Why CPO Won't Ship at Volume Before 2028
Explore the hurdles preventing co-packaged optics from reaching volume production before 2028. Understand yield issues, lengthy qualifications, and cautious customer behavior.
Every conference deck about co-packaged optics for the last three years has put the volume-production date about six months past the conference date. CPO has been "six months out" since 2023.
The reason isn't technology. The optics work. The packages work. The thermals can be engineered. The supply chain has the silicon. CPO hasn't shipped at volume because of yield, qualification, and customer behaviour — and none of those three gets solved by 2026.
This is the timeline I expect, with the disclaimer that timelines in this category slip by another year every twelve months.
What "production volume" actually means
The disagreement about CPO timelines isn't really about technology. It's about what counts as production.
The vendor-friendly definition: shipping units to a paying customer. By that measure, CPO has been in production since 2024 — a handful of hyperscalers run early-production deployments.
The operator-friendly definition: thousands of units per quarter, published prices, field-service expectations, replacement spares in inventory. By that measure, CPO is not in production anywhere as of mid-2026.
The gap between those two definitions is the whole timeline question. The "six months out" claim is true under the first and false under the second.
Yield is harder than the optics
A pluggable that fails comes out of a cage in five minutes. A CPO module that fails takes a line card — or sometimes a whole switch — out of service for the swap.
The yield economics aren't the same. A thousand pluggable modules with a 0.5% annual failure rate gives you five field replacements a year. Short, simple, spare from the pool. The same thousand optical channels concentrated into ten CPO line cards with a 0.5% line-card failure rate gives you five line-card swaps — longer, more expensive, more service-impacting, with different inventory.
CPO needs to ship at meaningfully lower module-level failure rates to deliver equivalent availability. Vendors are ramping toward those numbers. The qualification programs that prove them are multi-year exercises.
Qualification scales worse than expected
Pluggable optics qualify against host switches independently. Vendor A's optic plus vendor B's switch is one qualification event. Matrix scales linearly.
CPO optics don't exist outside the switch SKU they're integrated into. The matrix is per-switch-SKU, not per-optic. Every variant — reach class, fibre type, FEC mode — multiplies the SKU count.
Hyperscalers can absorb this overhead. They have engineering teams that pre-negotiate specific SKUs, qualify those, ignore the rest of the matrix. Enterprise buyers can't. They need catalog items. That's the structural reason CPO won't reach enterprise switching this decade.
Customer behaviour is the third pressure
A second-tier hyperscaler or regional carrier picking CPO in 2026 weighs three things: price premium, repair operations, vendor roadmap trust.
Price: CPO premium is currently 30–50% over equivalent pluggable. Closing slowly.
Repair: "send the line card back to the vendor". Acceptable for some operators, not for mid-tier teams running their own NOC.
Trust: the vendors who can ship CPO at all are the same ones who've missed previous launch dates. Operators planning 2027 deployment have to decide whether to bet on the next date being real. Most are betting on one more pluggable cycle.
The customers who would absorb the risk are leading hyperscalers, who have CPO already. The next tier won't sign until the first tier has demonstrated multi-year operational success. The first tier hasn't been at it long enough.
What 2028 actually requires
Three milestones for CPO to become operationally available outside hyperscaler AI fabric.
Yield at or beyond pluggable. Current CPO failure rates are between equivalent and 2-3x worse. Year-over-year improvement is real, gap hasn't closed. 2028 assumes the gap closes by 2027.
Standard SKUs through distributors. Today CPO is direct-vendor sale at project level. By 2028 it needs to be a catalog item with distributor pricing — meaning manufacturers commit to volume independent of specific customer orders.
Multi-vendor CPO line systems. Current offerings are tightly coupled to one vendor's switching silicon. Real product from at least three vendors is the threshold for a real market.
Two of three landing by mid-2028 = procurement option for tier-2 carriers. None landing = write the same post next year with 2029 in the title.
Where this leaves pluggable
The default for everything that isn't hyperscaler AI fabric, well into the late 2020s. The pluggable-vs-CPO question vendors push into 2026 RFPs is, for most customers, a question they don't yet need to answer.
If you're planning fabric for 2027–2028, pluggable optics are the safe assumption. CPO is a planning variable to track, not a baseline to commit to.
If you're hyperscaler-tier shipping AI training fabric, you've already committed and the rest of this analysis is observation.
Everyone in between has one more procurement cycle of pluggable before they need to decide. Vendors who tell you otherwise are selling the next CPO product launch. Pattern history says it slips.
What I'll watch for in 2027
Three signals that would change the timeline.
A second hyperscaler announcing CPO for general-purpose DC, not just AI fabric. A second vendor shipping catalog CPO (not custom-engineered per customer). Independent operator publishing real CPO failure rate data over 12 months.
Two of three landing by Q3 2027 = the 2028 timeline holds. Less than two = the timeline slips.
Plan for pluggable. Watch for the signals. Decide in 2028.