InnoLight Bets on NPO, but the Optical Packaging Race Is Still Open

InnoLight says near-packaged optics could begin shipping in 2027, despite an unsettled contest over how AI systems should replace copper connections. The company expects more customers to adopt the architecture during 2028. Its prediction turns InnoLight NPO development from a laboratory discussion into a test of customer demand.

The claim came from Zhongji Innolight during an August 23 conference call, according to a company briefing reported by 36Kr. The optical-module supplier said some customers have provided clear requirements and orders. It is now developing customized products around those requests.

The orders have not been independently detailed, and the customers remain unnamed. Shipment timing, volumes, product configurations, and qualification milestones are also unclear. That leaves a meaningful gap between a reported order and a production deployment inside an AI system.

The larger contest is between NPO and co-packaged optics, commonly shortened to CPO. NPO places removable optical engines near a processor or switching chip. CPO integrates optics more tightly with that silicon, promising better electrical efficiency but creating harder manufacturing and maintenance problems.

InnoLight is betting that customers will value a middle position. NPO shortens high-speed electrical traces without making the optical engine inseparable from the most expensive silicon. That balance could matter as AI racks demand more bandwidth while operators still require replaceable components.

This is not simply another optical-module upgrade. If the company’s schedule holds, NPO will become an early bridge between familiar pluggable modules and deeply integrated optical systems. If qualification slips, CPO and improved pluggables will have more time to close that opening.

InnoLight NPO Orders Move the Debate Toward Deployment

The important change is not a new NPO concept, but InnoLight’s assertion that customer requirements have become orders.

InnoLight says many customers are actively discussing NPO designs. Some have reportedly supplied concrete product guidance and placed orders, although the company did not identify them. It expects initial shipments in 2027 and broader adoption during 2028.

Those dates matter because optical architectures cannot enter AI racks through component availability alone. Accelerator vendors, switch designers, cloud operators, and module suppliers must agree on electrical interfaces, thermal limits, management functions, fibers, and service procedures.

A product must then survive qualification across that entire system. Testing normally covers signal integrity, temperature behavior, optical performance, connector reliability, manufacturing variation, and failure recovery. A customer request can start this work, but it does not guarantee volume deployment.

InnoLight’s existing business gives the claim more weight than a forecast from an early-stage component developer. The company already offers a broad 1.6T portfolio alongside 800G, 400G, and lower-speed transceivers. That manufacturing base could help it transfer established module processes into a near-packaged format.

The company argues that NPO can preserve those mature processes. Optical engines remain separate units rather than becoming permanent parts of the main compute or switch package. Suppliers can reuse testing knowledge, assembly methods, and portions of an existing component chain.

Customers also gain a recognizable maintenance model. A failed optical engine can potentially be replaced without discarding the accelerator, switch ASIC, or complete board. That distinction becomes financially important when the attached silicon is scarce and difficult to service.

However, “order” can describe several commercial stages. It might mean engineering samples, qualification units, reserved capacity, a framework commitment, or scheduled production quantities. InnoLight has not publicly clarified which meaning applies here.

That ambiguity should shape how the 2027 forecast is read. The statement establishes a customer-backed development direction, not a confirmed industry transition. Investors and system buyers still need production evidence.

The report also arrived through an RSSHub 36Kr feed, which is useful for discovery but not proof of the underlying commercial details. The company’s exact disclosures and later regulatory filings remain the stronger verification points.

Even so, the signal is notable. NPO has moved from a general architecture discussed across the optical industry into InnoLight’s stated product schedule. It now has a proposed commercial window against which progress can be measured.

That window places pressure on every supplier tied to traditional front-panel modules. They must show that pluggable optics can satisfy rising rack bandwidth without excessive electrical loss. CPO developers face the opposite pressure, proving that deeper integration can arrive without unacceptable service and manufacturing costs.

InnoLight’s schedule therefore creates a practical benchmark. By 2027, customers should be able to show qualified hardware, not only presentation slides. By 2028, broader adoption should appear in supplier revenue, deployment announcements, or visible production capacity.

Scale-Up Networks Are Reaching Copper’s Hardest Limits

NPO matters because scale-up networks combine extreme bandwidth with short distances, leaving less room for inefficient electrical connections.

Scale-up networking connects accelerators within one tightly coordinated computing system. These links let GPUs or other processors exchange data with low latency and behave like a larger computing resource.

Scale-out networking connects more systems across a cluster. Both domains need bandwidth, but scale-up links place stricter demands on latency, synchronization, reliability, and communication between neighboring processors.

Copper remains attractive at very short distances because it is familiar and inexpensive. Yet higher signaling rates increase electrical loss, equalization demands, heat, and design complexity. Extending a fast electrical channel across a large rack becomes progressively harder.

Optics can carry high bandwidth across longer distances with less channel loss. The challenge is deciding where the electrical signal should become light. That physical placement shapes power, cooling, density, manufacturing, and repairability.

Traditional pluggable modules perform the conversion at a system’s front panel. Electrical signals must travel from the main silicon across the board before reaching the module. At higher speeds, that path can require retimers or stronger signal processing.

NPO moves the conversion closer to the processor or switch. The shorter electrical path reduces the channel burden, while fiber carries the signal across the remaining distance. The optical engine stays physically distinct from the main silicon package.

CPO goes further. It places optical engines beside the ASIC within a shared package or similarly tight assembly. That arrangement can minimize electrical reach, but it also binds the optical and electronic manufacturing processes more closely.

The industry’s interest is visible beyond any single vendor. OIF is developing interface work for linear modules, CPO, and NPO applications at up to 224 gigabits per second per electrical lane. The work targets lower power, lower cost, and lower latency.

A separate industry group has made scale-up optics its central mission. AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI formed the Optical Compute Interconnect Multi-Source Agreement group in March 2026.

Its open optical specification uses non-return-to-zero signaling and wavelength-division multiplexing. These techniques send defined electrical states across several optical wavelengths within the same connection.

The initial specification describes 200 gigabits per second in each direction. Its broader roadmap aims to increase bandwidth while supporting multiple optical implementations. Those implementations can include pluggable, near-packaged, and co-packaged designs.

This flexibility is crucial. The transition from copper does not automatically decide the packaging winner. A common optical interface can create demand for light-based links while leaving suppliers to compete over placement and integration.

The consortium also exposes the strategic pressure behind the transition. Cloud operators want more than higher link speed. They want multiple suppliers, manageable integration risk, and less dependence on a single proprietary component chain.

NPO fits that goal when modules remain replaceable and independently manufacturable. It can let system designers combine processors and optical engines from different vendors, provided their electrical, optical, and management interfaces align.

However, an open optical physical layer does not make every NPO product interchangeable. Mechanical dimensions, thermal solutions, connectors, control software, diagnostics, and board layouts also influence compatibility. System-level standardization remains unfinished.

That unfinished work explains why InnoLight emphasizes customized development. Early customers will probably require optical engines designed around particular accelerators, switches, boards, and cooling systems. Initial products may therefore be customer-specific despite the industry’s open ambitions.

The timing also reflects a broader change in AI infrastructure. Larger accelerator domains require more links, and each link carries more data. Rack power limits make wasted energy harder to tolerate, especially when cooling must remove every additional watt.

NPO offers a way to attack electrical loss without waiting for every CPO manufacturing question to be resolved. It does not eliminate the integration challenge. It relocates that challenge into a modular product closer to the main silicon.

This is why the InnoLight NPO forecast deserves attention from cloud buyers and hardware designers. It connects a recognized physical bottleneck with a deployment window. The remaining question is whether the compromise works at production scale.

NPO’s Real Opponent Is Co-Packaged Optics

The primary contest is not optics against copper, but modular NPO against more tightly integrated CPO.

Both architectures shorten the electrical connection between an ASIC and its optical engine. Both seek better bandwidth density and energy efficiency than long board traces feeding front-panel modules. Their disagreement concerns how much integration customers should accept.

NPO keeps the optical engine close enough to reduce electrical loss, yet separate enough to replace. This modularity can simplify optical testing because suppliers evaluate the engine before it joins a complete accelerator or switch system.

A removable unit can also contain failures. If one engine develops a laser, connector, or receiver problem, a technician may replace that unit. The repair does not necessarily disturb neighboring optics or the attached ASIC.

CPO pursues the shortest practical electrical path. Integrating optical engines beside the switching or computing silicon can reduce the energy spent driving electrical signals. It can also provide high bandwidth around the package edge.

Broadcom is developing both approaches rather than treating them as mutually exclusive. At OFC 2026, it presented a 3.2T NPO design based on vertical-cavity surface-emitting lasers. The company also promoted 102.4-terabit Ethernet switching with CPO.

That dual strategy is instructive. A major silicon supplier sees value in NPO’s packaging position while continuing to invest in deeper integration. Customers may use different architectures across products, distances, and deployment generations.

CPO still has a compelling efficiency argument. Every millimeter removed from a very high-speed electrical channel can reduce the need for equalization and retiming. Integration also creates tighter control over signal paths and bandwidth placement.

However, tighter integration joins the failure economics of two complex systems. A defect in an optical engine can affect an assembly containing expensive switching or compute silicon. Repair procedures become harder when components are not independently replaceable.

Manufacturing yield creates another concern. Combining advanced electronic silicon, photonic components, lasers, fibers, and packaging raises the number of processes that must succeed. A weak yield in one stage can constrain the finished assembly.

Thermal behavior is equally important. High-performance ASICs generate substantial heat, while lasers and optical components require controlled operating conditions. Bringing them closer forces designers to manage interacting temperature limits.

External laser sources can move some heat away from the package, but they add fibers, couplings, power distribution, and failure points. Integrated lasers reduce certain connections but can complicate temperature management and replacement.

NPO does not escape these problems. Its optical engines still sit near hot silicon, and dense fiber routing remains difficult. Socketed or removable components require connectors that maintain alignment and signal performance through installation cycles.

The board also carries a short electrical link rather than eliminating it. At future signaling rates, even that shortened path can become restrictive. CPO advocates can therefore argue that NPO delays deeper integration rather than replacing it.

That criticism identifies NPO’s strategic vulnerability. The architecture could become a transitional design whose useful life ends when CPO manufacturing, reliability, and service models mature. Suppliers must recover development and capacity investments before that transition.

InnoLight presents a different interpretation. The company expects NPO to become broadly accepted because it combines maintainability, cost control, stability, reliability, and mature module processes. Under that view, modularity is a lasting product advantage.

The evidence does not yet settle the argument. There are announced designs, specifications, engineering programs, and reported customer requests. Public data about fleet failure rates, repair times, qualification yields, and operating costs remains limited.

A fair comparison must also use the same system boundary. Measuring only the optical engine can hide retimers, electrical drivers, cooling, connectors, and control electronics elsewhere in the rack. Measuring the complete system can change the apparent winner.

Workload and topology matter as well. A tightly integrated switch may justify CPO because many high-speed ports surround one ASIC. A compute tray with different service requirements may favor replaceable NPO engines.

Customer operating models will influence the decision. Hyperscalers that replace complete trays can tolerate integration differently from enterprises that service individual modules. Hardware ownership periods and spare-part strategies will shape total cost.

Open specifications may improve NPO’s position by encouraging multiple module suppliers. Yet the same specifications can support CPO implementations. Standardization expands the optical market without guaranteeing which packaging approach captures it.

The most realistic outcome is not an immediate universal winner. Pluggable optics, NPO, and CPO can coexist across several hardware generations. The competitive issue is which architecture receives the largest new scale-up deployments.

InnoLight NPO products will need to prove that their service advantages survive real system constraints. A module that is technically removable offers limited value if technicians cannot access it safely inside a liquid-cooled rack.

CPO systems must prove the reverse proposition. They need to show that improved efficiency compensates for more concentrated manufacturing and repair risk. Demonstrations must become repeatable production results.

The Maintenance Advantage Still Needs Production Evidence

InnoLight’s strongest NPO argument is serviceability, but serviceability must be measured inside complete AI systems.

The company says NPO offers easier maintenance, lower cost, stability, and reliability. These are company claims associated with its customer discussions. They should not be treated as independently verified deployment results.

Easy maintenance begins with physical access. A technician must reach the optical engine without removing several cooling lines, fiber bundles, or adjacent accelerator boards. The connector must also tolerate replacement without contaminating or misaligning optical surfaces.

Detection matters before replacement. Operators need telemetry that distinguishes an optical-engine fault from problems in the ASIC, electrical interface, fiber, connector, firmware, or remote endpoint. Otherwise, modular hardware can still produce slow diagnosis.

Software support is another requirement. Management interfaces must expose temperature, optical power, link quality, error counts, and component identity. Fleet tools need consistent methods for collecting that data across suppliers.

NPO can inherit portions of the optical-module management model, but proximity to compute silicon changes the environment. Products face different airflow, cooling, board placement, and power constraints than front-panel transceivers.

Reliability comparisons must include connectors and sockets. A removable design introduces interfaces that a fixed assembly may avoid. Those interfaces add flexibility, but they also create mechanical tolerances and possible failure points.

CPO has its own service burden. A failed optical engine can affect a larger assembly, depending on the packaging design. Yet redundancy, external lasers, and system architecture can reduce the operational impact.

Public demonstrations should therefore report more than link uptime. Useful evidence includes failure isolation, replacement time, recovered capacity, thermal margins, manufacturing yield, and performance after repeated service operations.

Cost claims need similar discipline. NPO may reuse mature module processes, which can reduce development risk. It also requires new sockets, board layouts, fiber routing, thermal hardware, and qualification work.

CPO can carry higher packaging complexity, but it can remove certain retimers or electrical components. The complete cost depends on yields, volumes, repair policies, energy use, and system life.

The uncertainty is especially important because early NPO products will likely be customized. Custom engineering can accelerate one customer’s deployment, but it can delay broad interoperability. It can also concentrate revenue around a small number of buyers.

Unnamed orders create commercial uncertainty too. A development order from one influential customer can validate a direction without establishing broad demand. A volume commitment from several customers would support a stronger conclusion.

The 2027 schedule leaves limited time for design completion, sampling, qualification, capacity planning, and system integration. Any delay in lasers, photonic chips, connectors, packaging, or customer platforms can move revenue into a later period.

There is also a dependency on the accelerator roadmap. An NPO engine designed around one electrical interface or board may not transfer directly when the host silicon changes. Product schedules must remain synchronized across several companies.

Competition could compress the opportunity before volumes arrive. Broadcom can connect optics with its switch and custom-silicon portfolio. Other photonics suppliers can offer engines, lasers, packaging, or complete modules.

Traditional pluggable optics will not stand still either. Linear pluggable modules, improved digital signal processors, better retimers, and refined board design can extend the front-panel model. These products preserve familiar service procedures.

Meanwhile, CPO developers are collecting reliability evidence and improving package designs. Every successful production generation reduces the argument that integrated optics are too difficult to operate.

Industry forecasts should be read with these uncertainties in mind. TrendForce expects the combined CPO and NPO market to grow from about $100 million in 2025 to more than $39 billion by 2030.

Its market forecast places faster growth during 2028 and 2029 as scale-up systems adopt optical links. That direction supports InnoLight’s timing, but it does not verify the company’s orders.

The forecast also combines two competing architectures. A large combined market can coexist with disappointing NPO share if CPO captures most deployments. Investors should avoid treating category growth as supplier-specific revenue.

The strongest evidence will come from customer qualification and production disclosures. Named deployments would show where NPO sits, which silicon it supports, and why operators selected it.

Until then, the company’s 2027 target remains a credible development marker rather than a guaranteed shipment ramp. The distinction protects readers from confusing a commercial signal with completed adoption.

Why InnoLight NPO Could Become the Practical Middle Path

NPO’s opportunity exists because customers need better electrical efficiency before they are ready to accept CPO’s full integration burden.

A transitional technology can still create a substantial market. Ethernet switches, memory systems, and server components often pass through intermediate architectures that remain useful for multiple generations.

NPO addresses an immediate engineering problem with familiar operational ideas. It shortens the difficult electrical path, converts data to light near the ASIC, and preserves a replaceable optical unit.

That combination makes it easier to divide responsibility. Silicon vendors can define electrical interfaces and system requirements. Optical suppliers can manufacture and test engines. Equipment vendors can integrate those components into serviceable hardware.

The arrangement also supports supplier diversity. A customer may qualify several optical-engine vendors around a common interface. Multiple sources can reduce capacity risk and improve negotiating leverage.

However, this model works only when standards cover enough of the product. A shared optical waveform does not guarantee a shared socket, thermal design, firmware model, or diagnostic system.

Early customization may therefore be unavoidable. InnoLight says it is communicating with customers and developing products according to their requirements. That process can produce deployable hardware before universal standards mature.

The risk is fragmentation. Several customers could request incompatible form factors or management systems. Suppliers would then support many low-volume variants, weakening NPO’s promised cost advantage.

A successful lead customer can prevent that outcome by establishing a design others adopt. The industry has repeatedly used dominant platforms and multi-source agreements to turn early proprietary choices into common module formats.

The OCI group creates one possible coordination point. Its founding membership spans accelerator developers, silicon suppliers, hyperscalers, and an AI company operating large computing systems. That mix gives the specification practical influence.

Still, OCI does not select NPO as the exclusive implementation. Its architecture can support multiple packaging approaches. InnoLight must win product decisions within the broader optical transition.

The company’s mature transceiver experience may help. High-volume optics require control over photonics, electronics, assembly, testing, and supply. Those capabilities remain relevant even when modules move away from the front panel.

NPO also offers a possible capacity advantage. Manufacturing optical engines separately allows suppliers to test them before system assembly. Known-good units can reduce the chance that an optical defect reaches an expensive final product.

CPO suppliers can use known-good optical components as well, but final integration remains tighter. The relative yield advantage will depend on actual package architecture and test coverage.

For enterprise buyers, NPO can preserve a recognizable replacement boundary. That could support longer hardware life and targeted repairs. Yet most early scale-up deployments will likely occur inside hyperscale or specialized AI infrastructure.

Those operators evaluate different metrics. They may prioritize rack availability, energy use, deployment speed, and automated failure recovery above individual module replacement. NPO must demonstrate value against those fleet-level priorities.

Developers should care because interconnect architecture affects the size and behavior of accelerator systems. Higher scale-up bandwidth can reduce communication bottlenecks in distributed training and inference.

It does not automatically improve every application. Software must exploit the larger accelerator domain, while models and workloads must justify the infrastructure cost. Better links cannot compensate for inefficient parallelization.

Enterprise buyers should care because optical packaging decisions influence hardware availability and vendor choice. A more modular supply chain could expand sourcing options. A tightly integrated architecture could deliver efficiency while increasing platform dependence.

Knowledge workers and general AI users will encounter the effects indirectly. Infrastructure improvements can influence service capacity, latency, and operating cost. Those outcomes depend on complete systems, not one optical component.

The most defensible judgment is therefore conditional. InnoLight NPO can become a practical middle path if it reaches production before CPO becomes easy to manufacture and service.

Its adoption will also require stronger results than advanced pluggables can provide. If pluggable designs remain adequate, customers can delay the packaging transition and preserve existing operational methods.

If electrical limits become urgent while CPO integration remains difficult, NPO gains its clearest opening. The architecture then solves the immediate channel problem with less change to manufacturing and maintenance.

That is the mechanism behind InnoLight’s forecast. It is not claiming that NPO delivers the maximum possible integration. It is arguing that customers will choose the best deployable balance.

Three Signals Will Test the 2027 Shipment Forecast

The next evidence should come from qualification, standardized hardware, and measurable production demand, in that order.

The first signal is a named customer qualification or a detailed system demonstration. It should identify the host silicon, optical-engine configuration, bandwidth, thermal environment, and service method.

Such a disclosure would strengthen InnoLight’s claim by showing that customer requirements survived physical integration. Engineering samples alone would offer weaker support because they do not establish production readiness.

The second signal is progress on interoperable NPO interfaces. Useful developments include agreed electrical test points, management specifications, mechanical designs, and qualification methods.

Broader standardization would strengthen the argument that NPO can become widely accepted. Continued customer-specific fragmentation would weaken the expected cost, supply, and maintenance benefits.

The third signal is production evidence tied to 2027 demand. Investors should look for capacity allocation, qualification revenue, shipment volumes, or customer capital plans that distinguish NPO from ordinary transceiver growth.

A reported order without volume or delivery detail cannot establish a ramp. Repeated guidance, visible manufacturing preparation, and subsequent revenue would make the forecast more convincing.

CPO progress must be watched alongside all three signals. A major integrated deployment with strong reliability and repair data would narrow NPO’s window. Delays or service problems would make the modular middle path more attractive.

Improved pluggable optics are the other boundary. If they meet scale-up requirements at acceptable power and reach, customers can postpone NPO. That outcome would weaken the urgency behind InnoLight’s schedule.

Readers should treat 2027 as a verification deadline, not a guaranteed inflection point. The company has supplied a clear sequence: active discussions, customer requirements, initial shipments, then broader 2028 adoption.

That sequence can now be tested against public evidence. Watch for named platforms, completed qualifications, interoperable designs, and production quantities. Each one moves NPO from a supplier forecast toward an industry architecture.

The central question is practical: will customers choose replaceable optics near the chip before tightly integrated CPO becomes operationally routine? Track those three signals, then compare them with CPO deployment data. That approach will reveal whether InnoLight NPO is becoming a durable scale-up solution or only a temporary bridge.