Generative AI’s spending boom has not only driven increased demand for data centers as hyperscalers work to expand capacity, but also is driving a surge in optic connections. This is due to the heavy data requirements needed to support genAI services and connections needed to link GPUs together in the clusters. Optical transceivers are becoming increasingly important in addressing bandwidth, a primary bottleneck in large-scale data centers, which refers to the speed of data transfer in the data center.
Corning is a centuries-old bellwether for materials such as advanced optics. The management team recently explained in its Q2 earnings call that “traditional data centers contain a network of interconnected switches and CPUs. GenAI requires a second network within data centers to connect every GPU to every other GPU in the cluster, creating a neural network. Now because GPUs have more processing capacity than CPUs, they need higher bandwidth links connecting them. The result is about 10 times the number of fiber [optic] connections in this new network versus a traditional data center.”
We’ve previously discussed memory bandwidth as a primary bottleneck at the GPU level, but scaling from singular GPUs to hyperscale data centers featuring hundreds of thousands of GPUs sees bandwidth arise as a primary bottleneck due to the immense data transfer requirements of AI training and inference.
GenAI to Drive Optical Growth
AI requires interconnected processors, to where thousands or tens of thousands of processors are connected. In turn, connectivity is needed for accelerated compute, which includes an increased number of switches, NICs, ports and also port speed.
For example, Chat-GPT was trained on a 25,000-accelerator cluster with roughly 75,000 optical interconnects. Increasingly powerful AI models, with escalating data and compute demands, are making bandwidth, data speeds and latency critical factors for future data centers to address.
Copper had long been standard for data center interconnects, but it cannot support network speeds of 800 gigabits (800G) to 1.6 terabits (1.6T) over long distances due to substantial signal loss. This isn’t to say copper is dead – Nvidia’s GB200 NVL72 utilized copper over optics (with more than 2 miles of copper cabling in the rack) to reduce power consumption by 20 kw (the system still draws 120kw of power). According to a representative from Marvell’s Cloud Optics division, “optical is the only technology that can give you the bandwidth and reach needed to connect hundreds and thousands and tens of thousands of servers across the whole data center.”
Optical transceivers are crucial in enabling high-speed data transfer, by transmitting and receiving data from optical (light) signals to electrical signals. In data centers, optical interconnects and transceivers are becoming the de facto standard to handle AI workloads, since they can function at significantly higher speeds than copper (currently at 800G+ speeds and moving quickly to 1.6T), with longer range, higher data capacity, and lower latency with minimal signal loss. One drawback, however, is that due to the electronic complexity of optical products, costs are higher as well as power consumption versus copper.
800G transceivers are a driver of growth industry-wide at the moment, with Marvell, Lumentum, Coherent, Mitsubishi Electric, Broadcom, Nvidia and others all shipping 800G transceivers and seeing high growth and demand. Coherent forecast in 2023 that the datacom transceiver market would more than double to $11.4 billion by 2028, with 800G transceivers taking more than 50% market share, up from the mid-teens in 2023, with the majority of that growth arising through 2026.
Source: Coherent
Mitsubishi Electric sees much more market share growth ahead for high-speed transceivers, forecasting the optical transceiver market to nearly triple from just over $4 billion in 2023 to $12 billion by 2029, with 800G and 1.6T transceivers accounting for more than 80% of the market.
Source: Mitsubishi Electric
Other industry forecasts suggest the broader optical transceiver market (including other end markets in industrial and telecom) will nearly double from $13.3 billion in 2024 to $24.7 billion by 2027.
Industry Executives See Strong Growth in Optics
Management teams from companies in the optical transceiver industry remain quite bullish about growth prospects in the future, catalyzed by AI data center demand.
Mitsubishi Electric, which reportedly commands nearly 50% market share in optical transmission devices for data centers, per Bloomberg, is rapidly expanding capacity to meet demand. Mitsubishi is “ramping up production capacity for optical devices to a level 50% above last year’s,” though CEO Masayoshi Takemi said that “won’t be enough to meet the strong level of inquiries we’re getting, [and] we may need double what we’ll have in September.”
Coherent CEO Jim Anderson similarly sees strong growth ahead for transceivers: “one of the most exciting growth opportunities is our optical transceiver technology, which underpins and drives the high-speed connectivity required by new AI data centers.” He added that the company “saw strong sequential growth in our 800G datacom transceiver revenue in Q4 and [is] also seeing increasing orders in backlog for the current and future quarters. We also delivered initial samples of our 1.6T datacom transceivers, which we expect to begin ramping in calendar 2025.”
Barclays analyst Tom O’Malley asked Anderson about the trajectory of the “big growth engine” for Coherent, the ramp of 800G and soon 1.6T transceivers, with Anderson saying that “It's stronger than what I had thought. And we've seen, just over the last, I would say, gosh, four to six weeks as I've spent a lot of time with our top customers across the – across all of our different product lines, but especially in our datacom business, I've gotten a much better sense for the opportunity that's in front of us, and I would say it's a very strong opportunity. And we continue to see demand strengthening, forecast strengthening, billings, backlog.”
Optics has also been a primary growth driver so far this year for Marvell, with electro-optics revenue exceeding expectations in Q2 and expected to grow in each quarter of this fiscal year. To note, Marvell is targeting to exceed its $1.5 billion AI revenue target this year, with optics contributing $1 billion or more of that sum.
Marvell CEO Matt Murphy explained in Q2’s earnings call that “demand has been extremely strong in the AI business, as we mentioned, both in custom and in our optics business. And that's the 800G products as well as traditional cloud, as well as DCI [data center interconnects]. So that's all going extremely well. And for next year, that should absolutely ripple through. We see continued strength next year above what we had communicated relative to the target for next year both in custom and in optics and the broader portfolio. … Demand has been strong. Bookings momentum has been extremely strong.”
Marvell’s management also noted in Q1 that its 800G PAM4 modules are currently the “primary interconnect enabler for state-of-the–art AI deployments,” while qualifications have begun for its 1.6T modules, which it expects will enable the next generation of AI chips. Marvell added in Q2 that “strong bookings continue for our market leading 800G PAM products and 400ZR data center interconnect, or DCI products,” while its 1.6T DSPs (digital signal processors) would begin shipments in Q3.
Marvell has the advantage of scale over its competitors. Per a Marvell spokesperson in June, “every single large language model today runs on compute clusters that are enabled by Marvell’s connectivity silicon.”
Lumentum is expecting the optical opportunity to quickly drive quarterly revenues to $500 million by the end of 2025, up 60% from last quarter’s $308 million, with management saying that transceivers will be the “number one growth area,” with EML chips and optical switching other growth drivers as the company works to rapidly boost transceiver capacity in Thailand. You can access our previous Lumentum deep dive here.
In terms of optics (and/or networking revenue), Marvell likely leads the three, with Coherent close behind. Marvell’s data center optics segment alone contributed more than $1 billion in revenue in fiscal 2024, and in Q2 FY25, optics and networking and switches likely accounted for revenue in the mid-$600 million range, given management’s comments that suggested between $200 million to $250 million stemmed from ASICs and storage. For FY25, Marvell’s optics, networking/switches and enterprise networking revenue could reach approximately $3 billion annualized (~54% of FY25 revenue estimate of $5.54 billion), with more than $1 billion in AI revenue (mgmt’s $1.5 billion AI revenue target has 2/3 coming from optics).
Coherent reported approximately $2.3 billion in networking revenue in the twelve months ending in June, or ~49% of overall revenue; Coherent has not provided an AI revenue target, but noted that datacom revenue rose 16% QoQ and 58% YoY due to AI demand. Lumentum reported $1.08 billion in cloud & networking revenue in FY24 (~80% of revenue), down 18% YoY, with management eyeing AI to drive revenue to a $2 billion-plus run rate by the end of calendar 2025.
Interestingly, it was announced on September 5 that Marvell, Lumentum and Coherent have demonstrated the industry’s first 800G ZR/ZR+ pluggable modules for 500 kilometer data center interconnects (an industry first distance for 800G modules). Utilizing Marvell’s Orion 800G DSPs, modules from the three are now interoperable, allowing regional data centers to take advantage of a multi-vendor solution and minimize vendor lock-in risk by having the ability to link together transceivers from different companies in a cost-effective and power-efficient manner.
Blackwell in Focus
Nvidia’s Blackwell is a force of its own, and with initial shipments expected to begin in Q4 this year (which ends in January), but will ship in volume come Q1 (ending in April), analysts are working to identify which companies will be primary suppliers on the optics side as Blackwell brings an enormous revenue opportunity for Nvidia and its suppliers.
Broadcom was the latest to field questions from analysts about optics tie-ins to Broadcom, though Marvell, Lumentum and Coherent all have been questioned as well – the common denominator is that the management teams are not commenting on individual customer engagements.
Interestingly, a Broadcom announcement from March 2024 noted that “Google and Nvidia will be the first adopters of 200G per lane optics for interconnecting GPUs and TPUs in AI Clusters” as 1.6T shipments begin by the end of the year (aligning with Blackwell). Broadcom also continued its partnerships with Innolight and Eoptolink in optics; this cross-checks with a report from SemiAnalysis on GB200 component suppliers, saying that “while Marvell was 100% share on Nvidia last generation with H100. This generation, Broadcom comes in a big way. We see both Innolight and Eoptolink looking to be adding Broadcom in volume for the DSP.”
Notably, in the most recent earnings report, the CEO stated he was not “directly” in the market of supplying Blackwell, so we will see if Broadcom is downstream or not come next year. Per the CEO of Broadcom: “We’re happy to be part of that ecosystem as I said. But directly, we’re not in that [Blackwell] market as you know.”
Marvell:Marvell:
Question (Atif Malik, Citi): “Curious when are you thinking about the volume adoption of 1.6T and what is holding that if it's not the DSPs. Are the lasers not ready? Or is it just waiting for the Blackwell?”
Answer (CEO Matt Murphy): “I think the way to think about it is just timing relative to the system builds our customers call, schedule their ramp, et cetera. … I remember all these issues in the past, right. There's a green laser problem. There's this problem or that. There's always some issue in this optical space, but this time it's really, everyone's going a million miles an hour trying to get their products ramped. Our module partners are ramping up with our solution. Our end customers that are driving this are going as fast as they can.
So it's just more of a timing issue that we need to intercept the platforms as they're ramping and we're doing that. So think of that as sort of shipments in the back half, but really contributing much more meaningfully next year on the 1.6T transition. But it's definitely underway, and we see a clear path to help enable right, this part of this next generation of accelerators to be able to ship in volume with the latest optical standards. And we're at the forefront and in the lead in that regard. So yes, it will be later this year and then more volume next year and it will — I think be a big product cycle for us.”
Quick Note on CXL: About two years ago, our firm covered Marvell’s CXL memory catalyst. Compute Express Link (CXL) improves how data centers add memory by offering a new switch that offers “cache coherent” memory pooling. Essentially, this means offering a new architecture that boosts memory bandwidth and helps to enable memory pooling through partially-disaggregated racks.
The new fabric required for disaggregated memory from the CPU is based on PAM electro-optics that Marvell specializes in. In July, new CXL memory-expansion controllers were announced called Structera with partners AMD, Intel, Micron and Arm participating in the press release. Custom CXL silicon is expected to sample in the fourth quarter and will represent an expansion to Marvell’s TAM assuming all goes well.
Lumentum:Lumentum:
Question (George Notters, Jefferies): “I'm just curious if you guys have an Nvidia qualification on this 800 gig single mode transceiver?”
Answer (CEO Alan Lowe): “Yes, we're not going to comment on who the customer is, George. I would say that — as I said before, most customers are working with us on products they don't already have. And so, for instance, we are designing 1.6T transceivers, and the performance is quite good. We plan on sampling customers this quarter on 1.6T. So, there's a few leaders that would be consuming that. And so, you can imply what you want from that, but we're not going to speak specifically about any individual customer.”
Coherent:Coherent:
Question (Vivek Arya, Bank of America): “Are you seeing any impact at all, positive or negative, because of changes in Nvidia’s product schedule or does that have no impact?”
Answer (CEO Jim Anderson): “On the first part, on the part that was about the order book, yes, we continue to see the order book strengthen. I think you asked about a particular customer. I can't comment on that particular customer, or really any particular customer. But I can say that in aggregate, we're continuing to see, again, the order book strengthen and demand growing, which is good.”
Margins May Determine the Winners
Given that optics is a fairly fragmented market with four major firms vying for market share in 800G and soon 1.6T transceivers, margins may ultimately determine the winners. This view is shared by Coherent, with management stating that on pricing and gross margins, “in general, what we would usually see in the transceiver market is the newer speed grades like 800G and then soon to be 1.6T generally carry higher gross margins than the older speed grades, right? The older speed grades are usually become commoditized over time.”
Based on Coherent and Mitsubishi Electric’s forecasts, shipment growth in 800G is expected through 2026 before shifting to 1.6T, leaving four to six quarters for these companies to drive shipments and revenues before commoditization potentially occurs with product margins shrinking.
Of the trio, Marvell has the best gross margin profile, at 46.2% last quarter, compared to 32.9% for Coherent and 16.6% for Lumentum. Marvell is expecting gross margin to expand to 47.2% next quarter, while Coherent guided flat QoQ and Lumentum guiding for some sequential improvement in future quarters.
Moving down the line, Coherent is the only one of the three with a positive operating margin, reporting a ~300 bp sequential expansion to a 4.8% margin last quarter. Marvell reported a (7.9%) operating margin, while Lumentum reported a (43.3%) operating margin.
With Marvell’s success likely equally tied to the ramp of ASICs in the coming quarters, and Lumentum deep in the red, Coherent is better positioned with a stronger bottom line profile to be able to withstand pricing competition, should the manufacturers prioritize capacity expansion. In this case, Coherent has leverage to boost market share gains by undercutting on price. Coherent also is showing slightly better sequential growth than Marvell, reporting ~10% QoQ growth in networking revenue to $680 million, while Marvell reported 8% QoQ growth in data center to $881 million (though Marvell is seeing growth arise from ASICs as well); Lumentum, on the other hand, reported an (11%) YoY decline as it struggles with weaker end market demand from telecom.
One primary theme evident in Big Tech’s recent earnings reports was the need for continual investments in AI infrastructure and physical data centers, with management teams positive on the long-term potential of generative AI products and services. Lead Tech Analyst Beth Kindig spoke with Yahoo Finance following Nvidia’s Q2 earnings report last month, saying that Big Tech is “in a race toward preventing extinction,” in that whichever company succeeds in AI first “could completely dominate to a level” to where competitors’ businesses will decline substantially. This is a view shared by Alphabet CEO Sundar Pichai: “the risk of under-investing is dramatically greater than the risk of over-investing.”
We’re seeing clear growth in Big Tech’s capex with no slowdown in sight – Bank of America is estimating that the Big 4 (Microsoft, Meta, Alphabet, and Amazon) will spend a combined total of $700 billion through the end of 2026. While a majority of AI capex is expected to flow to Nvidia and other AI accelerator beneficiaries including AMD and Broadcom, physical data center construction is surging, and outfitting new data centers requires AI server racks, cooling infrastructure, power systems, connectivity and other components. One component subsegment where we’re currently seeing growth arise, with positive forward-looking commentary from executives, is optical transceivers, with Marvell, Lumentum, and Coherent among the leading manufacturers we’re currently tracking.
Big Tech’s AI Spending, Physical Data Center Construction Surging
As we explained in our free newsletter in early August, “Big Tech Battles on AI: Here’s the Winner,” Big Tech’s capex spending is surging. Microsoft, Meta, Alphabet and Amazon committed more than $104 billion in the first half of 2024, up 47% YoY, with the four well on the way to spending more than $210 billion in capex for the year.
Spending is not expected to slow any time soon – UBS expects capex from the four to rise 25% YoY in 2025, well ahead of the current consensus estimates for 10% to 15% YoY growth, as AI demand still outpaces capacity as management teams push AI investments to the forefront.
The weight of four Big Tech CEOs speaking in unison on this topic (risking overinvesting and building AI capacity before it’s needed) is either a staggering coincidence — or they have important insights that are leading to the same conclusion, which is that AI’s primary risk is for those companies that are not early enough to capture it.
For a deeper understanding of Big Tech’s AI capex spending and outlook, and crucial comments on AI capacity and ROI, read our August newsletter here.here.
Stemming from this prioritization of expanding AI capacity comes a rapid uptick in physical data center construction, along with other signs that data center demand is rapidly increasing.
In North America, data center capacity under construction has soared more than 70% YoY to 3.87 GW through the end of June – for comparison, construction in all of 2023 totaled less than 3.1 GW. Preleased capacity surpassed 3 GW, with Big Tech and GPU renting startups such as CoreWeave accounting for more than 80% of this upcoming capacity.
In addition to this surge in construction activity, data from CBRE points to asking rental rates also rising, driven by tight existing supply and strong demand. Asking rental rates have increased 6.5% to $174 per kw/month this year, following an 18.6% rise in 2023 and a 14.5% increase in 2022. To put it a different way, asking rental rates have jumped 45% since 2021 (as construction began to accelerate), from ~$120 per kw/month to $174 per kw/month.
Building and optimizing these new data centers to meet the increasing performance and efficiency demands of Nvidia’s and AMD’s next-generation GPUs is placing more emphasis on fiber optics and optical transceivers for high-speed and high-capacity data transmission.
The I/O Fund strongly believes investors should look for demand signals for AI, which is why we steered away from best-of-breed software companies where AI revenue was entirely speculative and became a trap for investors, such as MongoDB or Snowflake. Big Tech does not use these software platforms, and therefore, the demand equation had not been solved. Instead, we focused tracking capex as it relates to AI semis and data center buildouts over the past 3 years, and that remains our strategy until we see capex contracts – with optics fitting well within this strategy.
Conclusion
As the industry gears up for Blackwell’s imminent ramp into 2025, we’ll be closely monitoring sequential growth in data center and networking segments, along with management commentary about the growth trajectories in optics as it unfolds. Optical transceivers and interconnects are becoming a key component in AI data centers due to transfer speed and other benefits, with the industry set to more than double over the next few years as data center construction surges along with hyperscaler AI capex.
Our Advanced members have received technical analysis updates and a possible buy plan for one of these optics stocks, as well as a handful of other AI stocks in explosive growth trends such as AI PCs.
Damien Robbins, Equity Analyst for the I/O Fund, contributed to this analysis.
Please note: The I/O Fund conducts research and draws conclusions for the Fund’s positions. We then share that information with our readers. This is not a guarantee of a stock’s performance. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis.
Recommended Reading: