TL;DR
· Morgan Stanley expects that the AI scaling network opportunity in 2030 will be about $70 billion, expanding more than 4x from last year’s estimate.
· In 2026-2027, scaling networks will still be dominated by copper cabling, with CPO only potentially reaching 20%-30% penetration by 2029-2030.
· Keysight Technology, Astera, Broadcom, and Semtech will benefit earlier, while Corning, Lumentum, and Coherent will see more resilience later.

In its latest report, Morgan Stanley estimated the 2030 AI scaling network market opportunity at about $70 billion, and put copper cabling’s lifecycle within AI clusters back in the spotlight.

This isn’t a story about a “CPO breakout right away.” AI clusters are moving from single racks to multiple racks, and GPU-to-GPU communication needs to be denser and faster. The total “back-end networking” budget is being enlarged. But before short-reach connections truly hit the upper limits in power consumption, distance, and bandwidth density, copper’s inertia for short connections remains strong.

The timeline in this report is fairly restrained: in 2026-2027, CPO penetration in scaling networks is near zero; modest introduction begins in 2028; and only by 2029-2030 could it reach a meaningful level of 20%-30%. The market opportunity is being raised significantly, but for optics to really capture the bulk of scaling networks, more GPU domains and a more mature supply chain still need to be in place at the same time.

The $70 billion opportunity comes from multi-rack setups—optics isn’t the first thing to be amplified

The core of this upward revision is that after AI clusters expand, the demand for connections inside servers and between racks increases clearly.

In traditional single-rack scenarios, GPU distances are shorter, and copper cabling still has advantages in cost, latency, and power consumption. For short-reach connections—especially within 7-9 meters—copper remains the most direct option. In recent years, stronger SerDes, retimers, PAM4/PAM6, and other technologies have continued to extend copper’s cable life, also repeatedly pushing back the timing of optical replacement.

The change happens after clusters continue to get bigger. As training and inference clusters expand from one rack to multiple racks, GPUs need to communicate across racks, and signal speeds move from 100G to 200G and 400G. As distance increases and rates rise, the difficulty of managing electrical losses, insertion loss, and noise also increases, and copper begins to approach performance limits.

2024-2030 back-end network revenue forecast; scaling network revenue rises quickly, with a 2030 market opportunity of about $70 billion.

For investors, this determines the order of benefits. It’s not necessarily CPO suppliers who benefit first, but chip and module companies that can keep copper running faster and farther; only when multi-rack clusters become more widespread will the upside of optical engines, passive photonics, lasers, and test equipment become more obvious.

2026-2027 is still the copper window; CPO won’t explode until after 2029

CPO’s appeal is that it brings optical components closer to the switching chip or compute chip, reducing the transmission distance of high-speed electrical signals on the board, thereby improving power consumption and bandwidth density. The challenge is that this isn’t just swapping one cable—it changes the split of responsibility across packaging, manufacturing, testing, maintenance, and the supply chain.

That’s also why CPO won’t break out across the board in 2026. In scaling networks, CPO penetration is near zero in 2026-2027, introduced modestly in 2028, and meaningful adoption is expected only in 2029-2030. At that point, if multi-rack GPU domain expansion progresses as planned, CPO penetration in scaling networks could reach 20%-30%.

CPO penetration by scenario/scale; scaling CPO only rises to 20%-30% in 2029-2030.

This leaves the copper cabling chain with at least a two-year window. Astera Labs’ Scorpio X-Series has already entered initial mass production shipments; Broadcom has connection opportunities in AMD MI400/Helios and the custom ASIC ecosystem; Semtech participates in the transition phase via CopperEdge low-power copper cabling and linear optical solutions.

More importantly, copper cabling and optics aren’t simple substitute relationships. Hyperscalers will mix DAC, ACC, AEC, AOC, NPO, and CPO based on distance, power consumption, cost, serviceability, and reliability. Even for short-reach, within-rack, and near-rack connections, a large amount of copper cabling may still be retained. CPO will more often handle the higher-density, longer-distance, and higher-power-pressure segments.

NVIDIA’s roadmap boosts optical demand, but the pace still depends on platform rollouts

CPO truly becomes important, and that’s directly tied to NVIDIA’s next-generation AI platform roadmap.

NVIDIA’s official technical blog shows that the Vera Rubin Ultra NVL576 will combine eight 72-GPU racks into a 576-GPU NVLink domain, using copper cabling and direct optical connections; the Kyber NVL1152 from the Feynman era targets larger-scale interconnects and uses a similar direct optical approach.

After GPU domains expand, optical engine demand won’t just increase linearly. In this report’s calculations, the number of optical engines per GPU could rise from about 2 currently to the 35-70 range. In other words, once an architecture switch occurs, the amount of optical content will increase significantly.

Compare XPU cluster scale vs. OE demand; GPU domains expand from 72 to 576/1152, and the number of OEs per GPU rises from 2 to 17-70.

This is also why Corning (GLW), Lumentum (LITE), and Coherent (COHR) are placed on the main line. Corning benefits from passive photonics and glass-related content; Lumentum and Coherent are more related to lasers, optical engines, and optical components. After Morgan Stanley includes scaling CPO adoption rates in its model, the earnings upside for related companies depends even more on the adoption pace.

But this is still “elasticity if adoption happens,” not revenue that has already been realized. There is also market disagreement about NVIDIA’s roadmap itself: some industry analysis says that parts of the Kyber or Rubin Ultra configurations may be delayed, while NVIDIA responds that the roadmap remains unchanged. For the optical chain, the key isn’t the name of a single product generation, but whether large GPU domains enter mass production as planned, and whether the XPU ecosystem outside NVIDIA adopts a similar connectivity path.

Keysight Technology is more like “selling shovels”; test equipment doesn’t have to bet on a single path

Within this main line, Keysight Technology (KEYS) differs from optical module companies. It doesn’t need to bet on whether copper cabling or CPO ultimately wins, because the more diverse the AI networking architecture becomes, the higher the demand for testing and validation.

Today’s AI back-end network hasn’t unified into a single standard. NVIDIA has NVLink and a follow-on expansion roadmap; the non-NVIDIA camp has UALink, SUE, PCIe, and different hyperscalers’ proprietary interconnect solutions. Every architecture requires signal integrity, bit error rate, interoperability, power consumption, and reliability testing.

As relayed by Investing.com, Morgan Stanley has upgraded Keysight’s rating from Equalweight to Overweight, and raised the target price from $350 to $400. The reasons include AI investment, networking architecture diversity, and higher testing demand for 800G, 1.6T, and 3.2T. Keysight’s AI-related revenue is about in the mid-teens percentage range of total revenue.

By contrast, optical component companies’ upside is more concentrated in CPO adoption rates and the specific platform rollout pace. If NVIDIA’s roadmap progresses smoothly, Corning, Lumentum, and Coherent will benefit more directly; if copper cabling continues to be extended in life during 2026-2027, Astera, Broadcom, and Semtech’s short-term certainty is higher.

CPO will eventually move into a core position, but hyperscalers aren’t ready to go all-in at once

The counterintuitive part of this report is that it both acknowledges that CPO will ultimately enter a core position long term, and stresses that in the short term, you can’t underestimate copper.

The barriers for CPO aren’t small. Hyperscalers worry about vendor lock-in: once optical components are deeply integrated into switching or compute packaging, subsequent replacement, repair, and multi-vendor procurement become more complex. Manufacturing yield, thermal management, maintainability, and quality risks also affect the adoption timeline. If the cost premium can’t be offset by power savings and improvements in bandwidth density, adoption will be delayed.

There are also architecture disagreements. NVIDIA’s roadmap may push a higher proportion of optical connections, but self-developed architectures such as Google TPU could use different topologies, potentially reducing reliance on traditional CPO solutions. While the non-NVIDIA XPU ecosystem creates opportunities for Broadcom, Astera, and others, standards aren’t unified, which also means the supply chain can’t ramp quickly with one single solution.

So, the $70 billion market opportunity raised here looks more like the back-end networking “total budget” is being expanded, rather than a single technology route having already locked in the winning outcome. In 2026-2027, copper cabling will still dominate within-rack and short-reach scenarios; after 2028, optics starts moving into more core positions; and only by 2029-2030 will CPO potentially form truly meaningful penetration in scaling networks. The easiest place for the market to misread is to equate “CPO will eventually arrive” directly with “CPO will breakout immediately.”

Click to learn about律动 BlockBeats recruiting positions

Welcome to join the official BlockBeats community:

Telegram subscription group: https://t.me/theblockbeats

Telegram discussion group: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia