Table of Contents
Co-packaged optics integrates optical engines directly onto the same package substrate as the switch ASIC, rather than placing transceivers in pluggable front-panel cages. The result is a much shorter electrical path between chip and fiber — less signal loss, and none of the power overhead that pluggable DSPs introduce.
That overhead is not trivial. QYResearch, in its 2025 report, found that DSPs in pluggable optics can drive overall system power up by as much as 25 to 30 percent. At hyperscale AI cluster scale, that figure translates to hundreds of megawatts across a single campus.
TechInsights projects that optical transceiver power consumption in AI datacenters will surpass 2.5 GW by 2030 under legacy pluggable architectures. CPO exists, in large part, to prevent that outcome.
Market sizing for co-packaged optics varies considerably depending on the research firm and what they count as "shipped." Semiconductorinsight.com values the co-packaged optics market at USD 1.45 billion in 2025 and projects it to reach USD 4.12 billion by 2034. HDIN Research puts the 2026 valuation range wider — at $2.2 billion to $4.2 billion — reflecting genuine uncertainty about how quickly hyperscaler deployments will ramp.
QYResearch, in its 2025 report, estimated the global market for co-packaged optics modules at US$ 44.6 million in 2024, with a forecast to US$ 1,236 million by 2031. The spread between these figures reflects how early the market still is, and how dramatically shipment volumes can shift on the back of a handful of large hyperscaler procurement decisions.
This is the question that matters most — and the one most market reports answer vaguely. Here is what is confirmed or credibly reported as of 2026.
Broadcom is the furthest along in production CPO hardware. TrendForce reports that Broadcom is continuing limited shipments of its 51.2T Bailly CPO switch. "Limited" is the operative word: Bailly is in customer validation at select hyperscalers, but it has not reached broad volume production. Broadcom controls its own silicon photonics platform, which gives it supply-chain advantages that fabless CPO startups simply don't have. The constraint right now is optical engine yield and the complexity of co-packaging at 51.2 terabits per second — not demand.
TrendForce also reports that NVIDIA has begun shipping its next-generation Spectrum-X CPO switch to select partners. NVIDIA's entry into CPO hardware is significant because it combines its own networking silicon with co-packaged optics in a platform already familiar to AI cluster operators. "Select partners" means early access rather than open availability, but hardware moving at all is a meaningful milestone.
Intel has been working on silicon photonics and CPO integration for several years through its Integrated Photonics Solutions group. Its approach centers on its own silicon photonics process, which it has used to supply optical engines to switch vendors. Intel is a component supplier in the CPO ecosystem rather than a finished-switch vendor — its shipment volumes are harder to track, but its role is foundational.
Marvell's CPO strategy is tightly linked to its Teralynx switch ASIC family. The company has demonstrated CPO integration on Teralynx and has partnerships with optical engine suppliers to bring co-packaged variants to market. As of 2026, Marvell is in advanced sampling and early customer evaluation rather than volume shipment.
Ranovus is smaller but technically notable. Its multi-wavelength laser-based optical engine is designed for co-packaging and has attracted attention from hyperscalers evaluating alternatives to silicon photonics. Ranovus is currently in the sampling and pilot deployment phase.
Ayar Labs takes a different approach: rather than targeting the switch, it integrates optical I/O directly into the compute package. Its TeraPHY chiplet is designed to replace electrical I/O between CPUs, GPUs, and switches entirely. Ayar has announced customer engagements with major compute vendors and has received funding from hyperscalers, but volume production is not confirmed for 2026.
| Company | Product | Status in 2026 |
|---|---|---|
| Broadcom | Bailly 51.2T CPO Switch | Limited shipments, hyperscaler validation |
| NVIDIA | Spectrum-X CPO Switch | Shipping to select partners |
| Intel | Silicon photonics optical engines | Component supply, ongoing |
| Marvell | Teralynx CPO | Advanced sampling, early evaluation |
| Ranovus | Multi-wavelength CPO engine | Pilot deployments |
| Ayar Labs | TeraPHY chiplet | Customer engagements, not volume production |
The honest read: two companies have hardware moving at any meaningful scale in 2026, and even those shipments are constrained to select hyperscaler programs. The rest of the ecosystem is in various stages of sampling and validation.
Several structural constraints are keeping CPO shipments well below what the market forecasts imply.
Optical engine yield. Co-packaging requires that both the optical engine and the switch ASIC pass quality tests before the combined package ships. A yield problem in either component kills the whole unit. At 51.2T and above, optical engine yields are still being refined.
Thermal management. Placing optical components directly on or adjacent to a hot switch ASIC creates thermal challenges that pluggable optics avoid entirely. Cooling solutions for CPO packages are not yet standardized, which complicates data center deployment planning.
Serviceability. Replacing a failed pluggable transceiver takes seconds. Replacing a failed co-packaged optical engine means returning the entire switch. That is a real operational concern in environments where uptime requirements are strict, and it slows adoption accordingly.
Packaging complexity. CPO assembly involves precision fiber coupling at scale — harder to automate than pluggable transceiver manufacturing. This is one reason semiconductorinsight.com notes that CPO solutions can cut bill-of-materials costs by up to 30 percent in theory, but realizing those savings requires manufacturing volume that most vendors have not yet reached.
If you are planning a 400G or 800G data center build today, CPO is not yet a mainstream procurement option. Broadcom's Bailly and NVIDIA's Spectrum-X CPO are real hardware, but they are entering hyperscaler programs first. Enterprise and mid-market operators will likely be working with pluggable 800G optics for the next 12 to 18 months before CPO reaches the supply availability that supports open procurement.
HDIN Research forecasts that annual shipments of 1.6T CPO modules will exceed 5 million units by the end of 2026 — a sharp acceleration from where the market sits today. Whether that materializes depends heavily on whether Broadcom and NVIDIA can resolve optical engine yield constraints in the second half of the year.
For most network engineers and infrastructure architects evaluating optics purchases right now, the practical path is high-speed pluggable modules in QSFP-DD and OSFP form factors at 400G and 800G, with CPO on the watch list for 2027 and beyond. TechInsights projects total optical transceiver market revenue to grow from $9.7 billion in 2025 to over $84.3 billion by 2030, and the bulk of that near-term growth will come from pluggable 800G and emerging 1.6T modules — not CPO.
If you are sourcing 400G or 800G pluggable optics today across QSFP-DD, OSFP, and other form factors, HYTOPTODEVICE covers the full range from 1.25G through 800G with OEM/ODM and white-label options for operators who need custom-coded modules.
CPO is not just a switch vendor competition. The companies controlling key upstream components will shape which switch vendors can actually scale.
Laser sources. Multi-wavelength lasers for CPO optical engines come from a small number of suppliers. Constraints here flow directly into switch vendor shipment timelines.
Silicon photonics foundries. Intel, GlobalFoundries, and TSMC all have silicon photonics process nodes, but capacity is limited and multiple CPO vendors are competing for the same wafer starts.
Fiber coupling automation. Attaching fiber arrays to co-packaged optical engines at high yield requires specialized equipment. The companies that crack reliable automated fiber coupling will become critical infrastructure for the entire CPO supply chain.
These dependencies explain why even well-funded CPO vendors are shipping in limited quantities. The bottleneck is rarely the switch ASIC itself.
Q1:What does CPO stand for and how is it different from pluggable optics?
A:CPO stands for co-packaged optics. It integrates the optical engine directly onto the switch ASIC package, eliminating the electrical path between the chip and a separate pluggable transceiver. Pluggable optics sit in front-panel cages and use a DSP to handle the electrical-to-optical conversion. CPO removes that DSP, reducing power consumption and signal loss at high data rates.
Q2:Which CPO companies are actually shipping hardware in 2026?
A:According to TrendForce, Broadcom is shipping limited quantities of its 51.2T Bailly CPO switch to hyperscaler validation programs, and NVIDIA has begun shipping its Spectrum-X CPO switch to select partners. Most other CPO vendors remain in sampling or early pilot phases rather than volume production.
Q3:Why is CPO not yet available for enterprise data center procurement?
A:CPO hardware is entering hyperscaler programs first because those customers can absorb the integration complexity and serviceability trade-offs. Optical engine yield issues, non-standardized thermal management, and complex deployment processes make CPO impractical for most enterprise environments in 2026.
Q4:How large is the co-packaged optics market in 2026?
A:Estimates vary significantly. Semiconductorinsight.com valued the market at USD 1.45 billion in 2025, while HDIN Research puts the 2026 range at $2.2 billion to $4.2 billion. The spread reflects how dependent total market size is on a small number of large hyperscaler procurement programs.
Q5:What are the main technical barriers slowing CPO adoption?
A:The primary barriers are optical engine yield at high switch capacities, thermal management of optical components co-located with hot ASICs, serviceability concerns around replacing failed optical engines, and the difficulty of automating fiber coupling in high-volume manufacturing.
Q6:Will CPO replace pluggable optics entirely?
A:Not in the near term. Pluggable optics in QSFP-DD and OSFP form factors at 400G and 800G will dominate enterprise and mid-market deployments through at least 2027. CPO is expected to grow significantly in hyperscale AI clusters where power density and bandwidth density requirements justify the integration complexity.
Q7:What should network engineers do now while waiting for CPO to mature?
A:Focus on 400G and 800G pluggable deployments using QSFP-DD and OSFP modules. Validate compatibility with your existing Cisco, Arista, or Juniper switching infrastructure, and evaluate third-party module suppliers that offer OEM/ODM options if you need custom-coded modules for your environment.
Co-packaged optics is real, and in 2026 it has moved from lab to limited production. Broadcom and NVIDIA are the companies with hardware actually shipping — both at constrained scale, inside hyperscaler programs. The rest of the CPO ecosystem is in sampling, pilot, or advanced development. For anyone planning a data center build today, pluggable 800G modules remain the practical choice, with CPO worth tracking closely as optical engine yields improve and manufacturing processes mature over the next 12 to 24 months.
Related Resources
1.Co-Packaged Optics (CPO): The Ultimate Answer to the AI Compute Boom
2.Top 20 Optical Module Leaders Paving the Way for CPO (Co-Packaged Optics)
3.AI Interconnect Revolution: Will LPO, NPO, and CPO Replace Traditional Pluggable Optical Transceivers?
