Bernstein: AI Data Center CPO Scaling Pushed Back Beyond 2030
nashnova research
Bernstein has pushed the co-packaged optics (CPO) scale-up timeline from the market's prior 2027–2028 consensus to after 2030, with full maturity possibly not arriving until 2035 — reshuffling the benefit window between optical supply chains and high-end copper-foil makers.
How far has CPO been delayed — and why does it matter?
Bernstein's Oct 15 expert-call transcript: Nvidia's CPO switches ship in summer 2026, but meaningful industry adoption starts only after 2030; full maturity may take until 2035 and beyond.
This means → compared with the prior market consensus of 2027–2028 scale deployment, the timeline has shifted back by two to three years. Near-term growth narratives for optical-module companies need recalculating.
The call's key guest, Brad Smith, formerly ran Nvidia's copper-cable and optical-interconnect portfolio. His framework: use copper inside the rack as long as possible, phase CPO in around network switches, and run single-mode fiber for long-haul rack-to-rack links.
How long can copper hold — where is the physical ceiling?
One core physics curve: the higher the signal rate, the shorter the distance copper can reliably carry — 25 Gbps ≈ 7 m, 200 Gbps ≈ 18 in, 400 Gbps ≈ 12 in — copper starts to "vanish."
Nvidia's Vera Rubin rack contains roughly 5,000 short copper cables, longest trace about 18 inches. Replacing all of them with optics would add roughly 60 kW of power and about $2 million in cost. In plain terms = as long as GPUs sit within 18 inches of each other, copper is the cheapest, lowest-power, lowest-latency option — there is no reason to switch.
Morgan Stanley flags the key constraint: at 200 G/lane, copper links struggle to cover a scale-up domain beyond roughly 100 accelerators. A single Nvidia rack with 72 or 144 GPUs stays inside copper's reach; NVL576 at 576 GPUs exceeds it. This reflects a crucial distinction — what triggers optical entry is not rack size but scale-up domain size.
What three bottlenecks are blocking CPO scale-up?
Yield: TSMC's COUPE optical engine achieves roughly 95% per-unit yield, but a switch needs 32 engines — system-level yield drops to about 20%. In plain terms = fiber cannot survive solder reflow heat, so it must be attached afterward; one misaligned fiber kills the entire switch.
Testing: single-mode fiber core cross-section is hundreds of times larger than a silicon waveguide — tiny misalignment causes sharp insertion-loss swings. Ring-modulator resonance drifts with temperature, and the buried oxide layer in silicon-on-insulator conducts heat roughly 100× worse than silicon. No unified test standard exists; Nvidia will add end-to-end opto-electronic testing only from 2027.
Serviceability: if any single optical engine fails in a CPO switch, the entire switch — or even the accelerator package — may go offline. This means → one component failure equals one machine failure, far higher operational risk than designs where modules can be swapped individually.
CPO is delayed — is optical-interconnect demand delayed too?
No. Near-package optics (NPO) — moving the optical engine from the front panel to a board socket right next to the switch chip, without co-packaging — lets operators attach fiber first, test it, then plug the engine in. Bernstein says NPO captures most of CPO's benefits and is easier to manufacture; industry reports suggest it could be ready by 2027.
Huawei data backs this up: its Hi-ONE NPO design doubled mean time between failures, lifted system availability to 99.8%, replaced 48,000 800G optical modules with 5,500 Hi-ONE engines, and cut system power by more than 550 kW.
This means → NPO is becoming the bridge technology while CPO matures. Optical-interconnect demand has not disappeared — it is arriving through a different technical path first.
How long will pluggable standards last?
On Oct 12, Nvidia and Cisco jointly released the USFP pluggable optical-module standard — the first pluggable form factor supporting 400 Gb/s electrical signaling per lane, with initial aggregate bandwidth of 6.4 Tb/s, scalable to 12.8 Tb.
Goldman Sachs noted on Oct 14: Nvidia personally championing the 400G-generation pluggable standard signals that pluggables will persist for at least one more generation. This reflects an effective "official endorsement" of the pluggable path until CPO matures.
Put simply = Nvidia is voting with its actions — it is not betting on CPO for all interconnects in the near term, and is letting pluggables run the next leg.
What does this mean for the materials chain — copper foil or optics?
CPO shrinks high-frequency traces from tens of centimeters to a few millimeters, allowing copper-clad laminates to drop to M6-grade and HVLP2–3 copper foil. USFP, by contrast, retains a tens-of-centimeters, rate-doubled 400G electrical channel from switch chip to front panel — demanding HVLP5-grade or higher ultra-low-loss copper foil.
This means → the longer pluggables persist, the more the copper inside switches must upgrade to high-end grades — stretching the benefit window for premium copper-foil makers.
The CPO delay ultimately redistributes the benefit timeline between optical supply chains and high-end copper-foil material chains: copper-foil makers gain an extra two to three years of elevated demand, while optical-module makers' breakout period shifts correspondingly later.
市场有风险，内容仅供研究参考，不构成投资建议。