In 2026, global AI computing power clusters are expanding rapidly, and a machine-room transformation from “copper” to “light” is quietly advancing deep within data centers. In June 2026, NVIDIA’s demand guidance revision to 25 million units for NPO spread widely across capital markets, pushing the technical term “near-package optics” into the spotlight. More critically, tens-of-millions-of-units orders from cloud giants including Google and Amazon have already materialized — NPO has formally shifted from a conceptual narrative into a performance-delivery cycle. With three core advantages — low deployment cost, compatibility with existing machine rooms, and hot-swappable maintainability — NPO may take over from CPO in absorbing the bulk orders of global cloud vendors from 2026 through 2027. An industrial transformation centered on optical interconnect technology routes has already begun.
What Is NPO? Why Is It the Answer for 2026?
To understand why NPO has become the optical interconnect mainline in 2026, one must first clarify its technical positioning.
NPO, short for Near-Package Optics, refers to an architecture in which the optical engine is positioned on the same board as, but physically separate from, the GPU or switch chip. The electrical signal routing distance is controlled within a range of 5 to 15 centimeters, and the optical engine retains an independent hot-swappable module without requiring 2.5D co-packaging with the chip. Put simply, NPO moves the optical engine from its traditional position at the front-panel pluggable slot to a location closer to the switch chip, but the optical engine and the ASIC chip remain two separate packaged units.
In contrast, CPO (Co-Packaged Optics) has at its core the use of 2.5D or 3D advanced packaging technology to integrate the optical engine — responsible for optical-to-electrical conversion — together with the switch ASIC chip onto the same substrate or interposer, further shortening the electrical signal transmission path to the millimeter scale.
Each approach has its own advantages and disadvantages. NPO’s strengths lie in unified industry standards, compatibility with existing data centers, and low repair costs; domestic manufacturers can fully participate in design and manufacturing, and at equivalent speeds, the value captured at the single-unit level is twice that of CPO. Its shortcomings are that bandwidth density and extreme-condition power efficiency lag behind CPO. CPO’s strengths lie in extremely low power consumption and ultra-high port density, making it suitable for ten-thousand-card supercomputing clusters, but it still lacks a unified global standard, faces extremely difficult thermal management challenges, carries high costs for whole-system replacement, and has supply chain dominance concentrated in NVIDIA and TSMC, with domestic enterprises only able to supply components.
At the current industry stage, NPO precisely balances performance with deployability. It sits between traditional pluggable optical components and CPO. Its manufacturing process closely resembles existing optical module technology, does not depend on cutting-edge chip co-packaging capability, and allows for decoupled design between the switch chip and the optical engine — making it more conducive to forming a mature, multi-vendor collaborative ecosystem. It is precisely this “pragmatic” quality that has made NPO the optimal intermediate form before mainstream manufacturers pursue large-scale CPO adoption. The market’s overly high expectations for short-term CPO volume ramp-up have, in fact, only underscored NPO’s irreplaceability in 2026.
The Route Debate: CPO Delayed, NPO Absorbs All Short-Term Incremental Demand
Optical interconnect technology routes have undergone significant expectation adjustments over the past year.
The market previously held high expectations for short-term CPO volume ramp-up, but real-world progress has not been optimistic. Affected by yield rate and thermal management bottlenecks, data from relevant institutions shows that CPO’s penetration rate in AI data center optical communication modules will be only approximately 0.5% in 2026, with this figure expected to rise to 35% only by 2030. A June report indicated that large-scale CPO mass production has been postponed to 2028-2029.
The main obstacles to CPO deployment come from multiple dimensions: the absence of a unified global standard, extremely difficult thermal management, and high costs for whole-system replacement. Besi, the global leader in high-end advanced packaging, views CPO as the most exciting development direction in the hybrid bonding field, yet its revenue model indicates that CPO was still in the R&D validation stage in 2025, only entering a conservative deployment stage in 2026. NVIDIA’s networking vice president also stated clearly at GTC that CPO is expected to become widespread alongside the mass production of Vera Rubin in the second half of 2026.
Against this backdrop, NPO has become the only optical interconnect solution capable of large-scale deployment from 2026 to 2027. The market’s previous expectations for short-term CPO ramp-up have been comprehensively revised downward, with funding and orders tilting entirely toward NPO. It is worth noting that different views still exist in the market regarding the industrial timeline of CPO versus NPO. Experts from leading overseas optical communication companies have stated that NPO remains in the customer solution evaluation and testing stage, with commercial deployment not expected until the second half of 2027 at the earliest. However, judging from publicly disclosed industry signals — including Google’s tens-of-millions-of-units order and NVIDIA’s demand guidance — the real-world delivery of NPO orders in 2026 is already an established fact. An industry consensus is forming: in the short term (2026-2027), NPO will see single-track volume ramp-up; in the medium term (2027-2028), NPO and CPO will run on dual tracks in parallel; in the long term (after 2028), CPO will penetrate high-end supercomputing while NPO moves down into general-purpose computing clusters, with the two not being fully substitutive of each other.
Orders and Scale: Tens-of-Millions-of-Units Confirmed Orders Open Up the Performance Ceiling
If 2025’s optical interconnect sector hype was primarily about samples and trial shipments, then the core change in 2026 is this: real orders have already been placed.
In April 2026, Google placed a bulk order for 12 million units of 3.2T NPO, with a total order value of RMB 12 billion to 15 billion (approximately USD 1.67 billion to USD 2.08 billion), with delivery scheduled from Q3 2026 to Q2 2027, dedicated to TPU v7 or v8 supercomputing clusters. On NVIDIA’s side, the number of NPO optical engines per GPU in the Rubin Ultra rack has increased by 78%; full-year 2026 NPO demand is approximately 10,000 units, expected to rise to a scale of 500,000 to 800,000 units by 2028.
Even more notable is the continued upward revision of demand guidance. According to recent North American research, NVIDIA has already revised its NPO demand guidance upward to 25 million units. Amazon has given clear guidance of approximately 10 million units. Demand from Google, Meta, and Microsoft has also become clear — Meta and Microsoft’s NPO demand is concentrated primarily on the Scale Up side, with clear demand guidance for the following two years expected in the second half of this year. At this point, all four major cloud vendors have essentially confirmed strong NPO demand. The latest June calculations from industry institutions show that the global NPO optical engine market was approximately RMB 14.2 billion (approximately USD 1.97 billion) in 2025 and is expected to surpass RMB 87 billion (approximately USD 12.08 billion) by 2027, representing a two-year compound annual growth rate of 19.3%, with all of the incremental growth coming from global leading cloud vendors’ full optical transformation of their machine rooms. GF Securities, after dual verification through research and supply chain checks, indicated that real effective demand from NPO cloud service providers in 2027 will be approximately 10 million units, with a market size of approximately USD 10 billion to USD 15 billion.
Taking Innolight as an example, market rumors suggest the company has secured an NVIDIA allocation guidance of approximately 10 million NPO units, corresponding to an order value of approximately USD 16 billion. Innolight’s full-year 2025 revenue was RMB 38.24 billion (approximately USD 5.31 billion) — the NPO segment alone is equivalent to three times its existing scale. Even applying a 50% discount, that would still amount to RMB 57 billion (approximately USD 7.92 billion), equivalent to recreating one and a half additional Innolights. This is precisely the core reason institutions have raised their target price for Innolight.
From “concept” to “orders,” NPO is undergoing the critical leap from laboratory to mass-production line.
Industrial Chain Restructuring: Comprehensive Benefits From Optical Modules to Upstream Components
The breakout of NPO does not benefit a single segment in isolation; it is driving a restructuring of the entire optical communication supply chain.
At the optical module level, Innolight and Eoptolink have been confirmed by the industry as the primary NPO suppliers. Overseas CSP customer demand for optical modules has been revised upward compared with the March OFC period, with the upward revision concentrated in the 800G category. A single 1.6T optical module is valued at approximately USD 700, while a single 800G optical module is valued at approximately USD 350. If Broadcom resolves TSMC’s capacity constraints in the second half of the year, there remains further upside potential for 1.6T demand.
At the upstream optical chip level, the rapid increase in silicon photonics penetration is another important mainline. An April brokerage research report noted that traditional EML optical chip production capacity is concentrated overseas and constrained by indium phosphide materials, with a significant supply gap expected in the EML route in 2026 — a gap that will primarily be filled by silicon photonics solutions. Silicon photonics is expected to account for more than 50% of 800G optical modules in 2026, and as high as 70% to 80% of 1.6T optical modules.
A core transformation within the NPO architecture is also that, once the optical engine and switch chip are decoupled, breaking through the bandwidth bottleneck no longer relies solely on increasing single-channel rates, but instead expands the number of parallel channels substantially through high-density silicon photonics integration — moving from a traditional 8 channels to 32, 64, or even 128 channels. This is directly driving upstream MPO fiber connectors into a qualitative transformation period of simultaneous volume and price increases. The number of optical connection points within equipment is growing exponentially; the price of 32-fiber MPO connectors is already 2.5 times that of 16-fiber connectors, while the value of 64-fiber connectors is more than 10 times higher still. However, the supply side faces severe bottlenecks — globally, only two or three companies, including US Conec and Senko, can supply high-fiber-count MT ferrules at scale, A2 thin-diameter optical fiber remains in persistent short supply, and a large number of second- and third-tier manufacturers have been forced to exit the market due to their inability to secure materials. The industry’s CR5 concentration has rapidly risen from 30% in 2025 to more than 50% in 2026.
The breakout of NPO is driving a comprehensive upgrade of the entire optical communication supply chain, from modules to chips, with industrial value accelerating its concentration toward high-barrier segments.
Outlook: 2026 Is Only the Starting Point
Looking back from the vantage point of June 2026, NPO’s journey from a technical concept to tens-of-millions-of-units order delivery has taken only one to two short years. But viewed against a longer industrial cycle, 2026 may be only the starting point.
NVIDIA’s technology roadmap shows that CPO will become widespread alongside the mass production of Vera Rubin in the second half of 2026. The overall CPO shipment trajectory is clear: approximately 10,000 complete units shipped in 2026, approximately 100,000 in 2027, and approximately 500,000 to 800,000 in 2028. This means that NPO and CPO will enter a dual-track parallel landscape in 2027-2028.
From a more fundamental perspective, AI computing power development is facing the challenge of a “connectivity wall” — its geometric growth curve far exceeds the arithmetic growth of the memory wall and the improvement of computing power itself. This determines that demand for optical communication and storage will continue to grow at a high rate, and the supply-demand gap will continue to widen.
The breakout of NPO is by no means accidental. It is the inevitable product of AI computing clusters transitioning from copper cables to optical interconnects, a technical choice forced by physical bottlenecks, and, more fundamentally, a collective decision made by global cloud vendors in the computing power arms race. CPO is built around a closed ecosystem led by NVIDIA, with the core optical engine segment monopolized overseas; NPO, by contrast, has seen cloud vendors including Google, Meta, and Alibaba form the CPX Alliance to unify global standards, adopting an open supply chain in which domestic optical module, passive component, and optical chip companies can fully participate in whole-system design, assembly, and delivery, capturing profit from the high-value segments of the supply chain. In 2026, NPO is moving from industry narrative to performance delivery. For the entire optical communication supply chain, this may be the beginning of a structural opportunity that will continue for years to come.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.