CPO Volume Shipments Delayed to 2028 as Hybrid Bonding Yield Becomes Core Bottleneck
nashnova research
Volume production of co-packaged optics (CPO) has slipped from 2027 to 2028. The crux: cumulative hybrid-bonding yield sits at roughly 40%, barely half the 90% needed for mass production — meaning CPO's ramp timeline now hinges on a vacuum tool that does not yet exist in mature form.
Why has CPO slipped again?
FUNDA's latest survey shows scale-out CPO volume delivery has been pushed to late 2027 through 2028, later than the market previously expected.
Two gates block mass production: wafer-to-wafer (W2W) hybrid-bonding tools are not yet mature, and the equipment supplier for the second test insertion (Insertion 2) has not been decided.
This means → CPO is not a "can't build it" problem. The technology works, but yield is too low and tooling is not in place for economical volume shipment.
Where exactly does the 40% yield fall short?
TSMC's COUPE platform uses hybrid bonding — a process that joins two chips face-to-face — in two steps: die-to-wafer (D2W) yield is roughly 80%; wafer-to-wafer (W2W) yield is roughly 50%.
Multiply the two steps and cumulative yield lands at about 40%. In plain terms = out of every 10 units produced, only 4 pass. Volume economics require each step to hit at least 90%.
W2W is the bigger problem: lens surface unevenness traps gas at the interface with the flat carrier, forming voids. During annealing, defects can spread to adjacent dies.
Who supplies the tools, and how fast can they deliver?
D2W tools come from Besi, at roughly $3 million per unit. TSMC's installed base already covers the 2027 target of 100,000 switches, requiring about 14 Besi tools.
If 2028 demand reaches 450,000 switches, roughly 61 D2W tools will be needed. This means → Besi's delivery pace will directly set the expansion tempo.
On the W2W side, Tokyo Electron (TEL) leads in development. EV Group has a platform too, but TSMC has concerns about its bonding-alignment performance. The current plan favors evaluating TEL while keeping EV Group as a backup.
Can vacuum tools fix the yield problem?
W2W tools equipped with a vacuum chamber can pump out residual gas before the surfaces make contact, potentially lifting yield from roughly 50% to around 90%.
But development timelines vary widely: the baseline estimate is two to three years; under favorable conditions it could shrink to 18–24 months. The earliest demo tools may appear in H2 2027; whether mature production tools are ready by 2028 remains uncertain.
Experts estimate each vacuum tool costs roughly $30 million; actual pricing and throughput are still unverified. In plain terms = this machine is both expensive and unfinished — the single biggest point-of-failure risk on the entire production line.
CPO's delay — whose window gets longer?
FUNDA notes: the later CPO matures, the longer the application window for near-package optics (NPO) — an approach that places optical modules close to the chip without packaging them together — in scale-up interconnects.
Nvidia's next-generation scale-up interconnect is expected to use an optical solution, but the choice between CPO and NPO has not been finalized. Existing manufacturing constraints have already reduced the likelihood of adopting CPO for early-stage scale-up.
This reflects a broader dynamic: CPO's delay is not just shifting its own timeline — it is reshaping the competitive landscape among optical-interconnect architectures.
Can the 2028 ramp actually materialize?
From PIC wafer (photonic integrated circuit wafer) to final qualified optical engine, experts estimate cumulative yield at only 30% to 40%.
Low yield and integration difficulty have already pushed some customers to consider delaying full CPO adoption to 2028.
This means → whether W2W vacuum-tool supplier selection can be completed in H1 2028 is the critical verification milestone for CPO's volume ramp. The entire chain's timeline depends on this single decision.
市场有风险，内容仅供研究参考，不构成投资建议。