Prepared remarks
Ladies and gentlemen, thank you for your patience. I would now like to hand the conference over to Mr. Dan O'Neil. Please proceed, sir.
Good afternoon, everyone. Thank you for joining our earnings call for the third quarter of fiscal 2026. Today, I'm joined by Bill Brennan, Credo's Chief Executive Officer; and Dan Fleming, our Chief Financial Officer. During this call, we will make certain forward-looking statements. These forward-looking statements are subject to risks and uncertainties discussed in detail in our documents filed with the SEC, which can be found in the Investor Relations portion of the company's website. It is not possible for the company's management to correct all risks nor can the company assess the impact of all factors on its business or the extent to which any factor or combination of factors may cause actual results to differ materially from those contained in any forward-looking statements. Given these risks, uncertainties and assumptions, the forward-looking events discussed during this call may not occur, and actual results could differ adversely and materially from those anticipated, implied or inferred. The company undertakes no obligation to publicly update forward-looking statements for any reason after the date of this call to conform these statements to actual results or changes in the company's expectations except as required by law. Also, during this call, we will refer to certain non-GAAP financial measures, which we consider to be important measures of the company's performance. These non-GAAP financial measures are provided in addition to and not as a substitute for or superior to financial performance prepared in accordance with U.S. GAAP. A discussion of why we use non-GAAP financial measures and reconciliations between our GAAP and non-GAAP financial measures is available in the earnings release we issued today, which can be accessed using the Investor Relations portion of our website. I will now turn the call over to our CEO. Bill?
Thanks, Dan, and thank you all for joining our third quarter fiscal '26 earnings call. I'll start by walking through our Q3 results, give an update on our business and share our view on our long-term opportunities. After my remarks, Dan Fleming, our Chief Financial Officer, will provide a detailed financial review of the third quarter and our guidance for the fourth quarter. We delivered record revenue of $407 million, a sequential increase of 52% and more than 200% from Q3 last year. We delivered a non-GAAP gross margin of 68.6% and generated approximately $209 million of non-GAAP net income. Over the past 18 to 24 months, maximizing network reliability and energy efficiency have been our core mandates as we built our roadmap and brought new products to market. In AI infrastructure, performance without reliability stalls clusters and scale without efficiency strains both economics and power envelopes. The strategy is clear: accelerate cluster bring-up, maximize XPU utilization and reduce total cost of ownership, all while providing our customers the highest reliability in the industry. Our recent performance reflects the most accelerated growth phase in Credo's history. From fiscal '24 to fiscal '25, we more than doubled revenue. And for fiscal '25 to the current year fiscal '26, we expect to triple revenue on top of that. That represents greater than 6x growth in just 2 years. Few companies, particularly in semiconductors, have scaled at that pace while maintaining consistent execution, healthy margins and product leadership. Our purpose-built SerDes MICs vertically integrated system model and deep hyperscaler partnerships win at scale. We established leadership in high reliability copper connectivity and built a strong position in optical DSPs and retimers. Now our strategy is to lead in reliability, power efficiency and signal integrity across the full spectrum of AI and data center connectivity from die-to-die links to chip-to-chip and board-level links to rack and long-distance optical and to resilient facility-wide optical solutions. By extending both inwards towards the silicon and outwards across the data center, we're positioning Credo to encompass the entire connectivity fabric of AI infrastructure. Each layer of connectivity is being fundamentally reshaped by demand for higher bandwidth and faster data rates. AI workloads continue to grow in parameter size, model complexity and cluster scale, driving sustained transitions from 100 gig to 200 gig per lane and the 400 gig per lane in the upcoming years. At the same time, architectures are becoming more complex, power envelopes are tightening and reliability requirements are rising. We believe the industry's persistent push towards higher speed and larger clusters continues to expand our long-term opportunity and our ability to win. I'll now discuss our business in more detail. Our AEC product line once again delivered strong growth, driven by existing customers and new wins, including our fifth hyperscaler. Demand is accelerating across both hyperscalers and emerging Neocloud providers. We continue to believe the industry is early in its AEC adoption. As AI clusters scale, reliability and power efficiency have become the primary design constraints. AECs are now the de facto standard for intra-rack and rack-to-rack connectivity up to 7 meters, increasingly displacing laser-based optical modules. Their reliability and power advantages are driving broad adoption. Our ZeroFlap AECs deliver up to 1,000x better reliability than commodity laser-based optics, while consuming roughly half the power. In XPU clusters where downtime can cost millions, network reliability matters. We're supporting large-scale deployments at 100 gig per lane today and expect a long tail deployment at those speeds. We're fully prepared to support strong industry momentum towards 200 gig per lane or 1.6 terabit ports. Our 1.6-terabit AECs will support Ethernet, UALink and ESUN protocols. Additionally, our PCIe Gen6 AECs are sampling now and will be released to mass production in the first half of fiscal '27. Our vertically integrated system-level model remains a key competitive advantage. We take end-to-end ownership from SerDes leadership in silicon innovation to system design and qualification, beat telemetry and supply chain execution, positioning us for sustained leadership. I'll now turn to our IC business, including our retimers and optical DSPs. Our IC portfolio spans both optical and copper connectivity across 50 gig, 100 gig and 200 gig per lane speeds. We expect strong optical DSP growth in fiscal '26 driven by 100 gig per lane deployments with increasing traction at 200 gig as customers prepare for 1.6T transitions. For Ethernet retimers, we're seeing significant growth with our 100-gig per lane solutions in both traditional switching fabrics and the rapidly expanding AI server segment. Our PCIe Gen6 retimers remain on track with fiscal '26 design wins expected to convert to production revenue in fiscal '27. Customer feedback has been consistently stellar. We're delivering an unequaled combination of industry-leading reach, latency and power efficiency. We're also excited about Blue Heron, our 200 gig per lane retimer that is purpose-built for scale of AI. It leverages our SerDes expertise to deliver long reach, energy efficiency and advanced telemetry with support for UALink, Ethernet and ESUN protocols. These IC solutions address a large and growing market opportunity. As the industry transitions to 200 gig per lane, we see substantial growth potential across multiple protocols. I'll now discuss our 3 most recent product families, where we've made meaningful progress since their announcement last year. At a high level, these products significantly expand our total addressable market by extending Credo's reach across the full spectrum of connectivity links inside the data center. I'm pleased to report that our progress with ZeroFlap Optics is ahead of schedule. As noted in our recent press release, we began production shipments with our first Neocloud customer, Tensor Wave. In addition, we're in qualification with 3 additional customers, including hyperscalers and Neocloud operators. At a high level, data centers today face major challenges with extended cluster bring-up times and uptime degradation created by the inherent link instability of commodity laser-based transceivers. Our ZeroFlap optics were designed to address these challenges directly. Through tightly integrated hardware, optics, firmware and our pilot software with switch level SDK integration, ZeroFlap optics deliver continuous link quality telemetry and autonomous detection and mitigation of potential link failures before they impact the cluster. This enables a step function improvement in network reliability. From a TAM perspective, ZeroFlap optics allows us to address optical connectivity spanning any length within the data center. Based on strong customer traction, we now expect to see a significant production ramp beginning in the first quarter of fiscal '27 and continuing throughout the year. Next, I'll discuss active LED cables or ALCs. ALCs extend our system-level ADC philosophy into mid-reach optical by combining Credo's connectivity architecture with the micro LED expertise gained in our Hyperloom acquisition. We're creating a new system-level product that delivers the reliability and power profile of an ADC with a thinner gauge optical cable capable of reaching up to 30 meters. This makes ALCs ideal for next-generation AI networks, where copper reach becomes limiting, and traditional pluggable optics introduce reliability, power and cost disadvantages. ALCs expand our TAM outward from short-reach copper into mid-reach optical, bridging the gap between AECs and conventional optical modules. We expect to sample and qualify our first ALC products in fiscal '27, and production ramp in fiscal '28. Finally, our OmniConnect line of products drives our reach inward towards the silicon to further expand our TAM. OmniConnect combines our purpose-built VSR SerDes with a family of gearboxes for XPU connectivity. Our first product, Weaver, enables up to a 10x improvement in memory beachfront I/O density which can reach up to 10 inches. By converting VSR to DDR, Weaver overcomes the physical fan-out constraints of traditional memory-to-compute interconnects. Our first OmniConnect customer, Pozitron, plans to leverage this architecture to deliver an inference XPU with 2 terabytes of memory capacity, enabling substantial bandwidth gains in memory-intensive workloads, such as real-time AI video generation. We expect the production ramp for the first OmniConnect gearbox to be in fiscal '28. We expect to introduce additional gearboxes over time to enable a composable architecture where the same XPU design can be optimized for inference or training workloads and be future-enabled as speeds or protocols change. We'll also develop an OmniConnect gearbox targeting near-package optics with micro LED that will address the reliability, serviceability and availability pitfalls of current CPO solutions, while at the same time, reducing power significantly. To wrap up on the business update, we're proud of our record performance and even more energized by the opportunity ahead. With continued growth in AECs and ICs and 3 new multibillion-dollar TAM expansions through ZeroFlap Optics, ALCs and OmniConnect, we've meaningfully broadened our near- to long-term opportunity. We remain confident in our ability to innovate, scale and grow in the expanding AI infrastructure landscape through our focus on delivering solutions with best-in-class network reliability and energy efficiency. I want to take a moment to express strong appreciation for our silicon operations and system product operations teams. They have done an outstanding job managing supply, scaling production and executing flawlessly in the face of significant upside demand from our customers. Their ability to respond quickly and reliably has not only enabled our record performance but has also become a distinct competitive advantage and truly the reason customers choose Credo. In an environment where execution matters as much as innovation, operational excellence is a differentiator. And with that, I'll turn it over to Dan Fleming for a detailed financial review of our Q3 and our Q4 guidance.
Thank you, Bill, and good afternoon. I will first review our Q3 results and then discuss our outlook for Q4 of fiscal year '26. In Q3, we reported revenue of $407 million, up 52% sequentially and more than tripling year-over-year and at the high end of our revised guidance range. Notably, our revenue again grew healthy double digits sequentially and reached new record revenue levels once again, a substantial year-over-year growth across 4 domestic hyperscale customers. Our top 3 end customers were each greater than 10% of revenue in Q3. As a reminder, customer mix will vary from quarter to quarter. We continue to expect that 3 to 4 customers will be greater than 10% of revenue in the coming quarters and fiscal year. And we continue to make progress in diversifying our customer base across hyperscalers, Neoclouds and other customers. Note that with product revenue representing the vast majority of total revenue, we will no longer break out product and IP as separate line items in our income statement. Our team delivered Q3 non-GAAP gross margin of 68.6%, above the high end of our guidance range and up 92 basis points sequentially. Total non-GAAP operating expenses in the third quarter were $77.4 million, above the high end of our guidance range due to our strong R&D investment and up 35% sequentially. Our non-GAAP operating income was $201.8 million in Q3 compared to non-GAAP operating income of $124.1 million in Q2, up demonstrably due to the leverage attained by achieving more than 50% sequential top line growth while OpEx growth was in the mid-30s. Our non-GAAP operating margin was 49.6% in the quarter compared to a non-GAAP operating margin of 46.3% in the prior quarter, a sequential increase of 327 basis points. Our bottom line once again demonstrated the substantial leverage we are delivering in the business. Our non-GAAP net income was $208.8 million in the quarter, a record high and a 63% sequential increase compared to non-GAAP net income of $127.8 million in Q2. Our Q3 non-GAAP net income quadrupled from Q3 of last year, which clearly demonstrates the magnitude of our top line growth, strong gross margins and our disciplined approach to scaling operating expenses. Our non-GAAP net margin was 51.3% in the quarter. Cash flow from operations in the third quarter was a record $166.2 million, up $104.6 million sequentially. CapEx was $26.5 million in the quarter, driven largely by purchases of production mask sets. And free cash flow was $139.7 million, up more than $100 million from the second quarter. We ended the quarter with cash and equivalents of $1.3 billion, an increase of $487.9 million from the second quarter, driven by the proceeds of our ATM offering, which began in October and ended in December and our strong free cash flow. We remain well capitalized to continue investing in our growth opportunities while maintaining a substantial cash buffer. Our Q3 ending inventory was $208 million, up $57.8 million sequentially. Now turning to our guidance. We currently expect revenue in Q4 of fiscal '26 to be between $425 million and $435 million. We expect Q4 non-GAAP gross margin to be within a range of 64% to 66%. We expect Q4 non-GAAP operating expenses to be between $76 million and $80 million, and we expect Q4 diluted weighted average share count to be approximately 197 million shares. These expectations are based on the current tariff regime, which remains fluid. As we look ahead to fiscal '27, we expect sequential revenue growth in the mid-single digits, leading to more than 50% year-over-year growth. And with that, I will open it up for questions.
Questions and answers
Our first question comes from Tom O'Malley from Barclays.
Bill, you mentioned that you saw a ZF Optics ramp in the first fiscal quarter of next year. And you talked about substantial size. Maybe you could compare what a ZF customer engagement looks like versus an AEC customer engagement? And then longer term, if you see kind of a similar pattern to what you've seen in AEC with the customers that you're mentioning, I think you mentioned 3 here, all representing some significant size? Or do you think there's more variation in the customer set when it comes to ZF optics?
Yes, I believe comparing customer activity with AECs is a good approach. We have been developing ZF Optics for nearly two years, and we are progressing well toward not only delivering the solution but also achieving a significant milestone. This is the first time an optical transceiver has been advanced to provide real-time telemetry data, enabling immediate decisions on identifying and addressing potential link failures before they occur. This enhances network reliability well beyond what is possible with standard laser-based optics. I want to emphasize that our products have undergone internal qualification, where we strengthen the solution before presenting it to customers for their qualification process. We provide customers with a thoroughly vetted solution for their evaluation. We have moved from providing samples directly into the qualification phase with customers, which is encouraging. Although we indicated last quarter that the ramp would occur in the second half of fiscal '27, we are now confident that it will begin in the first quarter, as we have already shipped production units. We are optimistic about this progression. We also announced our partnership with Tensor Wave regarding both AECs and ZF Optics. This announcement confirms that the portfolio we are delivering offers a higher level of reliability as our customers expand their clusters. Additionally, we are engaging with hyperscalers and other newcomers to the market. We are in the early stages of promoting this product and are very excited about the strong ramp we anticipate throughout fiscal '27.
Your next question comes from Tore Svanberg from Stifel.
Congratulations on the record results. Bill, maybe you could just level set us a little bit here. You mentioned we're still in very early stages of AECs. Obviously, there's a lot of excitement around CPO. So maybe you could just help us on what's driving some of the use cases for AECs right now. How should we think about those developing, especially in fiscal '27 and fiscal '28?
Yes. So I think the narrative on AECs is very similar to what has been played out up to this point. There are several areas within the data center network where AECs make a really compelling solution and really almost becoming de facto in an intra-rack as well as now more than ever, we're seeing rack-to-rack solutions that are within the reach of 7 meters. What's driving it is network reliability and power efficiency. And so I would only say one of our customers were really fully penetrated on all the swim lanes and those being GPU to host connections in the scale-out network, front-end connections within those same racks and then this aggregate in switch rack. Those are really the swim lanes that we've talked about. And so we see there's really great growth opportunity, not only for 100-gig per lane deployments as we see those increasing, but also as we see a shift to 200 gig per lane, it's even a stronger value proposition at those speeds. And so that's going to help us drive more volume as well as there's an uplift in ASPs. And so you mentioned the narrative on CPO. And look, this narrative has been one that's existed in different forms for the last decade, started with mid-board optical modules that have moved on to onboard optics, it's moved on to many different acronyms over time. And the bottom line is, just recently, I think there's been a bit of a signal-to-noise ratio issue in the market. And the noise right now is dominating the signal. So it's not an either/or type of situation. It's about deploying the right technology at the right reach and the right power. And we see the industry is evolving even more so to a heterogeneous mix of short-reach copper, pluggable optics, near package optics and eventually CPO. And so the strong interest we've seen in ZeroFlap optical is a kind of clear indicator that reliability matters more than ever now as AI networks are the bulk of the deployments and as these clusters scale. And so the bottom line is that the way we see it that until NPO and TPO solutions can deliver bulletproof reliability, deployments are going to be somewhat limited, which is why many of the forecasters show low single-digit share in the switching market over the next 3 years. Our investments are heavily focused on reliability. And so when we're talking about technologies that will deliver the higher density reach promised by CPO and NPO, our focus is on delivering the same reliability as AEC and ZF Optics. And so hopefully, that gives you color based on your CPO comment.
Joseph Cardoso from JPMorgan has the next question.
Congrats on the results. Maybe just wanted to get an update on how you're thinking about the composition of the 50%-plus growth heading into next year, as we think about the AEC opportunity continuing to ramp, but also confluencing with the material ramps of other areas of the portfolio like the PCIe solutions, optical products, et cetera. Can this be a year where we start to see a more material contribution from the non-AEC offerings in the portfolio and where they can drive a more material portion of the mix as early as fiscal '27? Or is the expectation really that's more of a fiscal '28 story and beyond?
I believe we will see a different mix between copper and optical in fiscal '27, especially as ZF Optics begins to ramp up. With that in mind, we do anticipate growth in AECs and ICs, with new growth emerging from ZeroFlap optics. This will include the PCIe business we are developing within IC and AEC. In fiscal '28, we plan to introduce our active LED cables and our first gearbox as part of the OmniConnect family. That's what we are looking at for fiscal '28.
Your next question comes from Vivek Arya from Bank of America.
Just a clarification to Dan first on what drove the upside? Almost $60 million plus upside was there a one-off or anything else right in your projects in the reported quarter? And then, Bill, I wanted to get back to this question of how complementary versus competitive is AEC versus optical solutions because over the last 3 months, we have seen this massive divergence in the performance of stocks of your optical peers, and this morning, we saw NVIDIA invest in 2 of your optical peers. So why isn't that a very important right incredible pushback that the market for AEC might be limited? So I just wanted to get your views on where copper versus optical is competitive and where they are more complementary?
Vivek, so let me address your first question regarding what drove that strength. And I'll answer a question that wasn't asked as part of my answer. If you look at our top customers for the quarter, we've just continued to see strength across all of our hyperscale customers. In fact, our top three customers all grew sequentially from Q2 to Q3. So that really drove that growth. And our largest three customers in Q3 were also our largest in Q2, as you would expect but in a different order. Let me just talk briefly about our largest customer. They were 39% of revenue, and they were also the same customer that was our largest customer in Q1. So that was quite a large increase quarter-to-quarter for them. The second largest customer was 32% and they were our largest customer last quarter. And then finally, our third 10% customer was 17% of revenue, and that was the first hyperscaler that we had to ramp.
And regarding the discussion about AECs and optics, the main story hasn't changed. You mentioned NVIDIA, which has clearly stated that where copper can be used, it will be used. The reasons for choosing an AEC over a laser-based optical module are straightforward: reliability is the primary factor, followed by power efficiency and total cost of ownership. This reasoning remains consistent. As we move towards 200 gig per lane and 1.6T deployments, we anticipate a slight decrease in connection length from 7 meters to 5 meters as speeds increase. Our investments over the past few years have heavily favored optical technology, reflecting the significant demand in that area alongside the demand for AECs. Our strategy is unique, focusing on delivering exceptional reliability by utilizing real-time continuous telemetry on each link to identify and proactively address degrading links. Additionally, our work on active LED cables (ALC) aims to deliver a new class of optical products that match the reliability, energy efficiency, and cost-effectiveness of copper, with initial reach of up to 10 meters, and future expansion to 30 meters. This will create a distinctive new product category in the diverse landscape between copper and various optical forms.
The next question comes from Quinn Bolton, Needham & Company.
Given the current market uncertainty surrounding CPO and optical technologies, could you provide more details on two products? First, regarding the Blue Heron DSP for scale-up AEC connections, are you noticing any interest? Secondly, Bill mentioned in the prepared remarks an OmniConnect gearbox with an ALC-CPO solution that may be available in the future. Can you share any information on when we might expect this ALC-CPO solution to be released?
Sure. I want to highlight that most of our AI revenue is currently from scale-out, and we don't see any revenue from scale-up, which is still a relatively small market at this time. However, there is significant potential for growth in the scale-up market, especially as it transitions from rack scale to cluster scale, which is prompting many discussions. We have announced our first customer for the Blue Heron product, which is a 200 gig per lane retimer that supports UALink, ESUN, and Ethernet. We will also be developing AECs with this product. As the scale-out opportunities develop, we will have a comprehensive product portfolio available. Regarding OmniConnect, extending its architecture to include a gearbox that converts VSR to micro LED is a clear path forward. The collaboration with ALC will be the key proof point. Ultimately, we anticipate having a gearbox designed for VSR conversion to connect micro LED with a pigtail, providing a straightforward and lower-risk approach to a near-package optics solution. This solution is expected to be reliable and operate at a power level significantly lower than that of laser-based CPO.
The next question is from Sean O'Loughlin, TD Cowen.
I will add my congrats on a really incredible set of results. I had a quick clarification. I think last quarter, you mentioned that you expected the fourth hyperscale customer to represent greater than 10% of revenues for the full fiscal '26. Obviously, you mentioned 3% to 10% customers this quarter. Is that still your expectation for the full fiscal year? And then on the OpEx guide, I was a little bit surprised to see that it was almost flat quarter-over-quarter, obviously after a pretty big step up last quarter. But with all the irons in the fire, including the acquisition this morning. Is there just some constraints around I don't know whether it's hiring qualified mixed-signal engineers? Or is there something else going on in OpEx? Or am I just overthinking all of this and you're just executing to your roadmap?
Yes, let me start with the first question. Regarding our fourth hyperscaler, we mentioned this last quarter, and we've seen significant growth primarily driven by our largest customer this quarter. While the figures align with our expectations from 90 days ago, they may not contribute 10% of our revenue for the entire quarter or year. As for operating expenses, we noticed a significant increase this quarter in R&D spending, which comes after a relatively low expenditure in Q2. Additionally, project-related spending was higher than usual due to various ongoing initiatives. If that spending decreases, it could lead to some additional hiring to balance out the year or quarter. These factors illustrate the dynamics of our R&D spending from Q3 to Q4, which I hope clarifies things for you.
Vijay Rakesh from Mizuho is up next.
I have a question about the 1.6T ramp. As you approach 1.6T, it seems that most of the major hyperscalers have not discussed CPO much. Do you expect that as 1.6T ramps into calendar '27 and '28, it will mainly involve copper? Additionally, you mentioned that the increase in ASP should provide significant upside due to the adoption of copper and the ASP changes.
Yes. For the 200-gig lane per market, we very much see that, that market is going to be addressed by AECs. And then a combination of laser-based models, we'll have the ALCs that ramp into that market as well. But that would be what I would consider the new product category. I think CPO is still sometime in the future beyond that. We see our customers ramping 200-gig programs really at very different schedules. Of course, NVIDIA is going to lead the charge with Vera Rubin, but many other customers will follow on a slower time line. So we do expect to see very strong business in all 3 categories that I just mentioned. So we'll have ZF optics that are going to be delivered in that time frame. I will say from an optical DSP standpoint, we're getting a lot of uplift right now for LRO. Power is becoming a much, much more important thing as our customers go to 200 gig per lane. So I think we have a really nice position. You mentioned ASPs, and that's right. There is going to be an uplift from 800 gig to 1.6T across the board, across the entire portfolio. So we feel great about the way we're positioned there.
Your next question is from Quinn Bolton, Needham & Company.
The follow-up, I just wanted to ask you guys announced the Chimera acquisition this morning. It looks like that's kind of more layer 2 stuff, right, MAC, TCS, Mastek security. Are you buying that just to kind of enhance the IC product that you've done historically? Or is this a move to try to get into more Layer 2 solutions down the road?
I appreciate the question. We didn't have an opportunity to address this in our prepared remarks since the acquisition was announced just in time. However, we are very excited about integrating Chimera into Credo. We have been working with Chimera as an IP partner since 2022, and they have a strong reputation in protocol IP, error correction, and security IP technologies. This acquisition will enhance our ability to offer comprehensive system-level connectivity solutions. You mentioned the possibility of expanding our offerings, and yes, that is definitely part of the opportunity. We are strategically optimistic about this move, as having Chimera dedicated to Credo projects will accelerate our end-to-end connectivity roadmap and broaden our overall platform.
The next question is from Sebastien Naji from William Blair.
There's been a lot of focus lately on supply chain constraints, including the high cost of memory. I guess what type of supply chain risks are you seeing for Credo, if any? And is there anything in the supply chain that can emerge as maybe a gating factor to your growth in some of the coming quarters?
Yes. So I think we got a little bit out in front on this topic last quarter. I feel great about our supply chain for Credo, and that includes wafers in all of the different product categories that we've talked about. And that encompasses 12-nanometer, 7, 5 and 3. So we did a lot of work over the last quarter to make sure that we are aligned with our supply chain partners, not only on the wafer level but also the packaging level. So I think it's clear that we're going to be able to support our plan as well as upside that we expect. In the market, we are absolutely in kind of uncharted territory where I think supply chain is going to become more and more of a differentiator. And as it relates to the supply chain issues that are outside of our normal IC builds, I would say, from a system-level, there's no issues from a supply chain standpoint there. I will say at an industry-wide level, memory, as all of a sudden, has been a concern. And if anything, we can look at the first OmniConnect product Weaver as almost a solution to some of the pain points where we enable the use of DDR over HBM, which I think is probably the tightest area within the memory market right now. Outside of that, there's been a lot of conversation about lasers. But from a ZF Optics perspective, we feel that we've more than underpinned our demand for '27 and really beyond.
The next question today comes from Jim Schneider, Goldman Sachs.
Bill, it was helpful to hear you outline the progression of your various product lines, especially the optical products over the next couple of years. I was wondering if you could give us a sense of how we should be modeling the strength of those optical products and where we might end fiscal '27 in terms of their contribution. Could these products account for around 15% to 20% of total revenues of the company at an exit rate, or should we be modeling something considerably lower than that?
Yes, we haven't provided specific details on that yet. However, based on our Q4 guidance, we anticipate ending this year just above $1.3 billion. A 50% growth would bring us close to $2 billion next year. Bill mentioned our expectation for AEC to continue growing from fiscal '26 to fiscal '27, which we believe will be a significant part of our fiscal '27 related to ZF optics. As we progress and increase customer engagement with that product line, we will provide updates as we enter the new fiscal year next quarter.
Your next question is from Suji Desilva, ROTH Capital.
Congrats on the progress here. Just quickly, how many customers do you expect to be ramping ZF Optics across in the coming fiscal year? And just a longer-term question on the gearbox. You talked about being able to handle training and inference in the same architecture. I was curious about if you could elaborate on that opportunity. It sounds interesting.
Sure. Sure. Absolutely. You got to be confused with the second question. What was the first again?
ZeroFlap Optics, how many customers do you think you'll be ramping it across fiscal '27?
My expectation is that throughout fiscal '27, we will exceed four customers, as we currently have four. I foresee adding more to that list, which includes both hyperscalers and Neoclouds. Regarding OmniConnect, the key enabler is our VSR SerDes on the XPU side of the connection. Gearboxes are designed to mirror that VSR SerDes and connect to other systems. For memory, the first DDR gearbox will be for 5. As the market transitions to LPDDR 6, you won't need to replace the XPU; you can just swap out the gearbox. This will provide inference capability with next-generation memory. You could also develop a scale-up gearbox, initially combining Gen7 and Gen6, where the XPU maintains the same VSR SerDes, and the gearbox would convert 100 gig lanes to either Gen7 or Gen6 PCIe. When 200 gig per lane is ready for a specific customer, you could simply install a new gearbox supporting that speed with any of the discussed protocols like Ethernet, UAL, or ESUN. This concept can extend to scaling out as well, allowing for upgrading from a gearbox that handles 200 gig per lane to one that can manage up to 400 gig per lane without major revisions. This approach enables the construction of a composable XPU tailored to varying market needs, easily upgraded through gearbox enhancements for increased speed or protocol changes.
The next question comes from Christopher Rolland, Susquehanna.
I guess the first one is probably to you, Bill. Just about AEC applications and kind of where this may be moving around, if you could talk about where you think you're being used in terms of front end versus scale out, scale up like traditional cloud where you're being used today? And what this looks like over the next couple of years in terms of changes?
Yes. So I'd say the part of the network that's probably where we're strongest is on scale-out. And so this is where we really see the full benefit of AECs, as we're talking about leading edge speeds, and we're talking about in the part of the network where reliability really means faster time to cluster stability as well as continuous uptime. And so we do very, very well scale out. Front end kind of comes along with it. And then we're also seeing a couple of customers now that are deploying in switch racks or disaggregated chassis. So it's really across the board, but I would say our real strength is in scale-out.
Next up is Karl Ackerman, BNP Paribas.
Bill, perhaps a follow-on to that question earlier. You indicated much of your AI revenue for AC products is for scale-out networks; how should we think about the $5 billion TAM for AECs split between front-end versus back-end links between the server NIC networking switches? And Dan, could you speak to why gross margins are guided down roughly 360 basis points at the midpoint of your outlook? Is it just conservatism? Is it near-term product mix? Anything around that would be helpful.
Yes. Let me address the gross margin question first. So overall, as you mentioned in Q3, gross margin at 68.6%, up 92 basis points sequentially. We've really, over the last, say, 7 to 8 quarters, really seen a significant benefit to increasing scale. But we've also been very persistent in saying that the gross margin expansion will always be linear as we continue to increase scale. There will always be differences from quarter to quarter in product mix, and we are conservative in the way we forecast. We believe that we have not changed our long-term expectation in the 63% to 65% range for gross margin. And we've clearly entered this phase where we're at or above that high end of that long-term expectation. So it's really just a function of how we view the world and how we forecast our gross margin, and it's a very conservative forecast.
You inquired about the AEC total addressable market and the $5 billion figure. We don't concentrate much on top-down forecasts; we leave that to the market forecasters. However, I can share my view on the market opportunity. I believe a significant part of the market opportunity comes from scale-out networks, which may eventually include some share of scale-up networks as they are implemented. The front end will likely be smaller than scale-out, probably around 20% to 25% of the total scale-out market. The aggregate and switch market remains to be seen, but it could represent a substantial total addressable market if that architecture is widely adopted, which I believe is a strong possibility.
Tore Svanberg from Stifel.
I just had a follow-up. So this pull in of the optics business, Bill, I mean is that just mainly because of certain technical milestones that prompted this, or are there market dynamics? And the reason I'm asking the question because obviously, there's concerns about the availability of commodity lasers. So just trying to understand exactly what's driving that pull in by a few quarters?
The pull-in is driven by customer demand. This indicates that, as we've mentioned previously regarding AECs, reliability is crucial, particularly in terms of the time taken to establish a cluster and the expected uptime following that stability. This translates to a direct boost in productivity. AECs have gained popularity as customers recognize their value. When we discuss ZF with our customers and the potential for extending reliability into the optical domain, it’s quite rare for anyone to express disinterest. This consistent customer demand has bolstered our confidence that we anticipate the ramp to occur early in '27, in the next quarter. From a supply chain perspective, it is important to note that we have been preparing for this for two years. Our strategy has been to extend the AECs model into ZF optics, and we have been actively securing supply during this time. We have made solid commitments to our supply chain partners, and we feel very confident in our ability to ramp up production, even with the timeline advanced by 6 months.
And everyone, there are no further questions at this time. Mr. Brennan, I'll hand the call back to you for any additional or closing remarks.
Yes. Thank you. I really appreciate the ongoing interest and support in Credo. We'll talk to you all very soon. So again, thank you very much.
Once again, ladies and gentlemen, this does conclude today's conference call. You may now disconnect.