembled largely from packaged electronic components and PCB-mounted pluggable cages.

The manufacturing process also has to validate the electrical and optical subsystems together. A fault can originate in the switch ASIC, package, optical engine, laser path, fibre interface, control electronics, firmware, or thermal environment, making production test more involved than qualifying a conventional PCB before independent optical modules are installed.

Micas’ Malaysian operation gives the company greater control over that integration. Its own corporate material describes a co-owned factory in Malaysia and presents manufacturing, supply-chain management, and custom engineering as part of its commercial model. The new expansion increases the relevance of that arrangement as CPO volumes rise.

Commercialisation still faces a serviceability problem. Pluggable optics are familiar to operators because a failed module can generally be replaced without removing the switch. When optics are integrated more deeply with the switching silicon, designers have to prevent an individual optical fault from turning into a disproportionately expensive system-level replacement.

Remote laser modules are one response, allowing the laser source to remain accessible while optical engines sit closer to the ASIC. Fibre routing, connectorisation, diagnostics, and field-replaceable elements will all influence whether CPO equipment can achieve the operational availability expected in large AI clusters.

Micas’ 51.2Tbps system has also been used in validated AI infrastructure with Supermicro and AMD hardware, providing a reference beyond isolated optical measurements. That moves evaluation towards the conditions operators actually care about: packet throughput, latency, cluster utilisation, thermal behaviour, software integration, and the ability to maintain the network over time.

The wider industry is already looking towards 800GbE and 1.6TbE connectivity, where copper reach and front-panel power become more difficult again. Micas describes CPO as part of that scaling path, while still selling conventional open-networking products alongside it rather than assuming every switch will move immediately to integrated optics.

Manufacturing capacity is therefore a meaningful milestone. An optical architecture cannot move into mainstream infrastructure if each system still depends on laboratory-style assembly, specialist manual alignment, or unusually intensive rework. Repeatable clean-room integration, automated test, process control, and acceptable manufacturing yield are prerequisites for commercial scale.

The Malaysian expansion does not resolve the remaining questions around interoperability, repair, lifecycle cost, and standards, but it shows where part of the CPO engineering burden is moving. The challenge is no longer confined to proving that optical engines can sit beside a switch ASIC; manufacturers increasingly have to demonstrate that those systems can be assembled, validated, delivered, and supported with the discipline expected of volume datacentre hardware.


Stories for you