Is CPO Overheating?

In the current era when AI infrastructure is being propelled to a development peak by the combined efforts of capital and the industry, any technological directions associated with "light", "interconnection", "bandwidth", and "power consumption" are highly likely to be labeled as "next-generation core technologies" and quickly become the focus of market pursuit. Among them, Co-packaged Optics (CPO) is the most representative example in this wave of enthusiasm.

However, just as the industry's discussion about CPO was heating up, Hock Tan, the CEO of Broadcom, poured cold water on this frenzy during the Q4 fiscal year 2025 earnings conference call.

"Silicon photonics will not play a substantial role in data centers in the short term."

Hock Tan didn't simply and crudely deny the future of this advanced technology. He explained in the subsequent Q&A session that silicon photonics and CPO are by no means "leapfrog" alternatives to existing technologies. Instead, they will be the ultimate choice that the industry will passively adopt only after the potential of the existing interconnection technology paths has been completely exhausted. He clearly sorted out the evolution logic of the entire interconnection technology: First, the large-scale implementation of rack-level copper interconnection, followed by the continuous iterative upgrading of Pluggable Optics. Only when both of these technology paths reach the dual limits of physical performance and economic cost will silicon photonics/CPO truly become an industry necessity.

"This day will surely come, but not now."

Of course, Broadcom is not the only one trying to cool down the hype around CPO. At the latest Barclays Global Technology Conference 23, a group of leading industry players including Arista, Credo, Marvell, Astera, and Lumentum all conveyed similar messages.

From "Insufficient Computing Power" to "Interconnection Anxiety"

In fact, in the past few years, the AI industry has gradually moved beyond the early stage of simply piling up computing power. Even NVIDIA, the king of computing power, is deliberately downplaying the emphasis on single-point computing performance and instead repeatedly highlighting the importance of interconnection, networking, and system-level architecture. Because in the era of large models, what truly determines the upper limit is no longer just the chip itself.

As the scale of model parameters and the cluster size expand exponentially in tandem, the core bottleneck of AI clusters has quietly shifted from computing power to interconnection ability. At this point, what tests industry giants is no longer just whether they can spend huge amounts of money to purchase more GPUs, but how to achieve efficient interconnection in ultra-large-scale systems. Factors such as whether the communication efficiency is high enough, whether the latency is controllable, and whether the system can operate stably and collaboratively are gradually becoming the key factors determining whether the computing power can be "put to use".

The industry has clearly foreseen this change. Hock Tan, the CEO of Broadcom, revealed in the earnings conference call that his customers are planning and deploying ultra-large AI clusters with over 100,000 GPUs. Arista further pointed out at the technology conference that the target of internal industry discussions has already pointed to cluster architectures with millions of GPUs. The judgments of multiple industry leaders tend to be consistent: When the scale of AI clusters expands from thousands of GPUs to hundreds of thousands or even millions, the network is no longer just a matter of "whether the bandwidth is sufficient", but has evolved into a set of systematic challenges, including interconnection reliability, power consumption ceiling, maintainability, and overall deployment rhythm.

For this reason, the keywords that the giants have repeatedly emphasized have also changed significantly in the past year:

Scale-Out, Scale-Up, Scale-Across, as well as the power consumption wall, link reliability, and system-level collaborative design... These seemingly "engineering-oriented" concepts actually mark a profound shift in the industry's understanding of the bottlenecks in AI infrastructure. The focus of AI competition is shifting comprehensively from computing power itself to interconnection and system capabilities.

Scale-Out Doesn't Require CPO

First, let's talk about Scale-Out. This concept is often referred to as a cluster system, which increases computing capacity by adding independent servers to distribute the workload, similar to NVIDIA InfiniBand.

For the giants, Scale-Out doesn't seem to require CPO at present.

Hock Tan's statement is the most representative: The 800G pluggable technology was launched in 2022, and its growth cycle will continue until 2026. The 1.6T products have entered mass production, and the growth is expected to continue until at least 2029. The 3.2T technology has been demonstrated. "In the next 5 - 7 years, pluggable optical modules will still dominate, which is why CPO technology in the Scale-Out field has not been widely adopted."

Yuan Wupeng from Lumentum further refined the market structure prediction for 2026: The total number of optical ports will reach 60 - 70 million, almost doubling year-on-year. Among them, 800G ports will account for about 55% - 60%, and 1.6T ports will account for about 15% - 20%. This data shows that even though the 1.6T technology is ready, 800G will still be the absolute mainstream in 2026, and there is no "radical leap" in the industry.

Chris Koopmans, the COO of Marvell, further strengthened this judgment: "The CPO technology in the Scale-Out field will eventually come, but it will still take several years. In markets with longer transmission distances and a need for interoperability, the adoption rate of pluggable products is slower." He specifically pointed out that customers have completed the software certification of 400G products, and when 800G products were launched, the migration from 400G to 800G was "almost instant". New deployment projects quickly adopted the 800G solution - this is exactly the ecological advantage brought by software pre-certification and the core moat of the pluggable architecture.

Arista also emphasized that at the 1.6T rate, "we are still confident in achieving the stable operation of Low Power Optics (LPO). At the same time, technologies such as Co-packaged Copper (CPO) are also under evaluation, and these are all potential differentiating advantages. Being the first to launch next-generation rate products and quickly commercialize them is one of our core differentiating strategies."

Is Scale-Up in No Hurry Either?

Scale-Up focuses on upgrading a single server or a chassis-based system by adding chips to the existing system to enhance computing power. It was once regarded as the "first battlefield" for CPO technology and the most promising application scenario for CPO.

However, in this scenario where CPO is most needed, the mass production schedule has also been significantly postponed.

After Marvell, the pioneer of CPO, acquired Celestial AI, it set a new revenue target: to achieve an annualized revenue of $500 million by the end of 2027 and double it to $1 billion by the end of 2028. It explained that this schedule is mainly based on the launch of the first product and the acquisition of the first customer. The 16-terabit chiplet product brought by Celestial AI has 10 times the performance of the most advanced 1.6T product in the Scale-Out field and has completely different form factors, densities, bandwidths, and technical characteristics. But even so, large-scale commercial deployment has been postponed to 2027 - 2028.

Jitendra Mohan from Astera Labs gave a more detailed time prediction: "As the system complexity increases, the data rate grows, and customers hope to expand the Scale-Up field from 1 cabinet to 2 - 4 cabinets, switching to optical technology will become an inevitable choice. We have carried out in-depth cooperation with customers on this intersection point and expect to achieve large-scale deployment in 2028 - 2029. However, the deployment of optical technology will not happen overnight. Some test deployments will be carried out in 2027 to prepare for large-scale deployment in 2028."

Why has the CPO application schedule been postponed even in the most needed scenario?

Yuan Wupeng from Lumentum gave an explanation from the perspective of the supply chain: "The CPO supply chain is still relatively new and needs time to increase production capacity to support large-scale demand. Therefore, the current supply limitation is not due to insufficient basic production capacity, but because the supply chain is not yet mature, and suppliers need time to adapt to the growing demand."

Bill Brennan, the CEO of Credo, analyzed from the perspective of production capacity: "The demand in the Scale-Up scenario will be several times the current demand. We are currently actively expanding production capacity, and the Scale-Up scenario will further significantly increase the production capacity demand. This requires the joint efforts of the entire industry. We are communicating with multiple customers on this issue and have already started relevant work."

The deeper reason is that the viability of various transitional solutions has exceeded expectations.

Mohan from Astera Labs admitted: "The reason why customers are reluctant to switch to optics is that optical technology requires higher power consumption and cost."

Broadcom emphasized: "We believe that CPO is the right technological direction, but we are not sure whether these products will be fully deployed because our engineers and industry peers will try their best to achieve Scale-Up through in-cabinet copper cables and pluggable optical modules. Only when pluggable optical modules and copper cables can no longer meet the requirements will silicon photonics technology become an inevitable choice."

Power Consumption and Reliability:

More Real Constraints Than Bandwidth

For the giants, power consumption and reliability are one of the important reasons why they are reluctant to fully embrace CPO.

Bill Brennan from Credo used a specific case to illustrate the severity of this challenge: xAI originally used laser-based optical modules to connect a cluster of 18 cabinets. Later, it planned to move to a liquid-cooled facility and reduce the number of cabinets from 18 to 6. The team asked Credo: "If you can produce 7-meter cables, we can build a 'zero-interruption' cluster." Because the reliability of the copper cable solution is well-known and is absolutely stable.

The concept of "zero-interruption" gave the industry great inspiration. Bill Brennan pointed out that since then, Credo's team has started to focus on solving the reliability problem, "especially for the link between the GPU and the first-level switch (T0). The core of the differentiating advantages of all our recently launched products revolves around reliability."

In the eyes of the giants, interconnection technology should first be reliable and controllable, rather than simply pursuing extreme performance. Predictability, diagnosability, and maintainability are often more important than the "theoretical optimum".

When talking about the ZeroFlap optical solution, Yuan Wupeng from Lumentum elaborated on the technical implementation of this concept in detail. He pointed out that Oracle, like xAI, faces the problem of link interruption, but their link length far exceeds 7 meters, so they can only use laser-based optical modules. The core of Lumentum's solution is to identify potential risks before link interruption occurs and mitigate them through proactive intervention.

Specifically, Lumentum redesigned the customized optical DSP to enable in-band communication - that is, while transmitting high-speed data, it can achieve two-way communication between DSPs to transmit telemetry data. Then, the pilot software is deeply integrated to convert the raw data into usable telemetry data. Finally, it is integrated with the customer's network through the switch SDK. "Now, we can provide customers with real-time and continuous telemetry data, including signal strength and receiving sensitivity, pre-forward error rate, and post-technical histogram. We can also identify electrostatic discharge (ESD) damage and detect dust on the optical fiber. This solution far exceeds the system-level capabilities of traditional laser optical modules."

At present, the upper limit of reliability for short-distance interconnection still lies in copper cables/AEC, and CPO is not a substitute. Its advantages mainly come from distance, density, and system observability.

Arista also mentioned the core issue of power consumption that the industry is concerned about in its speech. "The power consumption of our products is usually about 25% lower than that of similar products. When products are deployed on a large scale, this advantage will be particularly significant. Currently, power consumption is one of the core issues that the industry is concerned about, and low-power switches undoubtedly have strong market appeal."

Transitional Solutions

Are "Eating Up" the Narrative Space of CPO

It is worth noting that transitional solutions such as LPO, AEC, and ALC are continuously eroding and diverting the application space originally expected for CPO. More and more manufacturers have formed a subtle consensus in practice: the existing interconnection ecosystem is far from reaching its boundaries, and CPO is not the only answer, nor is it the endgame for now.

LPO (Linear-Driven Pluggable Optics) is regarded as an alternative solution. Arista said that LPO technology "can be said to be Arista's forward-looking innovation output to the industry." Its co-founder Andy Bechtolsheim proposed this concept and promoted its wide adoption in the industry. "Currently, 800G LPO optical modules have been deployed on a large scale. This technology brings significant cost advantages to customers: Since it does not require a digital signal processor (DSP), the capital expenditure is lower. At the same time, the power consumption is lower, and the operating expenditure is correspondingly reduced. Customers can use the saved power consumption budget to deploy more computing devices. We are confident in achieving the stable operation of LPO technology at the 1.6T rate."

The logic of LPO is "remove DSP → reduce power consumption and cost". For switch and optical module manufacturers, this is "architecture optimization" rather than "ecosystem reconstruction". Correspondingly, the logic of AEC (Active Electrical Cable) and ALC (Active Optical Cable) is to provide reliability close to that of copper cables and bandwidth close to that of optics in the most dense and critical interconnection range of 2 - 30 meters.

Bill Brennan from Credo emphasized that AEC has a very wide range of application scenarios in data centers. "In addition to various solutions to connectivity problems, there are a large number of innovation opportunities beyond the standard in both the Scale-Out/Scale-Up switch cabinets of the back-end network and the front-end connections. Therefore, more innovations will continue to emerge in the AEC field." He specifically pointed out that Credo's uniqueness lies in pioneering this market and "having to be deeply involved in every aspect of the industrial chain and taking responsibility for every aspect of the product" - this end-to-end control ability constitutes an insurmountable competitive barrier.

Marvell's "Golden Cable Project" is an extension of this AEC logic. Chris Koopmans from Marvell pointed out that cable products are essentially no different from optical modules, and hyperscale customers hope to achieve multi-source supply. "Our 'Golden Cable Project' is essentially a complete reference design, showing relevant enterprises how to build products that meet the requirements. The reason why we can achieve this is that our DSP has strong performance and powerful functions in PAM4 technology and can be adapted to various types of cables. We don't need to control and define the entire product end-to-end, which is exactly what customers want."

ALC technology is another important direction. Bill Brennan from Credo pointed out that when the industry finally needs a CPO alternative, "the micro-LED technology we invest in ALC will be directly applied to near-package optics. Its power consumption is only one-third of that of CPO, and it doesn't require the complex switch design as shown at current exhibitions. Therefore, we believe that when the industry finally needs a CPO alternative, our technical path will be more advantageous."

The List of Real Problems Facing CPO

Based on the statements of multiple companies, the challenges faced by CPO are highly consistent and concentrated in the engineering and commercial aspects rather than technical feasibility.

Maintainability is the most prominent pain point. Yuan Wupeng from Lumentum said bluntly that the core goal of CPO technology "after several generations of evolution is to reduce cost and power consumption. The industry is still solving problems related to maintainability. We believe that as the technology matures, CPO will be widely used." Hock Tan from Broadcom more clearly pointed out the three fundamental defects of CPO: "In terms of cost, CPO is definitely more expensive. In terms of reliability, laser-based CPO is far inferior to existing technologies. In terms of power consumption, it is not the lowest among all the technologies currently under discussion. These defects have led to the long delay in CPO mass production."

The complexity of system design is another obstacle. Chris Koopmans from Marvell pointed out that the optical interconnection technology required in the Scale-Up field is a completely different type of technology. It must be co-packaged directly with kilowatt-level XPUs and switches and has completely different form factors, densities, bandwidths, and technical characteristics. Jitendra Mohan from Astera Labs added that the optical link consists of three components: the electrical integrated circuit (EIC), the photonic integrated circuit (PIC), and the connector - "The connector is the key component responsible for coupling the optical output of the photonic integrated circuit to the optical fiber. Currently, it limits the large-scale implementation of optical technology (it is relatively easy to mass-produce millions of EICs and PICs, but there are challenges in reliably connecting optical fibers and other links)."

The uncertainty of the cost structure troubles all participants. Broadcom's judgment is the most direct: "In