Zhongji Innolight Targets a 2027 NPO Ramp, but Volume Depends on Two Customers

Zhongji Innolight says two influential customers should begin purchasing near-packaged optics in volume during the second half of 2027. The forecast, reported through the rsshub 36kr news feed, puts a date on one of AI infrastructure’s most important hardware transitions. It also exposes the gap between customer interest and production-scale adoption.

The company expects broader NPO volume in 2028. Cloud service providers, AI model developers, and neocloud operators are evaluating the technology, according to its July 28 investor-relations disclosure. However, the company did not identify the two expected buyers, disclose committed order volumes, or provide completed qualification results.

That leaves the central contest unresolved. NPO offers a middle path between replaceable front-panel transceivers and co-packaged optics integrated beside a networking chip. Its success depends on whether that middle path solves urgent bandwidth and power problems without inheriting too many packaging, cooling, and serviceability risks.

Zhongji Innolight’s 2027 NPO Forecast Has Two Stages

The forecast describes a controlled customer ramp, not an industry-wide conversion scheduled for 2027.

Zhongji Innolight’s reported timeline separates product introduction from material volume. The company expects several interested customers to complete NPO product introduction during 2027. It then expects two influential customers to begin purchasing in batches during the second half.

The distinction matters because “product introduction” can cover several steps. A customer may receive samples, validate electrical and optical performance, integrate the device into a system, and test reliability before issuing production orders. Each stage can expose problems that delay the next one.

The company’s larger volume expectation sits in 2028. That makes 2027 a qualification and early-deployment year under the current forecast. It does not make 2027 the point when NPO automatically becomes the dominant architecture across AI clusters.

The original Chinese-language report says multiple CSPs, large-model companies, and neocloud customers have expressed interest. CSP means cloud service provider, while a neocloud is a newer cloud operator built primarily around accelerated AI computing. These groups share a need for dense, low-latency connections, but they do not necessarily use identical system designs.

A hyperscale cloud company may design custom accelerators, switches, racks, and network protocols as one platform. An AI laboratory may rent capacity from several infrastructure providers. A neocloud may prioritize rapid deployment using equipment that already fits established operating practices.

Those differences affect qualification. An optical engine that works inside one customer’s tightly controlled rack design may require new thermal, connector, firmware, or management work for another. Customer interest therefore indicates a market opening, not interchangeable demand.

The rsshub 36kr item closely tracks a July 28 report about Zhongji Innolight’s investor-relations disclosure. That report supports the 2027 introduction window, the two-customer expectation, and the larger 2028 ramp. It does not establish signed purchase commitments from those unnamed customers.

The limited disclosure creates both strategic value and uncertainty. Naming a deployment window helps suppliers prepare lasers, photonic chips, connectors, substrates, assembly lines, and testing capacity. Withholding customer identities protects commercial relationships but prevents outside verification of system readiness.

The most defensible reading is narrow. Zhongji Innolight believes NPO can enter a phased adoption cycle led by major customers. Whether that forecast becomes revenue depends on customer qualification, production yield, and the final architecture selected for each scale-up platform.

Scale-up networking connects accelerators so they can operate as a larger computing domain. It differs from scale-out networking, which links more servers or racks across a broader cluster. Scale-up connections face unusually strict latency, bandwidth, and reliability requirements because accelerators repeatedly exchange data during one computing job.

That workload explains why the company is discussing NPO now. Increasing accelerator performance does not help if data movement becomes the limiting factor. Yet the remedy must also fit power, cooling, maintenance, and manufacturing constraints inside dense AI systems.

Why Scale-Up Networks Are Pulling Optics Toward the Processor

NPO is gaining attention because faster electrical links become harder to carry across conventional circuit-board paths.

Traditional pluggable optical modules sit at a switch or server faceplate. Electrical signals travel from the application-specific integrated circuit, or ASIC, across a printed circuit board before reaching the module. The transceiver then converts those electrical signals into light.

This arrangement has practical advantages. Operators can replace a failed module without replacing the switch ASIC. Vendors can source standardized modules from several suppliers, and technicians understand the operating model.

The drawback grows with lane speed. Long electrical traces, connectors, and interfaces weaken high-speed signals. Designers compensate with additional signal processing, retimers, and cooling, which consume power and board area.

Nvidia says a conventional pluggable path can incur electrical loss reaching 22 decibels on a 200-gigabit-per-second channel. Its CPO architecture shortens that path and claims roughly 4 decibels of loss, according to the company’s photonics analysis. These figures describe Nvidia’s implementation and should not be treated as universal results for every system.

NPO moves the optical engine away from the faceplate and closer to the processor or switch package. The shorter electrical distance can reduce loss and power while preserving more modularity than fully co-packaged optics. The optical engine remains near the package rather than being integrated into the same package.

That positioning is NPO’s central proposition. It tries to capture much of the signal-integrity benefit available from tighter integration without making optics inseparable from the most expensive compute or switching silicon.

The architecture is especially relevant when hundreds of accelerators must communicate inside a scale-up domain. These systems exchange model parameters, activations, gradients, and synchronization messages. Slow or unreliable links can leave costly accelerators waiting for data.

Bandwidth density also becomes a physical constraint. A rack cannot indefinitely add faceplate modules, copper cables, connectors, and cooling hardware. Moving optical conversion inward can provide more aggregate bandwidth within a limited area.

Broadcom describes NPO as a bridge between conventional pluggable modules and CPO. Its proposed VCSEL-based design uses vertical-cavity surface-emitting lasers, compact sources commonly used for short optical links. The company argues that this approach can address short-reach scale-up connections while CPO continues to mature.

The pressure extends beyond transceiver manufacturers. Switch-chip vendors must expose suitable high-speed interfaces. Accelerator designers must decide where optical conversion belongs. Packaging companies must assemble electronic and photonic components with acceptable yield, while system builders must provide cooling and field-service procedures.

Cloud operators face a different calculation. They want more useful computing capacity per watt, but they also value replaceable components and multiple suppliers. A design that saves link power can still be unattractive if failures require replacing an entire board or taking a large computing domain offline.

This is why the 2027 NPO forecast is more than a product announcement. It suggests major customers are willing to test a new location for optical conversion. Their decisions will influence which suppliers control value inside future AI racks.

The transition is not simply copper versus fiber. Copper can remain practical over very short distances, especially when cost and serviceability dominate. Pluggable optics can also improve through lower-power designs and faster modules. NPO must beat evolving alternatives, not frozen versions of older technology.

The rsshub 36kr Signal Points to a Modular Optics Bet

Zhongji Innolight is effectively betting that customers want tighter optical integration without surrendering replaceability too quickly.

The rsshub 36kr source signal matters because it captures a specific commercial claim: NPO adoption should proceed in phases, beginning with leading customers and spreading later. That sequence fits how infrastructure transitions usually occur.

Large cloud operators can absorb the engineering cost of an early architecture. They can co-design boards, cooling systems, firmware, optical engines, and cables around a specific workload. They also operate enough equipment to make incremental power savings economically meaningful.

Smaller buyers generally need a more standardized product. They may wait for interoperable form factors, established repair procedures, qualified second sources, and evidence from early deployments. A design that succeeds in one custom platform does not immediately become a broad merchant product.

NPO appeals to both groups in theory. It gives early customers room to optimize placement and signal paths. It can later support modular optical engines that suppliers manufacture and test separately from the main compute package.

That separation can improve production economics. If an optical engine fails during manufacturing, the vendor may avoid discarding an expensive switch or accelerator package. If it fails in service, the system may allow replacement at the module or board level, depending on the final design.

CPO follows a more integrated route. It places optical engines directly beside the switch ASIC inside the package. The shorter connection can produce stronger power and bandwidth-density results, but packaging becomes more complicated.

Nvidia and Broadcom are already moving CPO closer to production. TrendForce reported on July 27 that Nvidia had started shipping Spectrum-X CPO switches to selected partners. Broadcom was also continuing limited shipments of its 51.2-terabit-per-second Bailly CPO switch.

TrendForce described optical-engine supply, silicon-photonics capacity, and advanced packaging as key constraints on expansion. Its CPO production report also said Nvidia’s system uses TSMC’s COUPE packaging and offers up to 400 terabits per second of switching capacity.

Those deployments put pressure on NPO suppliers. If CPO reaches dependable volume before NPO qualifications finish, customers may standardize around the more integrated architecture. Conversely, production or serviceability problems with CPO can create a wider opening for NPO.

Zhongji Innolight’s advantage comes from its experience producing high-speed optical transceivers. NPO retains more of the modular manufacturing model familiar to transceiver suppliers. However, moving the optical engine closer to hot silicon changes the thermal and packaging problem.

The contest is therefore not NPO against an imaginary future technology. It is NPO against CPO products entering early production, improved pluggables, active electrical cables, and customer-specific copper solutions.

Industry coordination may prevent one architecture from excluding the others. AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI helped establish an optical scale-up consortium to develop an open connectivity specification. The group’s optical specification is intended to support pluggable, on-board, and co-packaged implementations.

That flexibility is important. A common physical interface can let system designers choose different optical placements without rebuilding every protocol layer. It can also create a larger component market and reduce dependence on one proprietary implementation.

Standards do not eliminate execution risk. Vendors still need reliable lasers, photonic integrated circuits, electronic drivers, connectors, fiber assemblies, substrates, and testing processes. They must deliver those parts at yields that support commercial pricing, even though exact prices remain confidential.

NPO’s modularity can distribute those tasks across specialists. It can also create more interfaces where performance or reliability problems appear. Every connector and boundary improves replaceability but adds potential loss, assembly work, and qualification requirements.

That trade is the heart of Zhongji Innolight’s bet. The company does not need NPO to defeat every optical design. It needs important customers to decide that NPO offers the right integration level for specific scale-up systems.

NPO’s Middle Position Creates Its Biggest Execution Risks

The same modularity that makes NPO attractive can prevent it from matching either pluggables on serviceability or CPO on efficiency.

NPO places optical components near high-power compute or switching devices. That environment raises thermal concerns because laser efficiency, wavelength stability, electronic noise, and component lifetime can change with temperature.

Cooling the accelerator or switch remains the first priority. Optical engines must fit around heat sinks, cold plates, power-delivery hardware, memory, and high-speed electrical interfaces. A promising laboratory layout can become difficult when converted into a serviceable rack design.

Fiber management adds another constraint. Dense scale-up systems may require many short optical connections in confined spaces. Installers need connectors that tolerate handling, maintain alignment, and avoid contamination.

Maintenance procedures remain unsettled across implementations. A replaceable NPO engine sounds simpler than replacing a CPO switch package, but actual service may still require removing a board or draining a liquid-cooling loop. The repair boundary depends on the customer’s mechanical design.

Reliability is particularly important for scale-up computing. A failed connection can interrupt collective communication across a large accelerator domain. The operational cost includes idle processors, recovery time, and potentially restarting work from an earlier checkpoint.

Broadcom argues that its 3.2-terabit-per-second VCSEL NPO engine addresses bandwidth density, power, cost, reliability, and manufacturability. Those remain vendor claims until customers disclose qualification and production results. Different NPO designs may also use different laser and photonic technologies.

Supply concentration is another risk. Early products may depend on a small number of photonic-chip foundries, laser suppliers, packaging partners, or test systems. A qualification delay at one component supplier can hold back an entire platform.

TrendForce sees similar constraints in CPO. Its analysis identifies optical-engine yield, silicon-photonics supply, and advanced packaging as factors governing production expansion. NPO reduces some integration intensity, but it does not escape the optical supply chain.

Market forecasts also require careful interpretation. TrendForce has projected that the combined CPO and NPO market will exceed $39 billion by 2030, with stronger growth during 2028 and 2029. That combined number does not specify how much value NPO will capture, and it depends on scale-up networks adopting optical interconnects.

LightCounting offers a more restrained adoption frame. It expects only 10% to 15% of scale-up interconnects to migrate to NPO or CPO by 2030 or 2031. The firm still sees substantial opportunity because the underlying number of accelerator links is growing.

Its NPO market analysis also emphasizes supply chains, technical readiness, and industry collaboration. Those conditions align with the uncertainty inside Zhongji Innolight’s 2027 forecast.

Customer concentration presents a commercial issue. Two influential buyers can validate the product category and create significant demand. They can also hold strong negotiating leverage and change forecasts when their accelerator roadmaps shift.

The unnamed customers may choose different architectures for different generations. A cloud provider could use NPO in one custom accelerator system while adopting CPO switches elsewhere. “NPO customer” therefore does not imply exclusive or permanent commitment.

The timing of customer decisions matters as much as technical performance. Accelerator platforms follow coordinated development schedules. Missing a design window can postpone meaningful shipments until the next hardware generation, even if the optical product later passes testing.

Zhongji Innolight’s statement should consequently be treated as management’s current expectation. It has not been independently confirmed by the expected buyers. No public evidence attached to the report identifies binding volumes, final acceptance, or revenue recognized from those programs.

That verification gap does not make the forecast unimportant. It defines the evidence investors and infrastructure buyers need next. Samples, qualification milestones, named systems, production orders, and repeat deployments carry progressively more weight than expressions of interest.

CPO, Pluggables, and Copper Will Keep Competing

NPO can become a mainstream scale-up product without becoming the only architecture inside an AI data center.

AI networks contain links with different distances, bandwidth requirements, failure costs, and service patterns. One design rarely dominates every layer. Copper may remain useful inside a chassis, while optics handle longer rack-level and data-center connections.

Pluggable modules still offer the clearest replacement model. Technicians can remove a failed transceiver from the faceplate without disturbing the central ASIC. Suppliers also benefit from established manufacturing, standards, and testing systems.

Their problem is the electrical journey from the chip to the module. At higher lane rates, that path consumes more power and requires greater signal conditioning. Dense faceplates also limit the number of modules and fibers that a switch can accommodate.

NPO shortens the electrical path while retaining a separately manufactured optical engine. This makes it a candidate for systems that need more bandwidth density but are not ready to bind optics directly to the main package.

CPO pushes integration further. Nvidia says its latest photonics platforms can reduce power consumption by up to 3.5 times and improve resiliency by 10 times compared with its conventional architecture. Those claims reflect Nvidia’s system assumptions, not guaranteed industry averages.

Broadcom’s CPO roadmap targets both scale-out and scale-up networks. Its 51.2-terabit-per-second switch has entered volume manufacturing, according to TrendForce, while newer platforms continue to increase capacity.

These developments challenge any simple claim that NPO is the inevitable next step. Some customers may move directly from pluggables to CPO for high-end switches. Others may retain pluggables until standardization and service practices mature.

Copper remains competitive over the shortest reaches. Active electrical cables can include signal-conditioning electronics while preserving familiar connectors. Their suitability declines as distance, lane speed, and aggregate bandwidth increase, but they can offer attractive economics inside compact systems.

Micro LED optical links represent another emerging route. These designs use tiny light-emitting diodes for short-reach data transmission. They are earlier in commercialization and must prove manufacturing scale, but they show that NPO and CPO do not own every post-copper option.

The competitive picture also involves protocols. Nvidia’s NVLink serves its tightly integrated accelerator systems. AMD and other participants support alternatives such as UALink, while Ethernet continues expanding into AI scale-up and scale-out roles.

An open optical physical layer can support several protocols, but vendors still differentiate through switches, accelerators, software, packaging, and network management. Optical suppliers must qualify across those environments or accept dependence on a smaller customer set.

For Zhongji Innolight, the best outcome is not necessarily one universal NPO standard. A healthy market could contain several NPO form factors serving large customer platforms, provided the company can reuse enough components and manufacturing processes across them.

Excessive customization would weaken that outcome. Unique optical engines, connectors, thermal limits, and test requirements for every customer could raise development costs and slow production. It could also make second sourcing difficult.

Too much standardization can create another problem. If products become interchangeable before suppliers differentiate through yield or scale, customers gain pricing leverage. Optical vendors need common interfaces that expand the market without erasing their manufacturing advantages.

This balance explains why early leading customers matter. Their specifications can become reference designs for later buyers. A successful deployment can pull component partners and standards bodies toward one implementation.

A failed deployment can have the opposite effect. Reliability problems may send customers back to pluggables or forward to CPO. A platform delay can also leave suppliers holding capacity built for demand that arrives later than expected.

The 2027 period will therefore test more than Zhongji Innolight’s NPO product. It will test whether the surrounding ecosystem can convert technical interest into repeatable systems that operators are willing to maintain.

Three Signals Will Determine Whether 2028 Becomes the Volume Year

The next decisive evidence will come from customer qualification, manufacturing performance, and architecture choices in new accelerator platforms.

The first signal is whether either expected lead customer confirms production qualification. A named system, supplier acknowledgment, or customer deployment would strengthen Zhongji Innolight’s forecast. Continued anonymity through late 2027 would leave uncertainty around both timing and order quality.

Qualification should include more than peak bandwidth. Buyers will evaluate bit-error performance, latency, thermal behavior, optical power, connector reliability, firmware management, and failure recovery. They will also test whether manufacturing variation remains acceptable across a meaningful production batch.

The sequence matters. Engineering samples show that a design can function. Qualification establishes that it meets a customer’s requirements. Production orders indicate intended deployment, while repeat orders show that the system works well enough to expand.

The second signal is manufacturing yield across optical engines and advanced packaging. Yield measures the share of manufactured units that meet specifications. Low yield can restrict supply, complicate schedules, and consume excessive testing capacity.

NPO may hold an advantage because suppliers can test optical engines separately before system integration. That benefit must appear in real output, not only in architecture diagrams. Investors should look for capacity additions accompanied by stable yields and customer acceptance.

CPO production provides a useful comparison. TrendForce says Nvidia and Broadcom have moved CPO switches into early shipments or volume manufacturing, while component capacity remains a bottleneck. If those vendors improve output quickly, CPO will apply more pressure to NPO’s 2027 window.

If CPO ramps slowly because of package yield or service concerns, NPO’s modular design gains strategic value. Customers could use it as a bridge for several product generations rather than a temporary compromise.

The third signal is the optical architecture selected for upcoming scale-up accelerator systems. Product roadmaps from Nvidia, AMD, hyperscalers, and custom-ASIC developers will reveal whether optics moves onto the package, near it, or remains at the faceplate.

The Optical Compute Interconnect effort deserves attention because it includes major chip suppliers and AI infrastructure buyers. Concrete specifications, compliance tests, and interoperable components would reduce qualification risk. Repeated delays would favor proprietary designs and customer-specific engineering.

The reported Zhongji Innolight forecast will be strengthened if a lead platform adopts modular NPO engines, suppliers demonstrate repeatable yield, and a customer confirms volume deployment. It will be weakened if CPO captures the same system sockets first or if NPO remains stuck in sample qualification.

Readers should also separate market growth from supplier success. AI infrastructure can consume more optical connectivity while individual vendors lose programs, face price pressure, or misjudge the winning form factor. A growing category does not remove company-specific execution risk.

For enterprise buyers, the practical issue is system availability rather than the name of the optical architecture. They need accelerators that remain fed with data, operate within power budgets, and recover predictably from component failures.

Developers have a less direct but still important stake. Faster scale-up fabrics can reduce communication bottlenecks in large training and inference systems. However, hardware improvements only help when software frameworks and workload designs use the added bandwidth effectively.

Knowledge workers tracking this transition should preserve primary disclosures, customer statements, and architecture diagrams instead of relying on repeated news summaries. A searchable engineering knowledge base can help teams compare claims as specifications and deployment dates change.

The rsshub 36kr headline provides an early marker, not the final evidence. It identifies a 2027 introduction cycle, two expected lead buyers, and a possible 2028 volume ramp. The next task is to test each part of that sequence against disclosed customer and manufacturing milestones.

Watch for a named production platform first. Then examine whether qualified optical engines ship with repeatable yield. Finally, compare NPO’s share of new scale-up systems with CPO, improved pluggables, and short-reach copper.

If all three signals align, Zhongji Innolight’s forecast will look less like supplier optimism and more like the start of a durable architecture shift. If they diverge, 2027 may remain another extended qualification year.

The question is no longer whether AI systems need better interconnects. It is which integration boundary customers will trust at production scale. Follow the qualification evidence, not only the adoption language, as NPO approaches its proposed 2027 test.