Description

History Repeats for Disadvantaged Data Center Connectivity Supplier Priced to Perfection

1) Introduction

The recent AI boom has created a rising tide for just about all semiconductor companies, and this enthusiasm is warranted for those with defensible businesses focused on data center infrastructure buildout. A rapid influx of AI compute servers inside the data center has created networking bottlenecks. The upcoming 800G infrastructure ramp is the first step in alleviating this constraint, which brings with it opportunity for advanced forms of both optical and copper electrical interconnects. However, one connectivity firm, Credo (CRDO) has been perceived as an AI “winner” with shares priced to perfection, up over 180% since NVDA’s first big AI earnings report in May 2023 – despite the fact that AI is accelerating the process by which its core product is losing its market-leading position.

Credo is a semiconductor connectivity firm, selling an active electrical cable (AEC), its core earnings driver, as well as SerDes (serial/deserialization) IP, line cards, optical digital signal processors (DSPs), and SerDes chiplets, all on a N-1 strategy. The company went public in early 2022 with revenue growing almost 100% year-over-year from marquee customers like Microsoft purchasing AECs. As data speeds increase, the use of direct attach cables (DACs) becomes untenable at longer lengths as the copper cable requires more material to send electrical signals, giving rise to increased data error rates from interference and potentially blocking air cooling to the servers. AECs solve these issues by introducing signal reconditioning chips at either end of a thinner copper cable, allowing for faster data rates with lower power requirements and cost than an equivalent optical cable. Credo was the first company to not only create the AEC but to also get the product to market. 

Unfortunately, the company’s one leading position is waning as its strategy in AECs and the connectivity market overall, is in complete opposition to how the data center industry is evolving. In this first wave of AI investment, Nvidia’s position in GPUs is translating to sales of their networking hardware, pushing out single products like CRDO’s AEC in favor of their homogenous optical-based tech stack. The counter to NVDA’s position includes custom silicon solutions, in which Credo plays a minimal part. In addition, competitors with superior SerDes IP (Marvell, Broadcom) are entering the AEC market CRDO created, and at lower margins. Importantly, Credo experienced an almost entirely similar situation years before going public in the active optical cable (AOC) market with these same players who dominate that space. Moreover, the AEC market is being compressed between short reach copper cables which are still capable at 800G, and optical cables making significant cost/power/reliability advances at shorter reaches.

In all, Credo is getting an unwarranted AI-winner valuation bump from investors unfamiliar with its history, trading at ~90x next year’s adjusted EBITDA with undeserved expectations of a massive increase in revenue over the next several years. In direct conflict to the company’s narrative of substantial growth, CRDO’s AECs are most competitive at 50Gbps/lane, a bit rate that’s already at a late stage of adoption in data centers. At 100Gbps/lane, the relative advantages of this cable diminish compared to substitutes. In addition, competition and the weaknesses inherent in Credo’s business model should result in disappointing revenue growth and earnings as early as this year, in contrast to the rosy expectations apparently embedded in Credo’s lofty stock price. We believe CRDO should trade on a more appropriate multiple for its quality, a generous ~17.5x EBITDA, and on a lower revenue base than consensus expectations, equating to 40%-80% downside.

 2) Investment Thesis

We recommend shorting CRDO on our expectations that elevated enthusiasm is likely to diminish when competitors’ 800G ramp begins in size later this year, and Credo’s slower pace of revenue growth and margin expansion from modest AEC adoption should fail to meet the market’s current overly optimistic expectations. While Credo portrays itself as a leader in the AEC space, its strategy puts its competitive standing at risk as the industry disintermediates whole-product suppliers, and as Marvell/Broadcom enter the market at lower prices. In fact, there is a very real risk that CRDO gets pushed out of the AEC market almost entirely, severely harming its equity value. It’s happened to them before…

Giving confidence to our variant view is the fact that this dynamic between Credo’s physical products and competitors Marvell/Broadcom has played out before, in active optical cables. Before going public, CRDO did an exemplary job carving out a niche in AOCs with its physical (PHY) optical devices, a market that was neglected by MRVL and AVGO in the mid-2010’s. However, once the market expanded, those two behemoths pushed Credo out. The company was left selling optical DSPs, only a mere component of optical modules. In fact, DSPs are 20%-40% of the cost in an optical module. Even if Credo had managed to maintain share in optical modules, which it didn’t, this disintermediation would have resulted in a revenue decline of 60%-80%. By 2018, CRDO successfully pivoted to its internally developed AECs, a new cable type not yet on Marvell and Broadcom’s radars. We believe history is poised to repeat itself.

Compounding the problem, as we’ll show, the AEC market as a whole may possibly remain niche as cost/power advantages wane with each new network node transition. In that scenario, AECs would end up squeezed between passive DACs and AOCs. We anticipate the years-away transition to 1.6T would only continue this compression, relegating Credo and its AECs to a small and highly-commoditized market, much as DACs are today.

We also believe Credo has exaggerated its AEC’s power advantages compared to AOCs. This is a minor point at 200G/400G, but for longs counting on AEC’s lower power consumption to be an advantage as network technology advances, we suggest a refresh of this assumption.

At Credo’s current price of ~$23/share, the market values it at ~90x next year’s adjusted EBITDA on an expected 50% increase in revenue to ~$300M. This is optimistic for several reasons, but even giving credit for growth less than 3 years out with $875M in revenue by FY 2027 and 20% adjusted EBITDA margins (current consensus), the name trades at almost 25x EBITDA.

This high valuation becomes truly egregious when factoring in stock-based compensation. Over the last two years, CRDO has generated ~$26M in adjusted EBITDA, while it issued ~$54M in SBC. That’ over 2x EBITDA. If the company maintains this pace of SBC, we anticipate almost half of its near-term adjusted EBITDA margins of 20%+ would be from this add-back. If the company doesn’t scale adjusted EBITDA margins above 30%, it should fail to generate any free cash flow over the next 3-5 years with capex needs running at ~10% of revenue.

With the market anticipating a 4x-5x increase in revenue on AI and 800G hype in less than 3 years, we believe Credo trades well in excess of fair value and requires perfect execution just to maintain its current price. Moreover, we think perfect execution is highly unlikely. The long thesis has multiple points of failure, each of which has a high probability of occurring in the next 6-12 months.  Materialization of one or all of these points would force investors to recognize between a 40%-80% write-off of company value.

There are several reasons for the current mispricing:

1. 1. Investors incorrectly view Credo as an AI beneficiary, when actually Nvidia’s GPUs and InfiniBand stack push Credo’s AECs to second-tier status. In February 2023, Microsoft, Credo’s largest customer at ~45% of sales, signaled to the company it intended to reduce its level of AEC purchasing. Subsequently, the stock dropped almost 50% that day and revenue decreased through mid-2023. While CRDO cited higher inventory levels as the cause to investors, in reality this was only partly the case, and a more concerning trend started to emerge: Microsoft was integrating more Nvidia GPU clusters in its data centers with low-latency InfiniBand cables. Nvidia, as part of its incredible moat, sells interconnects for its prized compute products that optimize the hardware for speed and efficiency. That means the best way to get the most out of the GPUs for AI applications is to buy almost everything from Nvidia.

However, industry experts had long pointed out InfiniBand’s reliability issues, but once Nvidia resolved those issues, it made sense for Microsoft to phase out CRDO’s AECs and go with NVDA’s passive DACs for the shortest reaches and multi-mode optical cables for longer runs. In our view, Microsoft’s core rationale for implementing Credo’s AECs as an interconnect was only as an Ethernet backup option to InfiniBand. By early May, Credo shares dropped into the low $7 range, a drawdown of ~65% in 3 months.

However, by the end of May, Credo’s structural concerns were completely forgotten. The share price more than doubled as enthusiasm swept over investors following NVDA’s earnings report that showed demand for AI-related hardware had started surging in the first quarter of 2023. To date, Nvidia has sold over half a million GPUs, with Meta and Microsoft buying 3x the number of chips in aggregate compared to Google, Amazon, Oracle and Tencent. As anticipated, these compute products have driven considerable sales of NVDA’s InfiniBand cables. With Nvidia’s stranglehold on not only AI compute (with 95%+ share in GPUs) but also between-server interconnect cables, Credo’s stock price appears to be driven by AI, but in actuality this first wave of investment is a threat to the company, with AECs being pushed out or relegated to a slower, lower-value subsegment experiencing less demand.

To that point, Credo’s revenue is only up ~10% since the start of the AI boom early last year, while NVDA, AVGO, and MRVL’s AI-derived revenues are up ~410%, ~500% and ~200%, respectively.  

2. 2. The company’s AEC leadership position is threatened by competition and disintermediation, just as it was in the AOC market years ago. In the mid-to-late 2010’s, Credo carved out a solid leadership position with their 56G-112G PHY optical devices for AOCs, less than a $1 billion market in 2017. However, as the market grew, Marvell and Broadcom came in with better engineering and lower prices. Other competitors with much smaller shares but greater flexibility developed their own modules and outsourced components like DSPs to provide customers with MRVL/AVGO alternatives. Credo was unable to keep up on the module front, and now only develops optical DSPs. Today there are many AOC suppliers, with Marvell and Broadcom at the top generating a combined billion-plus dollars in revenue from their best-in-class transceivers. Meanwhile, CRDO likely generates less than $50M in revenue from optical DSPs and zero from complete AOC modules, which they used to have.

What was just described for AOCs can be said for AECs today. Marvell and Broadcom have just entered the market, and smaller players like Astera Labs (ALAB) have some clever tech that goes beyond a basic AEC. Credo’s AEC market share is at risk – both its prices and the sale of the whole cable itself – and its AEC revenue could fall by as much as 60%-80%. But even worse, unlike the AOC market today, which has expanded to connecting everything in a data center outside of the server racks, the market for AECs could remain relatively small.

3. 3. The low-power advantage of Credo’s AEC all but vanishes compared to AOCs when data speeds exceed 100Gbps/lane, putting expected revenue growth from the 800G network transition at risk. Credo’s forecast of a large increase in AEC sales growth is predicated on the assumption that as the data center industry transitions from 400G (8x50Gbps/lane or 4x100Gbps/lane) to 800G (8x100Gbps/lane or 4x200Gbps/lane), DACs will become obsolete and hyperscalers are not likely to buy more expensive AOCs at these shorter lengths. The key selling points for Credo’s AECs, according to the company, are improved reliability and smaller size over thick copper DACs as well as 50% lower power demands compared to AOCs at similar reaches. But these claims are not entirely accurate across network generations.

At 50Gbps/lane, CRDO AECs dissipate over 4.5W/end, a 2x advantage over AOCs at 9W/end. But at 100Gbps/lane, the AECs consume 10W/end, while an AOC equivalent at 100Gbps/lane generates only 15%-20% more (~12W/end) on average. Some AOC providers state consumption of only 10W/end. Therefore, AECs power advantage is only truly at 50Gbps/lane, putting them in better standing against AOCs at 50Gbps/lane (200G/400G), but with only marginal advantages at 100Gbps/lane (400G/800G). If this increased power consumption stems from inefficiencies in Credo’s SerDes design, and if Marvell/Astera’s AECs consume less power at 100Gbps/lane, then the AEC market goes to these competitors. At present, CRDO’s AECs offer little benefit over AOCs at 100Gbps/lane other than upfront cost, being cheaper by less than half. If 5%-10% of data center capex is cabling, and less than half of that is for short reaches, then saving less than 50% on a portion of that amounts to only 1%-2% in overall budget savings.  

Even worse, Credo, in its marketing materials, is not being entirely accurate with regard to the power needs of its AECs at 50Gbps/lane or 200G/400G. It cites 4.5W of power dissipation (i.e. heat), but power dissipation is only one component of total power consumption. Assuming typical electrical resistance for a 7mm wide AEC copper cable with a 3-meter reach, total power consumption would be ~5.3W/end at a minimum, or 15%-20% higher than marketed.

As such, we expect Credo’s AEC-based revenue growth to face stronger headwinds than most expect, even though the company is signaling high adoption rates as the 100Gbps/lane (800G) transition heats up in the second half of the year. This same playbook of large-scale adoption was pitched at Credo’s IPO, but for the 400G transition, which has largely concluded. Ultimately, at 200Gbps/lane (800G/1.6T) and beyond, Credo’s main product faces an existential threat: AOCs with shorter and shorter reaches continue making inroads, shrinking the AEC market by 60% compared to 100Gbps/lane and leaving AECs viable at only 0.5-2 meters (as DACs are today).

4. 4. AI demand is compressing the 400G network lifespan, with hyperscalers quickly going to 1.6T. Almost all AI systems today are at 800G, while the remainder of a data center is generally at 400G. AI networks also appear to be upgraded every 18-24 months, as opposed to 3-4 years for traditional networking equipment. So, in the next 1-2 years, as AI stacks move to 1.6T with almost all optical connections, many network hardware providers see data centers jumping the rest of their network to 800G early, cutting off the useful life for 400G products. In fact, 400G hardware volume is expected to reach its peak in data centers later this year and decline rapidly thereafter as 800G hardware ramps in volume. Given that Credo’s AECs have their largest relative advantage at 50Gbps/lane, this curtailing of useful life at 400G puts them in a precarious spot.
   5. The cooling advantages of smaller-diameter AECs versus DACs are likely overstated, squeezing Credo’s AEC market between DACs and AOCs. At present, the AEC market size is estimated at ~$2 billion. The DAC market is less than $7 billion, and it’s implied by the company and others that at 800G and above, AECs will replace passive DACs. In opposition to this point, Nvidia just released a passive DAC for its 800G networking products at 2 meters and below. Clearly, passive DACs are good enough for NVDA at 800G. Notably, NVDA has an active copper cable (ACC) similar to AECs but prefers passive DACs at short reaches. This preference calls into question AECs’ competitiveness versus DACs at the shorter lengths left to them by AOCs. We believe the conflict in narrative stems from (1) the oversold cooling advantages of AECs and (2) passive DACs being 20%-50% cheaper across nodes.

Passive DACs consume ~0.1W/end of power. Meanwhile, AECs dissipate 4.5W/end of heat, and consume more power in total. So, while an AEC is ~7mm thick and a DAC is 12mm, the power dissipation differences are so vast that both copper cables at a 3-meter length have similar temperature rises of 4.8 and 5.2 degrees Celsius above ambient, respectively. But at 100Gbps/lane, CRDO’s AECs consume ~10W/end of power and generate temperature rises of ~10 degrees Celsius above ambient, twice that of a passive DAC. Even though their thinner cables allow greater airflow to servers, by generating the twice the levels of heat compared to DACs at 100Gbps/lane, AECs necessitate more cooling, not less.

This requirement of greater cooling makes AECs less attractive as a replacement for DACs at the shortest distances, even at higher data speeds.  Additionally, air cooling advantages in general are not expected to be a competitive advantage in AI server racks.  NVDA’s latest GB200 requires complete liquid cooling, and almost all AI new server racks should become 100% liquid cooled in the next 1-2 years. In sum, we believe Credo’s AECs are stuck between a rock (AOCs) and a hard place (DACs) with a market size of only a few billion dollars, not a combined DAC/AEC market of ~$9 billion.

6. 6. Credo’s N-1 strategy is the result of lagging competitiveness and is not a strategic advantage; the trailing edge still faces the same high costs as the leading edge, but with lower margins. Management itself will tell anyone that Credo products, from cables to IP, sell because of the quality of the company’s SerDes technology behind them. However, Credo started to lag in SerDes performance as data speeds increased, chips got smaller, and the traditional methods of designing a SerDes solution changed from analog to mixed-signal engineering. This transom was crossed a few years ago at 7nm, and since then CRDO has fallen behind even with mixed-signal solutions. Competitors today are developing designs on 2nm (not yet in production), while Credo’s SerDes are still at 4nm, putting Credo 3-4 years behind the leaders. Moreover, the R&D expenses required to stay competitive keep skyrocketing: The cost to develop an optical module at 65nm 10-15 years ago was ~$24M. At 7nm in 2020, it was $250M. Today, at 2nm, costs are estimated to be ~$750M.

Given this cost reality and lack of resources, Credo created its N-1 strategy, hoping convince investors its inability to keep up with the leading edge is a competitive advantage. Basically, the N-1 strategy relies on developing and selling IP and products just behind the leading edge. However, once leading-edge node technology is no longer leading-edge, it becomes more widely adopted, and customers expect lower gross margins and volume discounts. Credo argues that these lower margins are offset by lower costs – the supposed advantage to its N-1 strategy. In reality the trend of rising costs at each new node mentioned above remains intact at N-1 as photomask sets become more expensive and suppliers are less inclined to lower prices. Even a 25% discount on an N-1 node photomask still costs 125% more in capex than it would have 4 years ago, yielding a ~22% CAGR. In sum, Credo’s N-1 strategy puts revenues and margins under pressure, while capex remains elevated just to stay in the game, limiting the company’s ability to generate free cash flow.

7. 7. Recent share issuance and insider selling cast doubt upon Credo’s growth forecasts. There’s a risk to basing a short thesis on a company’s expected growth being insufficient to justify its valuation: Growth below trend could still satisfy the majority of investors in the name, if they remain assured in the future prospects. One reason why we’re confident Credo’s growth won’t satisfy most investors is that shares traded largely flat following its FY3Q 2024 earnings release that guided to year-over-year growth of 90% in FY4Q 2024 (February-April 2024). That’s an annualized revenue run rate of ~$240 million after recovering lost ground from lower MSFT sales, with revenue up only ~10% since Q1 of calendar year 2023.

Additionally, we note that management itself isn’t acting as though it believes strong growth is in the cards: Despite having ~$240 million in cash and burning just ~$25 million in free cash flow in the last 12 months, Credo issued ~10 million shares (~5% dilution) at $17.50/share in December 2023. Gross proceeds of ~$155 million went to CRDO, and insiders sold a little less than $20 million.

Following the secondary issuance in December, insider ownership dropped from ~35 million shares to ~30 million (~18% ownership). Insiders now own ~15% of the company, or ~26 million shares, down 25% from December. Most notably, the COO, Job Lam, has sold down his position by almost 50% in the last 4 months, from 9+ million to ~5.2 million shares, with a visible acceleration in sales after December with prices above $20/share. Since the end of Credo’s IPO lockup in mid-2022, insiders have sold down their ownership in the company by almost 60%, from ~35% of shares outstanding to ~15%.

If cash burn isn’t exorbitant and the company is expecting a large increase in revenue on the horizon, then it makes little sense to issue more shares, or for insiders to sell so aggressively.

The largest micro risks to our thesis are the scaling up of AEC sales 4x-5x as the market expects, strong sales of IP and optical DSPs, and a sale to a strategic buyer at a premium to normalized earnings. The largest macro risks are slower networking node transitions prolonging 400G hardware life, rapid development of Ethernet AI networking solutions, and less short-reach optical cable content than expected in data centers.

3) Business Analysis

Overview: Solid engineering firm that had head start in connectivity products, only to lose the leading edge

Credo was founded in 2008 by a pair of engineers from Marvell, Lawrence Cheng and Job Lam, currently its CTO and COO, respectively. The company had only 15 employees from 2008 to 2013. During this time, it developed SerDes IP and DSPs. In 2017, it debuted an active optical cable physical layer (PHY) device. AOCs were relatively new then, with the AOC market still less than $1 billion. Over time, the market expanded into the billions, and competitors Marvell and Broadcom came in. As all this was happening, CRDO developed its other physical product, the Active Electrical Cable (AEC), launching it in 2018.

The company since lost ground in the AOC market, both in market share and total addressable market. Today it sells only optical DSPs, a component of the optical PHY device. This optical PHY device or history in the AOC market was noticeably not mentioned in the company’s corporate history, nor in any sell side research, when Credo had its IPO. Management can hardly be blamed: If you’re trying to go public, it’s not a great idea to call attention to your legacy optical cable product that was disintermediated and lost its market-leading position right before saying that sales growth will be driven by a new cable with the same market dynamics and competitors as the previous device.  

Importantly, we understand from industry experts that Credo originally sold a complete AOC cable (just as today it sells a complete AEC cable), only to get pushed into modules by Marvell and Broadcom, and then pushed out again into DSPs as described above. However, we have not yet found any hard evidence that this whole AOC cable actually existed. If we did, we would use Credo’s AOC experience as a perfect analogy for what is now happening to it in the AEC market. Instead, we’re alluding to the similarities of the disintermediation of its physical optical modules because we have evidence that that actually happened.

While we don’t know how much Credo’s optical module-based revenue rose and fell, its lack of scale was evident by 2020, as the company that June raised a Series D preferred round worth $100 million. Notably, this was two years after the debut of its AEC, with the product and market still in their infancy. In the Series D press release, Credo said the capital raised would be used to “accelerate deployment of our current 400G products.” That 400G deployment mainly occurred from 2021-2023 and should be done by this year, with 800G deployments ramping aggressively.  

CRDO sells multiple IP products, chiplets, optical DSPs and of course AECs. AEC revenue over the last 12 months amounted to ~$100 million, or 50%-55% of total revenue. Microsoft is Credo’s largest customer at ~45% of total revenue, followed by Amazon at 15%-20% and Google at less than 10%. While more volatile, two large Chinese hyperscalers (Alibaba and/or Tencent) have amounted to ~20% of revenue in the past, though less than 10% today.

SerDes IP

SerDes is a method of high-speed data transfer in which data are channeled through specially engineered transceivers at each chip connection point. It’s a necessity for any semiconductor chip. SerDes takes all the data on the chip that is to be transferred, packages it up (i.e. serializes it) in the transceiver and sends it through an I/O port. It’s then broken back down (i.e. deserialized) in the receiving chip. This process results in less power consumption and EM interference, fewer I/O ports required to send a given amount of data, and ultimately a smaller, less complex chip.

Credo has SerDes IP solutions for 28nm down to 4nm with a wide range of reaches. Below 7nm, analog SerDes methods break down substantially, necessitating mixed-signal designs, with costs increasing almost exponentially. When the industry hit 7nm, Credo appropriately developed a DSP-based SerDes, but notably started to lag behind the leading edge. The current leading-edge node in development is 2nm.

Credo’s SerDes IP is used in almost every one of its products. Potential customers consider its performance versus that of competitors in deciding which company’s products to buy.

HiWire AECs

Active electrical cables (AECs) are copper interconnect cables with signal-reconditioning chips on either end. They were designed as a high-performance replacement for passive direct attach cables (DACs), and they’re power- and cost-competitive with active optical cables (AOCs). AECs come in two types: (1) Distributed disaggregated chassis (DDC) and (2) Server NIC (network interface card)-to-Top-of-Rack (ToR). Most use PAM4 signal modulation technology, with some utilizing NRZ. Use cases are shown in the diagram below.

Source: Credo.

DDC AEC cables are typically 0.5m-2.5m in length, for the shortest reaches between racks in a cabinet. These compete directly with DACs.

NIC-to-ToR AEC cables are mostly 3m-5m, specifically for switch-to-server racks inside a cabinet. These are likely to replace DACs, albeit with AOCs encroaching. There is a risk that as switch ASICs with higher I/O port density (radix) come to support a greater number of network device connections, an entire layer top-of-rack switches and connecting cables could be eliminated.

HiWire P3

The new Pluggable Patch Panel (P3) is a media conversion box that sits between the servers and coherent optical connections to offload power and reduce cooling requirements in a server rack.

Source: Credo.

Line Cards

Credo sells three retimer and gearbox line card products: Bald Eagle (400G gearbox), Black Hawk (800G retimer),  and Screaming Eagle (800G gearbox, 1.6T retimer). Credo also has 4 Media Access Control Security (MACsec) gearboxes and retimers: Owl 400/800 and Osprey 400/800.

Optical DSPs

A digital signal processor (DSP) converts analog signals into digital ones and compensates for the waveform distortions that accumulate in optical fiber transmission. Credo has two types of DSPs: 53Gbps/lane Seagull products (50G-400G) and 106Gbps/lane Dove products (100G to 800G). The company has only one transimpedance amplifier (TIA), the Teal 200 (200G).

Source: Credo.

Chiplets

The company has two SerDes chiplets, the low-power BlueJay (64x56Gbps/lane) and the Nutcracker (32x112Gbps/lane). Notably, Tesla appears to be the main customer for its chiplet products.

Management: CEO with minor position; insiders’ ownership down almost 60% since end of IPO lockup

CEO Bill Brennan joined Credo in 2013 and aided in scaling the company up from 15 employees to ~450 today. Prior to joining, he was VP of the Storage business unit at Marvell for 11 years. His compensation consists of ~$440K salary, a matching cash bonus, and ~$3 million in share awards, for a total of ~$4 million (~75% equity). He currently holds ~3.5 million shares (~2% ownership), which has declined ~18% since December 2023.

Founder and CTO Lawerence Cheng was Engineering Director of Analog Design at Marvell for 12 years before starting Credo in 2008. His compensation consists of ~$200K salary, a small cash bonus, and ~$1 million in share awards, for a total of ~$1.3 million (~75% equity). He currently holds ~10 million shares (~6% ownership), which has declined ~10% since December 2023.

Co-founder and COO Job Lam was Engineering Director for the Read Channel division at Marvell for 11 years before starting Credo in 2008. His compensation consists of ~$300K salary, a ~$150K cash bonus, and ~$1.2 million in share awards, for a total of ~$1.75 million (~70% equity). He currently holds ~5.2 million shares (~3.5% ownership), which has declined ~40% since December 2023. Since the IPO, Mr. Lam has sold 50% of his shares.

Bonuses are paid when the sum of revenue growth and non-GAAP net income margins exceed 45%. This non-GAAP net income margin incentive explains the issuance of stock-based compensation that has averaged ~16% of revenue the last two years, with non-GAAP net income margins of ~3%.

Insider ownership dropped about 25% from ~35 million shares in mid-2023 to ~26 million shares (~15% ownership) today. Since the end of the IPO lockup in mid-2022, insiders have sold down their ownership in the company by almost 60%, from ~35% of shares outstanding at the time of the IPO to just ~15% today, with selling continuing.

Customer Dynamics: Hyperscalers focus on power/performance over marginal cost savings, prefer disintermediated suppliers and second sources, capping supplier margins

Hyperscalers are massive-scale cloud computing platforms, providing a wide range of services over the internet. The biggest include Amazon, Microsoft, Google, Meta, IBM, Oracle, Alibaba, and Tencent. These customers have substantial negotiating power over suppliers.

In areas where a hyperscaler does not have the expertise or the desire to devote resources to internal development, they outsource not only hardware, but also software and other IP. In other areas, they may strike a strategic partnership or alliance, partly funding development by a third party of a bespoke product for their use only. Sometimes, they will take equity positions in these firms, with commitments on volumes. However, they will also generally either develop a second source internally or fund multiple suppliers. They almost never have just one source for a key product. Also, hyperscalers do business in multiple geographies, and can have differing architectures and tech stacks in each of them. Almost none has a single design spread throughout its data centers.

The higher cost of optical cables led hyperscalers to push for a disaggregated supply chain, similar to how supply chains across many industries have evolved in the last 50 years: Certain suppliers make components (lasers, drivers, DSPs, TIAs, etc.) that go into pluggable cable ends, while another provides the actual cabling. Hyperscalers then get an assembler of choice to put it all together, eliminating as much added cost as possible while also getting discounts on price for buying in massive volume. (Credo sells its products to OEMs, ODMs, and optical manufacturers, who in turn sell completed devices to the hyperscalers.) Notably, the transition from leading to trailing edge node typically results in a 25%-30% drop in prices.

Over time, anyone designing, building, and selling a complete product gets disintermediated. As shown above, this is essentially what happened to Credo and their optical PHY device product in the late 2010s, and there is no reason to suspect this same phenomenon won’t occur in the AEC market, which MRVL entered just last year.

Hyperscalers are increasingly becoming more power- and performance-conscious. As such, if a product doesn’t bring substantive improvements on those fronts, it’s unlikely to interest them by competing on price alone. This is especially true for cables, which make up only a small portion of a data center budget and have multiple alternatives.

Supplier Dynamics: SerDes and PHY engineers in short supply, cable components readily available

Credo relies on engineers with a specialty in analog and mixed-signal analysis and design. Throughout the semiconductor industry, analog experts (i.e. those skilled at translating physical phenomena into digital signals) amount to a small subset of practitioners, as most of the industry focus is on digital chips.

SerDes engineers are also increasingly rare, because they deal with the physical layer (PHY) on the silicon itself. This lack of engineering supply is also another factor in why many customers outsource SerDes IP. Their ability to hire engineers is paramount for success.

As for physical inputs to AECs, there are plenty of providers of copper cables. Wafers, which Credo’s cables’ microprocessors are built on, are manufactured at TSMC. Interestingly, Credo utilizes a handful of Chinese firms for packaging, assembly, and testing of AECs.

Competitor Dynamics: SerDes is what wins; market share to shift to big players, who can compete at hyperscale volume on lower margins. Minimal advantages for Credo AECs at 100Gbps/lane vs. AOCs, DACs

Cable Competitors

Credo’s main competitors in AECs are Marvell, Broadcom, and Astera Labs. Marvell started selling its Alaska 100Gbps/lane PAM4 DSP for copper cables in early 2022. Amphenol, Molex, and TE Connectivity sell AECs with Marvell Alaska DSPs. With testing taking a year or so, Alaska-based AECs most likely started production and delivery to hyperscalers in mid-2023.

Importantly, in December 2023, MRVL debuted its Perseus 110Gbps/lane PAM4 optical DSP for AOCs, specifically designed to replace passive 5-meter copper cables, consuming less than 12W/end at 800G. Previous iterations of AOC at 800G assumed consumption of 20W/end, so Marvell believes they have achieved a 40% power savings.

Microchip debuted its AEC retimers for 800G, a component of an entire cable, in late 2023.

Astera Labs sells its Taurus AECs for switch-to-switch and switch-to-server applications with reaches up to 3 meters. With a recent IPO, ALAB is likely to aggressively sell the product to keep share prices high during the insider lockup. Further, the company embeds smart monitoring tech in either end of the cable that should aid in maintenance and performance tracking, something Credo does not have. ALAB also appear more precise on power consumption trends in its marketing materials, stating its AECs offer “up to 50% less power than an AOC.”

While these physical products can have multiple suppliers for individual components, the companies that win with hyperscalers are those that provide a comprehensive suite of solutions (MRVL, AVGO, AWE.LN). Companies with one scalable product and little else to offer quickly get pushed to second-tier status. Competitors strategically offering a complete suite of networking products explains why they’re not only entering the AEC market today, but also plan to remain there even if it shrinks.

Notably, Nvidia also has an AEC alternative, an active copper cable (ACC). An ACC is a passive DAC with an equalizer IC in each end. At 100Gbps/lane it consumes only 1.5W/end at a 3- to 5-meter length. That’s 85% less power than an equivalent AEC, and on a thinner wire. In terms of technology, ACCs differ from AECs only in their means of conditioning the signal coming through them, which does introduce other issues with signal integrity.

Among DAC cable manufacturers, CRDO technically competes with Amphenol, TE Connectivity, Molex, Cisco, Dell, and HPE. At 100Gbps/lane, experts predict 2 meters may be the practical limit for passive DACs.

InfiniBand versus Ethernet

While all data center suppliers are seen as beneficiaries to increased investment from AI workloads, differentiation in networking protocols shows that is not entirely accurate. Nvidia’s GPUs are certainly the AI workhorse of choice with 95%+ share of AI servers in data centers. The company’s InfiniBand architecture is built to suit these GPUs with low latency, higher bandwidth, and now also with superior reliability to Ethernet. InfiniBand is estimated to comprise 90% of AI deployments. To get an AI network up and running, InfiniBand is the quickest and easiest to implement choice. Notably, Ethernet is the network protocol for the rest of the data center. 

Optical cables are the technology of choice for InfiniBand interconnects beyond a few meters. For the shortest reaches, Nvidia provides InfiniBand passive DACs, even at 800G (recently debuted). The latest NVDA GPU clusters use a combination of InfiniBand and Ethernet. However, Ethernet is only used for front-end networking, while InfiniBand (along with NVLink copper cables) is used throughout the AI cluster, both front- and back-end. Notably, the back end is more data/compute intensive, so the interconnect there is of higher importance in an AI workload.

InfiniBand cables are the connector of choice for AI deployments, given that Nvidia has a high share of the AI compute market. (It’s estimated to have a 3–5-year lead in AI over rivals.) That’s why AI enthusiasm for Credo is mostly misplaced: Credo’s AECs don’t support InfiniBand protocols, so it’s almost entirely missing out on the current AI investment wave. CRDO will be an AI beneficiary only when Ethernet evolves beyond its current latency issues to become the AI network technology of choice – a development that’s likely years away. It’s also possible by then that AECs will be relegated to the shortest cable lengths, only partially replacing DACs at sub-2 meter reaches.

Eventually, we anticipate Ethernet networking to rival InfiniBand. Ethernet is better for scaling, but first the AI ecosystem outside of Nvidia must develop solutions, which will take time. Demand is high for such an alternative, given hyperscalers’ preference for having second sources and multiple technologies in their stacks. And development is already underway: At Broadcom’s recent “AI Day” event, the company referenced a case study with Meta showing that AI clusters with Ethernet can provide performance superior to that of InfiniBand.

Cable Technologies

Given that almost 55% of Credo’s revenue is derived from AECs, which are also its largest growth driver, it’s critical to understand how technology formats change with each network generation. Below is a highly informative graphic from Marvell circa 2022, showing that as data speeds per lane increase, AOCs gain share while DACs lose share. Notably, AECs fill the gap between the two cable types at 100Gbps/lane, with DACs competitive at 1 meter, AECs at 1-5 meters and AOCs at 5 meters and above.

Astera Labs believes DACs are competitive at 100Gbps/lane at 2 meters, and Nvidia’s InfiniBand passive DACs have reaches up to 2 meters as well. If that is indeed the case then at 100Gbps/lane, AECs may only have a viable market of 2-5 meters, a ~25% smaller market than displayed below.

Source: Marvell circa 2022.

The industry isn’t stopping at 100Gbps/lane, the next generation (200Gbps/lane) is only a few years away. Given our discussions with a few industry experts, we think cable technologies at 200Gbps/lane would look like this.

Source: Internal estimates.

DACs would likely not die out entirely, with a viable reach of 0.5 meters, or possibly even 1 meter. AOCs would continue to improve in power and efficiency at shorter lengths, encroaching into rack-to-rack cabling below 5 meters. These developments would likely leave AECs viable at only 0.5- to 2-meter cable lengths, ~60% smaller market size than the current market at 100Gbps/lane estimated by Marvell. Ultimately, copper will only be used inside racks, with passive being the primary (cheaper) choice, and optical for everything else.

Power Comparison of AOCs versus AECs

Credo’s AECs boast a 50% reduction in power versus a competing AOC. The chart below shows that at speeds below 100Gbps/lane, this is largely true. However, at 100Gbps/lane, the difference is compressed:

At 100Gbps/lane, Credo’s AEC power consumption increases from ~5W/end to ~10W/end. At 400G, both 50Gbps/lane and 100Gbps/lane are used, so an average was taken. At 800G, we can see the spread between the two is only 15%-20%. This takes both Credo’s and Marvell’s latest spec sheets for AECs and AOCs into account. AOCs previously were consuming 15+W/end at 3- to 5-meter lengths, and the latest specs for 800G designs consume ~12W/end. A few competitors even claim that 800G AOCs at 3 meters could get down to ~10W/end of power consumption. (This analysis also uses Credo’s own power dissipation/consumption figures, which may be higher than those of competitors given that experts claim Credo’s SerDes designs are inferior.)

These results are due to the fact that copper as a transmission medium becomes less efficient when you’re trying to push more bits through it. That means you need more power to send the data through, and also to adjust for interference with chips at either end.

Evident in this analysis too is that optical cables have not reached a speed limit, and their advantages over competing media grow with each new network generation. There is a high probability that as we get beyond 800G into 1.6T, copper as a cable medium may be reserved only for use cases at the rack level. With new network generations, we anticipate the power relationship between AECs and AOCs will be flipped the other way around.

A TCO analysis at 400G (50Gbps/lane) shows a 50% savings from using AECs compared to AOCs. However, with our estimates on the power and cost at 800G, the TCO of an AEC ends up representing a savings of less than 30% compared to that of an AOC. (This result assumes a tighter cost and power differential at 800G, based on estimates we’ve sourced from multiple cable suppliers and industry experts.) This is still a cost savings to AOCs, which explains the market opening AECs currently have.

Bear in mind that the above analysis doesn’t factor in the decline in failure rates of optical cables over the years, although they’re still well above AEC failure rates. Co-packaged optics promise to drastically reduce module failure rates even further. If that occurs, optical cable adoption at the shortest reaches should accelerate meaningfully.

Cooling Comparison of DACs versus AECs

Passive DACs consume ~0.1W/end of power even at faster data speeds per lane. As speeds increase, the thickness of the cable increases to compensate. (Think of this like widening a freeway by several lanes to accommodate the greater traffic flow associated with a higher speed limit.) While the thicker cable can block airflow, the cable itself generates minimal heat, and more material in the wire means lower electrical resistance. Conversely, a thinner cable has more electrical resistance than a thicker one and can therefore generate more heat if power dissipated is equal.

All that means an AEC at ~7mm of thickness has more electrical resistivity than a DAC at 12mm. In order to push more cars down its freeway, an AEC doesn’t add lanes; it just forces more cars down a narrower path and uses power-consuming chips at either end to put any jumbled cars back in proper order.

As detailed above, Credo’s AECs below 50Gbps/lane dissipate 3.5-4.5W/end, but at 100Gbps/lane consume ~10W/end of power. A thinner wire with more power pushing on either end creates more heat than a thicker DAC.

The electrical resistivity and power dissipation differences between AECs and DACs are vast but offsetting at 50Gbps/lane, so both copper cables at a 3-meter length have similar temperature rises (Celsius above ambient). At 100Gbps/lane, an AEC generates a temperature rise of ~10 degrees Celsius above ambient, almost twice that of a passive DAC. In fact, AECs are only competitive on heat generation at 100G/200G, shown below.

While thinner cables create greater airflow to servers, AECs appear to require more cooling for data centers above 200G, not less. This partly explains the resilience of passive copper interconnects within the server racks even as data speeds increase. Additionally, DACs are cheaper than AECs by 20%-50% across generations.

Further undermining the view that AECs will be used in AI workloads, there is an accelerating trend to liquid cool these specialized server racks as single chips surpass 1kW of power consumption. Air cooling is reaching its limits for these high-performance chips, which operate better at lower temperatures.

As such, the tradeoff between DACs and AECs for better air flow appears marginal at 200G/400G for the regular data center connections, and moot for AI use cases (800G and beyond).

SerDes IP

SerDes is the key to efficient data transmission and power consumption, and industry experts have noted that Credo started to fall behind competitors in SerDes several years ago. That’s why Credo pushes its N-1 strategy, and it’s also one of several reasons why the company expects to lose share in the AEC market. To give Credo credit, SerDes scaling is slowing and becoming more difficult, and an N-1 strategy is more viable than before. Nevertheless, it’s still lagging.

Many companies have internal SerDes teams/IP, but only a few independently and actively sell IP: Alphawave and Synopsys. Alphawave Semi is the leader in high-speed SerDes, with IP sales in excess of CRDO’s. Additionally, it’s targeting Credo’s customers on IP and now PAM4 optical DSPs.

Optical DSPs

Credo’s optical PAM4 DSPs have commendable market share at 400G and 800G, competing with offerings from Marvell, Broadcom, and Alphawave. However, DSPs are a smaller market within networking, and these competitors sell comprehensive solutions that include all the components of an interconnect. As with IP, small building blocks of critical components are difficult to scale as a business and are not highly valued by strategic acquirers. As such, the company’s admirable optical DSP market position does not translate into a sustainable amount of revenue given the company’s current structure or is of strategic value to a potential acquirer.

Additionally, PAM4 as a modulation technology is starting to show its limits, with coherent modulation a likely replacement when networks get to the 1.6T generation. PAM6 and PAM8 run into several scaling issues. In keeping with its N-1 strategy, Credo is noticeably behind here as well.

Finally, linear pluggable optics (LPO), i.e. optical communication without a DSP, are expected to take a growing position in the data center market. Nvidia is pushing this solution as an LPO cable at 3-5 meters generates half the power of an equivalent AOC at 800G. At present, LPOs have interoperability issues, performing best with hardware optimized for them. (That’s another way Nvidia locks in its competitive advantage.) Growth in LPO sales would negatively impact CRDO’s sales of not just DSPs, but AECs as well.

Market Trends: Strong data center spending amplified by AI, 2x higher network speeds every 2-3 years

Data Center Spending and Cables

Capital expenditures on data centers by hyperscalers were $250-$300 billion in 2023 (over $125 billion from the top four) and were expected to grow by double digits in 2024 and for the foreseeable future, according to Dell’Oro Group. According to New Street Research, AI-related capex was ~$90 billion last year and was expected to grow considerably by 2027 with hundreds of billions in spending. All in all, the data center market (including AI), could reach $1 trillion by the late 2020’s. Most of this spend goes to high-dollar compute hardware.

Cabling comprises somewhere between 5%-10% of data center capex. According to Gartner, the wired connectivity market is expected to grow ~10% per annum to ~$17 billion in 2025. The AEC market is expected to grow from ~$2 billion today to ~$5 billion by 2025, although we’ve pointed out weaknesses of these estimates above. Currently, the DAC market is ~$7 billion and the AOC market is ~$5 billion, according to MSA.

SerDes/Line Cards/Optical DSPs

The connectivity IP market is expected to grow ~14% per annum to almost $2.5 billion by 2030.  

The line card market is expected to grow almost 20% per annum reach ~$250 million by 2025.

The optical transceiver market for Credo’s products is expected to reach ~$1.4 billion by 2025. DSPs account for 20%-40% of a transceiver’s costs, equating to a $275-$550 million market for optical DSPs by 2025.

Networking – 400G/800G and Beyond

Data center networking has exhibited data rate increases of 2x every 2-3 years, as shown below.

Source: Alphawave Semi.

The bulk of the 400G transition in the data center occurred from 2021-2023. Google was the first to move in late 2018, followed by Amazon (which is largely done with 400G deployments), and Microsoft and Meta concluded their 400G deployments last year. 400G hardware volumes are expected to peak this year. 800G deployments should scale later this year and be the dominant switching technology node by late 2025. Should the industry continue to follow this pattern, we would expect the 1.6T transition to occur in 2027 and 2028. This shift is shown in the chart below for optical transceivers, which should be the dominant networking medium for 800G and beyond:

Source: Coherent Corp.

4) Why now?

Credo is a viewed as an AI “winner” with a current valuation assuming near perfect execution and growth, while real competitive threats to its main product gain traction. However, lurking below the surface are plenty of signs suggesting that AI is actually a threat to the company. We believe that as the 800G transition occurs this year, the company’s last talking point on super-high growth should come and go (as did its “ramp” from the 200G/400G transition), causing the stock to reprice to a more accurate, lower valuation. The next two earnings reports are critical if Credo intends to show enough revenue growth to support the consensus estimate of a 4x-5x increase in revenue over the next 2-3 years. We would argue the flat stock price reaction to the latest FY3Q 2024 earnings report (which guided to 90% year-over-year growth from increasing Amazon AEC purchases and a third new hyperscaler customer) was telling.

Without large scale adoption from at least 6 of the 7 largest hyperscalers (over 2x MSFTs ~$60 million current purchasing level), we fail to see how Credo meets consensus estimates for FY 2027. Should AEC sales increase by just tens of millions instead of hundreds, Credo would need its other products and IP (~$95 million LTM) to increase 7x-8x in 2-3 years to meet the FY 2027 consensus.

Furthermore, a company secondary offering in December 2023, now 25% below the current price, and insiders accelerating stock sales above $20/share appear to contradict the company’s assertion on a near-term growth ramp in calendar year 2H 2024. Additionally, interactions between investors and the company at technical conferences show a lamentable heavy-handed control of messaging, along with oversensitivity to optics and the current share price. 

With CRDO’s stock price correlated to that of serious AI players (NVDA, MRVL, AVGO et al), downside risks to the near-term price become twofold at this critical juncture: First, Nvidia, up massively in the last year, could “sneeze” on earnings against harder comps, causing a downdraft for all AI names. Second, the spread between CRDO’s stock price and reality could become too much to bear, causing investors to wake up. This divergence is highlighted below:

• The company’s quarterly revenue decreased from $55 million in early 2023 to $30-$35 million as Microsoft reduced AEC purchases and is now guiding to a ~$60 million, up only 10% year-over-year since the start of the AI boom.
• In stark contrast, NVDA, AVGO, and MRVL have seen AI-based product revenues increase 410%, 500%, and 200% year-over-year, respectively. However, since early May 2023, CRDO’s stock price is up over 180%, comparable to rises in NVDA (up over 200%), AVGO (~110%) and MRVL (~80%).

Over the longer term, we see structural issues with Credo’s business strategies and products introducing existential risk, leaving minimal value for a strategic acquirer. The company’s own history in the AOC market reinforces this contrarian view.

Credo went from selling an optical module in 2017, to selling only optical DSPs within a few years. DSPs are 20%-40% of the value in an optical module. So, if CRDO was selling $100 million in optical modules, being disintermediated to then sell only a component would reduce revenue by 60%-80%, to $20-$40 million. It’s entirely possible the same thing will happen again to Credo in the AEC market today.

At present, Credo generates less than $200 million in revenue and (including excessive stock-based compensation) has generated just 6.5%-7.5% adjusted EBITDA margins for the last two years (with a 10%-15% peak in early 2023). The company trades on ~90x EV/FY2025 EBITDA (~$300 million in revenue, ~12% adjusted EBITDA margins). Even accounting for growth, applying consensus estimates of ~$875 million in revenue with 20% adjusted EBITDA margins by FY 2027, Credo trades at almost 25x EBITDA. Given CRDO’s characteristics, we fail to see why 25x on even this year’s (FY 2025) EBITDA is merited. Mellanox (MLNX), arguably the best comp to CRDO, was superior in every way, and was acquired at ~16x EV/EBITDA in 2019 by Nvidia.

Ultimately, Credo should re-rate closer to ~17.5x EBITDA on a normalized basis, with revenue below consensus equating to ~40% downside. We believe this Base Case is generous on almost every assumption, which is how we approach valuation when we’ve discovered a highly asymmetric downside situation.

Should CRDO see a repeat of its experience in the AOC market with share loss and product disintermediation, the company’s current AEC revenue would be at risk of a 60%-80% drop, normalizing at a structurally lower level and leaving CRDO likely worth ~10x EBITDA, equating to ~80% downside. We view the odds of our Base and Downside Case to be 50% and 40%, respectively.

Additionally, the competitive environment allows for an interesting pair trade: An undervalued name competing and determined to push Credo out of design wins is Alphawave Semi, up only 45% since May 2023 compared to CRDO’s unwarranted ~180%. AWE.LN is trading just slightly above 10x on ’24 EBITDA (8x lower than CRDO), with ~80% of bookings tied to real AI investments, bookings set to increase rapidly, and multiple hard catalysts over the next 12 months culminating in a possible US listing in 2025.

 As to Credo’s structural problems, the points below should illustrate some issues with any long thesis:

1. The company is betting on AECs, copper cables with speed limitations at newer nodes bound by the laws of nature, and possibly extinct outside of the smallest market: Copper is used in data centers because it’s cheap and effective at certain data speeds. As shown above, as data speeds increase, power needs and interference increase, too. While AECs are pitched as a remedy to this, they are still copper. At the next generation (1.6T) we should see copper cables, primarily active ones, lose ground to optical. Beyond that, it seems that AECs or ACCs could be the last copper cable standing. But even then, active copper likely only partially replaces passive DACs, which are still competitive at 0.5 meters.

Another reason to limit copper connections in data centers is the increased need for cybersecurity. Copper cables emit electromagnetic radiation that carries information about the data being transmitted. Tapping devices could be installed or wires spliced to steal sensitive information. Optical cables are impenetrable to these kinds of tampering and are a much more secure way to transmit data.

Credo’s value is therefore contingent on one product, AECs, and is essentially a bet on the direction of actively powered copper cables in data centers, a challenged market over the next 3-5 years with multiple alternatives and low switching costs.

2. Early adopter Microsoft tested AECs at 400G and decided to maintain current levels; other hyperscalers are likely to do the same: While we discussed Credo’s AOC market experience to show that its history may be repeating itself with AECs, current AEC customer behavior is already foreshadowing future sales growth. Microsoft was an early adopter of CRDO’s AECs, and once it started to deploy more NVDA servers with InfiniBand, it essentially stalled on AEC purchases. We’ve shown AECs are most competitive at the 50Gbps/lane (200G/400G), with more headwinds at 100Gbps/lane (400G/800G) than most would believe from a marketing deck. Credo generated ~$100 million in AEC sales in the last year, with ~$60 million from a scaled Microsoft and less than $40 million from Amazon. Should Google (the third AEC customer currently testing) and essentially the entire remainder of the top seven hyperscalers follow suit, total AEC revenue 3-5 years out would likely only hit ~$425 million, far below the ~$875 million expected by the market. We view winning all the hyperscalers as highly unlikely, and other CRDO products like IP, optical DSPs, line cards, and chiplets as insufficient revenue drivers to fill the gap. On top of all that, rival AWE is targeting SerDes IP and PAM4 Optical DSPs markets aggressively.
3. The cable product revenue base exhibits poor visibility and is subject to inventory cycles: When it comes to data center buildout, cables are not a high priority. Expensive, complicated servers and architectures are designed, and then power and cooling needs are determined. Connectivity in data centers is purposefully designed to have redundancy and to facilitate ease of removal and replacement. Purchases of cables are usually the last thing done before switching on a data center, and the cables themselves may be sourced from many vendors. Businesses built on just cables have poor visibility into their customers’ plans beyond a 6-month horizon. Cables are virtually commodities, with low switching costs courtesy of the evolution in pluggable designs. The only cable products with strong customer visibility are those on the leading edge that have unique technologies or attributes. Additionally, the ease of storage also translates to inventory cycles if overbought by a customer. These dynamics were evident when Microsoft paused its AEC purchases: Credo was blindsided and had to issue an off-cycle 8-K citing elevated inventory (not the only reason as we’ve shown). Any long-term forecasts by the company are likely highly variable and uncertain. Notably, management hasn’t provided clear guidance on long-term revenue targets in the company’s presentations or earnings releases. The same goes for long-term operating margin targets, not printed in releases or presentations, although articulated to investors in conversations.
4. Surprising level of Chinese revenue exposure given current geopolitical tensions, though that might explain CRDO’s Cayman Islands registration: Credo generates over 50% of its revenue from mainland China and Hong Kong, where it has two offices. That revenue may not represent the true end-market geographies, but with AEC packaging and assembly done in China, a western hyperscaler using Credo as a supplier may be exposed to more geopolitical risk than previously anticipated. Additionally, given that Credo is a predominantly US company, the Cayman Island registration for its holding company is a bit odd, though with Chinese subsidiaries, revenues, and employees, it may make sense.
5. New designs may phase out some top-of-rack switches, reducing overall cable intensity in data centers: Linear extrapolation of market trends in cables does contain a potential flaw: the trend of increasing I/O port density (radix) in switch ASICs. Higher-radix switches could eliminate a layer of top-of-rack switches, lowering the overall number of switches in a cloud network. By extension, this would lower the number of cables, improving latency and power consumption but undermining the impact of increased data center buildouts on cable demand.
6. Marginal strategic value to an acquirer, given its N-1 strategy: A “last hope” for CRDO longs might be to have the company get bought out by a strategic acquirer. Marvell’s acquisition of Inphi (IPHI) for ~60x EBITDA in 2020 comes to mind. But that high multiple was due to IPHI having an almost insurmountable lead in optical modules. In fact, some would argue most of MRVL’s value today comes from IPHI. By contrast, Credo’s N-1 strategy means it does not have a lead, and there are multiple competitors with similar or even superior products. As a result, Credo doesn’t have much strategic value to a potential acquirer, and it would be unreasonable to apply elevated “M&A multiples” to determine a normalized valuation for the business as a going concern.

5) Risks/Thesis Pressure Points

Competition Risk. The company’s main product is at risk of losing share and the sale of an entire physical product to bigger and better competitors, which hyperscaler customers already prefer. Additionally, we have already seen a similar situation occur in the company’s history, with AOCs. However, there may be elements of AECs and the market that make Credo’s position more defensible than previously believed, allowing for sales increases of 4-5x as the market has priced in. We consider this risk to be minor, as the company has already admitted it expects to lose share.

New Market Risk. Credo’s engineers did an admirable job of developing the AEC. This allowed them room to pivot when they were pushed out and disintermediated in the AOC market. While they could do this again, we see no signs of another market-making product, and costs to develop even trailing-edge devices are increasingly high.

Technology Risk. There is a possibility the company could develop add-on products that differentiate its AECs, similar to Astera Labs’ smart functions, prolonging or even accelerating sales.  Additionally, the dynamics between copper and optical may evolve differently, keeping AECs competitive against alternatives.

Strategic Buyout Risk. The semiconductor industry has been consolidating for years, and small players typically get bought out. Inphi, acquired for ~60x EBITDA, is a company several newer firms like Credo have tried to get investors to compare themselves to. Should CRDO carve out a sustainable niche in AECs, it could be worth a considerable amount to a larger player like AVGO or MRVL whose strategic advantage is to provide comprehensive solutions to hyperscalers. We think this risk is mitigated, however, because Credo’s N-1 strategy means competitors have had similar or better products in market for years by the time Credo’s new products/IP are launched.

6) Valuation

Credo essentially has one material growth driver, AEC cables, with the remaining product portfolio in small markets. It also has minimal competitive advantages with its N-1 strategy. In almost any reasonable scenario, there is significant downside risk for the equity, in the range of 40%-80%. While we assume various levels of EBITDA margins in each scenario, we keep gross margins in the range of company guidance (63%-65%) even though increasing competition is likely to put margins under pressure.

The only way to craft scenario with upside for the equity is to assume that the company grows to match Mellanox prior to its purchase by Nvidia (~$1.5 billion in revenue) but assume almost 10% higher normalized EBITDA margins (~35%) and apply an EV/EBITDA multiple ~40% higher (~22.5x).

The following are our Base Case assumptions for Credo:

1. Short-term guidance:
   1. FY4Q 2024 revenue: $59-$62 million (64%-66% GM), Non-GAAP opex: $33-$35 million
   2. Non-GAAP EBITDA margins: 7.5%-9.5%.
2. Long-term guidance:
   1. No revenue guidance given. IP should be 10% of revenue structurally.
   2. Gross margins: 63%-65%. Products: 50%-55%, IP: 100%
   3. No operating margin guidance given in presentations/transcripts. Margins only mentioned in talks with investors. Non-GAAP operating margins of 30%-35%. (SBC assumed to be 10%-15% of revenue)
3. Revenue should grow from ~$195M in FY 2024 to ~$650M in FY 2027 and ~$1.1 B in FY 2030. Our FY 2027 estimate is ~25% below consensus of $875 million.
4. Adjusted EBITDA margins should trend upwards from ~6% in FY 2024 to ~22.5% by FY 2027, and ~30% by FY 2029. Our adjusted EBITDA margins are ~10% above consensus for FY 2027 (20%) and in the range of the company’s stated margin goal by FY 2029.
5. We assume SBC declines from ~15% of sales to ~10% by FY 2025.
6. Capex: ~14% historically, ~12% of revenue in FY 2024-2027, ~10% of revenue thereafter.
7. ~17.5x normalized EV/EBITDA (or ~35x EV/FCF) multiple. Our reasoning behind these multiples is in the Peer Analysis section below.
8. Discount rate of ~15% (small-cap).

Five-Year Operating Model

We use a simple five-year operating model to determine value:

Base Case: Our Base Case assumes the company grows to $1.1B in revenue by FY 2030, essentially being the AEC provider of choice for the 7 top hyperscalers ($100M+ in AEC revenue per hyperscaler), with adjusted EBITDA margins of ~30%, on a 17.5x EV/EBITDA valuation (~35x EV/FCF). These assumptions yield approximately 40% downside, or a negative ~10% IRR over five years.

• Base Case Valuation: $14.25/share.

If using GAAP EBITDA by eliminating the SBC add-back at 10% of revenue, CRDO would be worth $10.75/share in this Base Case, ~55% downside.

Downside Case: The Downside Case assumes the company gets pushed out of selling AECs, instead selling AEC components that generate 40% less revenue, even assuming the 7 top hyperscalers as customers and generating ~$425M by FY 2027 (50% below consensus) and $750M by FY 2030, with 20% EBITDA margins and a 10x EV/EBITDA multiple to reflect the inherent weakness in the business. The result is downside of ~80%, or a negative ~25% IRR over five years.

• Downside Case Valuation: $5/share.

Upside Case: Our Upside Case assumes the company grows to $875M in FY 2027 (consensus) and then ~$1.5B in revenue, with adjusted EBITDA margins of ~35% on 22.5x EV/EBITDA valuation (~40x EV/FCF). This yields upside of only ~8%.

• Upside Case Valuation: $24.75/share.

This valuation scenario is far beyond reasonable, as it assumes revenue at maturity in line with Mellanox circa 2020, a highly unlikely outcome for CRDO (which has essentially one cable, versus a portfolio of leading InfiniBand/Ethernet products). On top of that, we apply a ~40% higher EBITDA multiple than MLNX’s ~16x at the time of its acquisition. This was done just to have a scenario that contains any upside at all to the current price.

Peer Analysis
Credo has few direct public comparables, as other semiconductor connectivity and IP firms don’t have a single product comprising over 55% of revenue like Credo’s AEC. We believe the best comparable was Mellanox, but we show other datacom cabling companies like COHR and LITE in the analysis as we expect Credo to look more like a cable supplier if its AEC is the growth driver.

Alphawave Semi is a competitor, but not a great comparable because it now has custom silicon solutions and is truly an AI beneficiary. We also view AWE as a materially mispriced long at only 10x-15x EBITDA on the LSE.

While Inphi (IPHI) is a popular comparison for Credo, it shouldn’t be. Outside of its AEC, Credo is unlikely to generate $1 billion in revenue from its line cards, IP, and optical DSPs. Moreover, IPHI was leading-edge in PAM4 and coherent optical modules, DSPs, TIAs, and drivers, while Credo has a cable product with the rest of the business on an N-1 strategy. In PAM4 optical DSPs, Credo has decent market share next to Marvell, and yet likely generates less than $50 million in revenue from this venture and is expected to lose out with better competitors entering the space. By the transitive property, Marvell is also not a good comparable, as it has Inphi and a custom silicon business, not to mention real AI-derived revenue.

Lastly, we did not include Astera Labs even though it has the Taurus AEC that is likely going to take market share from Credo. The firm just went public with a small float and an eye-popping valuation, and its main growth driver should be its Aries PCIe retimers in AI accelerator cards. ALAB is more closely tied to NVDA, which throws it some business to keep Broadcom from grabbing more share. So Astera is truly tied to AI spending, with an implicit NVDA blessing.

Trading Comparables

EV/EBITDA

1. Coherent (COHR): ~12.5x normalized EV/EBITDA, currently ~15x.
2. Lumentum Holdings (LITE): ~15x normalized EV/EBITDA, currently 17.5x.
3. Mellanox (MLNX): ~13.5x normalized EV/EBITDA, acquired at ~16x.

Coherent is a vertically integrated materials and optics company that makes lasers used in data center and telecom optical cables. COHR is most likely to benefit from the shift to greater optical content in 800G/1.6T networks, and its data center revenue is trending upward.

Lumentum is a provider of optical products, transceivers, and other hardware for data center applications, with a much smaller exposure to industrial markets. Like COHR, greater optical content in connectivity is expected to drive considerable growth for the company. This explains a slightly higher normalized valuation for LITE.

Mellanox was a provider of InfiniBand and Ethernet switches and cables for high performance computing, including copper and optical cables. The company had dominant market share for both protocols and was purchased by Nvidia in 2019 after a competitive bidding process between multiple industry leaders.

The normalized EV/EBITDA multiple of these peers comes in at ~13.5x and ranged from 7x to almost 20x from 2010 to 2024. Current valuations for the optical cable companies appear to embed AI enthusiasm, which is deserved as the 800G transition begins.

EBITDA Margins

1. Coherent (COHR): ~25% normalized non-GAAP EBITDA margins, currently 20%.
2. Lumentum Holdings (LITE): ~20% normalized non-GAAP EBITDA margins, currently 15%.
3. Mellanox (MLNX): ~32.5% normalized non-GAAP EBITDA margins.

Coherent’s margins are ~25% and are expected to stay rangebound. Its gross margins are ~35%. SBC averages ~2% of sales. COHR’s capital expenditures should average ~8% of revenue.

Lumentum’s margins are ~20%, with higher R&D than COHR. Notably, its gross margins are also ~35%. SBC averages 6%-7% of sales. LITE’s capital expenditures should average 7.5%-10% of revenue.

Mellanox’s margins scaled to ~32.5% by 2019. Notably, its gross margins were 65%-70% as it maintained its leadership status across multiple products. At scale, R&D was ~30% of sales and SG&A was ~20%. Stock-based compensation was ~8.5%. MLNX’s capital expenditures averaged ~5% of revenue.

Notably, in our Upside Case we give Credo superior EBITDA margins to MLNX at 35% and in our Base Case we assume ~30%. In both cases, these assumptions may be overly generous as we keep gross margins above 60% despite competition that sells products closer to 35%-45% gross margins.

Valuation Multiple Conclusion

For Credo, we assume a ~17.5x EV/EBITDA normalized valuation, a 30% premium to the group we analyzed above. The multiple is ~10% above Mellanox’s acquired valuation, which is generous given that CRDO is lagging in all other product lines and its leading AEC product is coming under multiple angles of attack from competitors.

Also, a multiple based on Mellanox pre-2023 excludes any AI benefits and would be acceptable for CRDO as they are not materially exposed to the trend. The company is exposed to the trend of faster connections in the data center outside of AI but are under threat. Notably, if Mellanox were still publicly traded today it would be valued higher than ~16x EV/EBITDA.

The other optical cable companies are lower-margin businesses (possibly foreshadowing AEC’s future gross margins) but are well-positioned for changing data center trends. That’s why they have higher current multiples than the normalized estimate we compared Credo to, as we attempted to make assessments without any AI multiple expansion implied in our valuation.

Given Credo’s N-1 strategy, we did not present an acquirer or a precedent transaction analysis.

7) Market Expectations/Perceptions

Credo is covered by 12 analysts. Their ratings include 10 Buys, 1 Hold, and 1 Sell, with an average price target of ~$26/share Most appear to be excited about recent Amazon AEC purchases, Google’s current testing, and strong optical DSPs sales. However, none of these analysts appears to have noticed the power increases at 100Gbps/lane. They continue to state that AECs are half the power consumption of AOCs. Consensus forecasts average to a 4x-5x revenue increase by FY 2027 with 20% EBITDA margins.

The sell side, collectively, is confident in the 2H ramp for AECs on the 800G transition. In fact, some analysts incorrectly say the 400G ramp is coming later this year, which is off by a wide margin. According to most industry experts, 400G deployment really occurred in 2021-2023 and is trailing off this year as 800G starts up aggressively.

Though if one were to read these same analysts’ initiation reports, they repeated the same thesis for the 200G/400G transition. Recent reports show analysts believe Credo’s AECs will comprise ~55% of the copper cabling in data centers within a few years, accelerated by AI and 100Gbps/lane speeds. There appears to be no mention of Credo’s previous PHY optical devices from the 2010s, or the strength of the competition that’s entering the market. Lastly, the sell side has hardly mentioned the dynamic between AI spending and networking dollars going to NVDAs InfiniBand.

8) Downside Protection – Where’s the Margin of Safety?

The company’s downside is considerable. Expectations are high, with large increases in revenue supposedly coming, and the equity reflects this. Credo’s growth engine is based on one product, AECs, with other products lacking the share and TAM to fill the gap in expectations if AECs fall short of plan. Its N-1 strategy puts the business in a potentially hazardous situation, being only a secondary supplier. And with multiple substitutes in the market at the trailing edge, Credo has no unique assets to garner a high valuation in a buyout.

Adjusting growth and margins to consensus estimates in our model implies a 40x EV/EBITDA terminal valuation for the business, based on the current market price of ~$23/share. This is at least 2x what would be close to reasonable, reflecting both poor earnings capabilities and elevated market expectations. Even Marvell, a winner in AECs, AI investment, and the data center connectivity market in general, has been trading at ~30x.

Credo appears to not discount any fears about the business and is clearly being given an AI winner valuation.

9) Conclusion

Credo is priced to perfection as the market gives it undeserved credit for being an AI beneficiary. Meanwhile, its leading-edge product, the AEC, is facing various competitive threats, and its remaining business is lagging behind those of peers.  Some investors observed these dynamics and the high valuation but held off going short just in case the company sees a sizable AI-driven increase in revenue. One year into the AI boom, with revenue up just ~10% since early 2023, we don’t believe AI “manna from heaven” is coming to CRDO in force, and further evidence to support our thesis should come out in the next few earnings reports.

As we’ve shown, AI developments in this first wave of investment are Nvidia-dominant with InfiniBand the network interconnect of choice, leaving out CRDO’s AEC. As a reminder, the company’s revenue went down through part of 2023, as Microsoft deployed more NVDA GPUs. This highlights the threat, not benefit, AI could be to Credo.

Over the longer term, the advantages to AECs diminish above 50Gbps/lane (400G), and data centers appear to be inclined to cut short the useful life of 400G hardware, skipping to 800G/1.6T more quickly than in prior cycles. And AI trends in power, cabling, cooling demands and methods, all point to headwinds for Credo’s AECs.

Furthermore, competition and industry maturation in the AEC market could take place in the same way they did in the AOC market years ago, resulting in a potentially large loss for Credo.

As of April 11, 2024, Credo is trading at ~$23.02/share. With a generous Base Case valuation of $14.25/share, we believe there is ~40% downside to the equity. Should the company be disintermediated and lose share in the AEC market as it did in the AOC market, the equity could be worth $5/share, equating to ~80% downside.

Our expected probability distribution of the cases is: 50% Base, 40% Bear, and 10% Upside, resulting in a blended target price of ~$11.50/share, ~50% downside.

Catalyst

• Revenue increases less than expected, or even declines
• Customer and market share losses
• Product disintermediation
• Insider sales