Table of Contents
- What Is Co-Packaged Optics and Why Is It Reshaping AI Data Center Switches?
- Who Are the Key CPO Suppliers and What Are Their 2026 Roadmaps?
- What Are the Three Supply Chain Bottlenecks for CPO in 2026?
- How Do TSMC, Intel, and Samsung CPO Foundry Approaches Differ?
- What CPO Components Must Procurement Teams Source?
- How Does Vertical Integration Reduce Merchant Market CPO Supply?
- What Lead Times Should You Expect for CPO-Related Components?
- How Do You Build a CPO Sourcing Risk Plan for 800G to 1.6T Transition?
- What Are the Compliance and Standards Boundaries for CPO?
Co-packaged optics sourcing in 2026 is defined by scarcity: NVIDIA and Broadcom control the first production platforms, TSMC’s COUPE capacity is booked by AI accelerators, and the merchant market for optical engines barely exists.
What Is Co-Packaged Optics and Why Is It Reshaping AI Data Center Switches?
Pluggable optics served data centers for two decades; at 51.2T the copper between ASIC and module wastes power the industry can no longer afford.
Co-packaged optics (CPO) integrates the optical engine directly into the switch ASIC package, converting electrical signals to light at the silicon instead of at a faceplate module — cutting power per bit by up to 70 percent.
At 800G and 1.6T, pluggable modules burn 13 to 15 watts each on DSPs and long electrical channels; a fully populated 51.2T switch spends nearly half its power budget moving bits from ASIC to optics. CPO changes what you procure: there is no module to multi-source — the optics are fused to the platform, so vendor selection happens at system level, years before deployment. For the substrate side, see our CoWoS bottleneck analysis.
| Dimension | Pluggable Optics | Co-Packaged Optics |
|---|---|---|
| Power per link | 13-15 W per 800G module | Up to 70% lower (TrendForce) |
| Bandwidth density | Faceplate-limited | 10x higher (Tom’s Hardware) |
| Sourcing model | Multi-vendor modules | Platform-locked |
| Field service | Hot-swappable | Not replaceable |
| Supply chain risk | Module allocation | Foundry and OSAT queues |
Who Are the Key CPO Suppliers and What Are Their 2026 Roadmaps?
The CPO supplier list is short, and most build for their own ecosystems first and merchant customers second.
Five vendors lead CPO in 2026: NVIDIA with Spectrum-X Photonics, Broadcom with the 51.2T Bailly, Cisco, Intel with OCI chiplets, and Marvell via custom silicon.
NVIDIA’s Spectrum-X CPO uses TSMC’s COUPE packaging, targets 400 Tb/s aggregate switching, and ramps in the second half of 2026 with backend scale-out switches first; SemiAnalysis projects only 10,000 to 15,000 units in 2026. Broadcom’s Bailly is in volume production, and the Broadcom Bailly CPO platform anchors most third-party deployments. Cisco pursues CPO through Silicon One, Intel contributes 200G-per-lane OCI PICs, and Marvell embeds CPO in custom hyperscaler ASICs. Only Broadcom sells a merchant CPO switch today, and our Broadcom PCIe switch analysis shows how tight even its conventional allocation runs.
| Supplier | CPO Platform | Capacity | Production Status |
|---|---|---|---|
| NVIDIA | Spectrum-X Photonics | 102.4T, 400 Tb/s aggregate | Ramp H2 2026, 10-15K units |
| Broadcom | Bailly | 51.2T | Volume production, Meta validated |
| Cisco | Silicon One CPO | 51.2T class | Development |
| Intel | OCI chiplet | 200G/lane PIC | Sampling for 800G/1.6T |
| Marvell | Custom CPO ASIC | Program-dependent | Hyperscaler programs |
What Are the Three Supply Chain Bottlenecks for CPO in 2026?
Few CPO articles mention that its three critical inputs are capacity-constrained at the same time, for the same reason: AI.
TrendForce names three CPO bottlenecks: optical engines, silicon photonics (SiPh) wafers, and 2.5D/3D advanced packaging — the last shared with every AI accelerator in production.
Optical engines need precision assembly of lasers, modulators, and fiber attach at tolerances few OSAT lines handle at yield. SiPh wafers concentrate at TSMC, GlobalFoundries, and a few specialists. Third, CPO assembly uses the same CoWoS-class 2.5D and 3D lines as AI GPUs, so CPO orders queue behind accelerator volume. Our packaging alternatives analysis quantifies that fight, and the CoWoS to CoPoS roadmap shows TSMC’s relief plan through 2027-2029. CPO lead times follow AI accelerator cycles, not networking cycles.
How Do TSMC, Intel, and Samsung CPO Foundry Approaches Differ?
Foundry choice determines CPO architecture, and platform bets made in 2026 lock in through the decade.
TSMC pairs its COUPE photonics platform with CoWoS, Intel pursues OCI chiplets from its own fabs, and Samsung targets full CPO in 2028 with a turnkey flow in 2029.
Per Tom’s Hardware (August 3, 2026), TSMC’s COUPE plus CoWoS reaches 6.4 Tbps board-level optical interconnect with 2x energy efficiency and 10x bandwidth density on the platform behind NVIDIA’s Spectrum-X CPO — so TSMC capacity is spoken for. Intel’s OCI integrates lasers on-chip but serves Intel-aligned ecosystems. Samsung is later: true CPO in 2028, turnkey in 2029, the second source for buyers unwilling to depend on TSMC. Most buyers evaluate CPO platforms on switch specs and ignore the foundry. A platform on a sold-out foundry is a paper platform. Our AI power analysis covers the power side of the same build-outs.
| Foundry | Platform | Key Metrics | Timeline |
|---|---|---|---|
| TSMC | COUPE + CoWoS | 6.4 Tbps board-level, 2x efficiency, 10x density | In production for NVIDIA ramp |
| Intel | OCI chiplet | 200G/lane PIC, 800G/1.6T | Sampling |
| Samsung | Full CPO | Turnkey flow | 2028 CPO, 2029 turnkey |
| GlobalFoundries | SiPh wafers | 300mm photonics | Active foundry supply |
What CPO Components Must Procurement Teams Source?
CPO does not eliminate the BOM — it reshuffles it. Buyers focused only on the switch platform miss the component families around it.
Beyond the switch itself, procurement must cover external laser sources, photonic and electronic ICs, fiber connectivity, retimers, power delivery, and packaging materials like TSV interposers.
The photonic and electronic ICs are fused to the platform — you source them by choosing the vendor. External laser sources stay separate because keeping lasers off the hot ASIC improves reliability, and they are a recognized merchant opportunity. Around the platform sit retimers, precision fiber connectors, and a power tree migrating to new architectures per our 800V HVDC analysis. The CPO engine is single-sourced by design, but the surrounding BOM — power ICs, clocking, passives, connectors — is where OEM and EMS buyers keep multi-sourcing leverage and where line-down risk lives. A $0.40 clock buffer stops a $400,000 switch as dead as a missing optical engine. Run those lines through our BOM tool.
How Does Vertical Integration Reduce Merchant Market CPO Supply?
The companies driving CPO fastest — NVIDIA, Google, the hyperscalers — have no intention of selling you components.
Vertical integration pulls CPO supply captive: NVIDIA pairs TSMC COUPE capacity with its own switches, Google runs in-house optics, and the market grows at 37 percent CAGR mostly inside vendor walls.
IDTechEx projects the CPO market past $20 billion by 2036 at a 37 percent CAGR, with North America holding 44.5 percent of demand, per DataM Intelligence. Yet most of it moves inside vertically integrated stacks. NVIDIA’s 2026 output of 10,000 to 15,000 CPO switches, per SemiAnalysis, is essentially pre-allocated to its own ecosystem. Google designs its own optical interconnect rather than buying merchant parts. This mirrors AI accelerators, where merchant buyers queued behind captive demand for two years — our Q3 2026 lead time outlook tracks the same dynamic in MCUs. Buyers outside the hyperscaler club should engage platform vendors 12 to 18 months early and treat merchant CPO parts as allocation products.
What Lead Times Should You Expect for CPO-Related Components?
No distributor publishes CPO lead times yet: this is a negotiated-allocation market, not a quoted one.
Expect CPO platforms at 26 to 52 weeks under allocation, photonic components at 20 to 40 weeks where merchant supply exists, and packaging-adjacent parts tracking AI accelerator queues.
NVIDIA’s ramp of 10,000 to 15,000 CPO switches in 2026, per SemiAnalysis, against demand in the hundreds of thousands of ports, means platform allocation runs deep into 2027. Advanced packaging — the gating input — stays booked by AI accelerators, as our CoWoS capacity analysis documents, and SiPh wafers add 20-plus-week foundry cycles. CPO programs fail on schedule, not technology: an OEM that designs CPO into a 2027 product without locked allocation discovers the queue after design-in. For EMS partners, buffer the surrounding BOM — power, clocking, connectors — at 26-to-40-week exposure, consistent with our Q3 2026 MCU outlook. Once parts are secured, our semiconductor logistics guide covers photonics-sensitive handling.
How Do You Build a CPO Sourcing Risk Plan for 800G to 1.6T Transition?
The 800G-to-1.6T transition concentrates CPO sourcing risk: two optics generations, overlapping platforms, and suppliers rationing both.
A workable risk plan has four parts: dual-generation qualification, early platform allocation, BOM buffer stock, and standards-based exits.
First, qualify pluggable 800G alongside CPO: pluggables remain the fallback if CPO allocation slips, and hybrid designs keep that door open through 2027. Second, place 1.6T platform allocation requests now — with NVIDIA ramping only 10,000 to 15,000 units in 2026, mature supply may not arrive until 2028. Third, buffer the surrounding BOM at 26 to 40 weeks for power, clocking, and connectivity — where EMS builds actually stop. Fourth, align designs to OIF and IEEE 802.3 interfaces so a vendor exit does not force a board respin. Most buyers plan CPO as a technology adoption and forget it is a two-year supply commitment.
What Are the Compliance and Standards Boundaries for CPO?
CPO interoperability is not optional — multi-vendor optical links only work inside standards, and early platforms test those boundaries.
The key frameworks are OIF and IEEE 802.3 CPO interoperability standards, covering optical interfaces, external laser definitions, and link budgets that keep multi-vendor CPO deployable.
For procurement, standards compliance preserves your second source. OIF’s implementation agreements define how external lasers and optical engines from different vendors interoperate; IEEE 802.3 governs the links. A compliant platform lets you dual-source lasers and eventually engines; a proprietary one locks you to the vendor’s roadmap and pricing for the product’s life. First-generation CPO platforms ship with vendor-specific management interfaces even when the optics comply, so verify at the interface level, not the datasheet. For OEM buyers writing supply agreements, require OIF conformance evidence in the contract — it is the difference between a platform and a trap. Our industry solutions pages offer application-specific sourcing support.
Building AI data center switches and need CPO-ready component sourcing? SupplyICs supplies optical networking, switching, and advanced-packaging components from our vetted global network, with full traceability and QA inspection. Contact our sourcing team to discuss CPO component availability, lead times, and alternative sources for your 800G to 1.6T roadmap.
Frequently Asked Questions (FAQ)
What is the difference between CoWoS and CPO?
CoWoS is TSMC's substrate-level 2.5D packaging that stacks logic and HBM on a silicon interposer. CPO integrates optical engines directly with a switch or compute ASIC, moving data as light instead of copper. They solve different problems and are increasingly used together.
When will CPO switches be widely available?
Broadcom's 51.2T Bailly CPO switch is in volume production with Meta deployments validated. NVIDIA's Spectrum-X Photonics ramps in the second half of 2026, with SemiAnalysis projecting 10,000 to 15,000 units this year. Wider availability arrives through 2027 and 2028.
Who are the main CPO suppliers?
NVIDIA, Broadcom, Cisco, Intel, and Marvell lead in 2026. Broadcom and NVIDIA ship switch platforms, Intel supplies optical compute interconnect chiplets, and Marvell focuses on custom silicon. Foundry capacity comes mainly from TSMC's COUPE platform.
What are the CPO supply chain bottlenecks?
Three bottlenecks dominate, per TrendForce: optical engines, silicon photonics wafers, and advanced packaging. The 2.5D and 3D lines CPO needs are the same lines AI accelerators compete for, so CPO orders queue behind GPU demand through 2026 and 2027.
Can independent distributors source CPO components?
Only partially. Optical engines and photonic ICs are mostly vertically integrated into NVIDIA, Broadcom, and Google supply chains, so merchant supply is thin. Lock vendor allocation early and use qualified independents for surrounding parts like retimers, power ICs, and connectors.
