# Catching Both Copper and Optics: ALAB and CRDO ### Two Bets on the AI Connectivity Transition Astera Labs ($ALAB) and Credo Technology ($CRDO) trade as a pair. Both are chip companies that solve the connectivity bottleneck inside AI data centers, and both get lumped under the single word “connectivity,” so their stocks ride the same thematic swings together. When one rises, the other rises. When one drops, they drop together. But open up a rack and the two sit in entirely different places. ALAB binds tightly clustered GPUs into a single block. CRDO links servers and racks that sit far apart. They do different jobs, they win customers in different ways, and the structure of how they make money is different. So the question becomes this: the market’s habit of grouping them under one theme and assigning them similar multiples is right up to a point, but where exactly does it break down? This article unpacks what each company actually builds, rack by rack. It then walks through, step by step, how their positions get rearranged at the transition where copper ends and optics arrive. Along the way it weaves in what’s being heard on the ground, and it uses public disclosures to calculate, on the same yardstick, how much revenue a single accelerator generates for each company. It then ties this to both companies’ public results to determine how far revenue and margins can run, and to separate the scenario the market has already priced in from the variable it has not. If you already hold ALAB or CRDO, or you’re weighing whether to, this piece should give you a solid analytical foundation for putting the two on the same table and deciding where, and with what view, to place your bet. * * * **Contents** 1. What actually happens inside an AI rack 2. ALAB, from a retimer company to a fabric platform company 3. CRDO, the company that built $1.3 billion by putting chips in cables 4. Who benefits when optics arrive (Pluggable, NPO, CPO) 5. The revenue one accelerator generates: putting both companies on the same yardstick with CDA 6. It isn’t a two-way fight: the competitive landscape 7. What the market has already priced in, and what it hasn’t 8. So how do you invest? * * * ## 1. What actually happens inside an AI rack First, fix one picture in your head. Everything that follows sits on the same set of coordinates only if you do. A large language model can’t fit on a single GPU. The model’s size far exceeds the memory capacity attached to one GPU, so the model is split across tens to hundreds of GPUs, each computing its own share. But whether training or inference, throughout the entire job these GPUs have to keep exchanging intermediate results with each other. What one side computes has to reach the other before the next step can begin. The key point here is that if this exchange stalls, the GPU just sits idle. You’ve laid down hundreds of expensive GPUs, and if half of them are stuck waiting for data, that’s money burning. SemiAnalysis’s Dylan Patel described the interconnect as the place where GPU utilization goes to die. Connection is performance, and connection is cost. The way GPUs connect splits broadly into two. One is the very fast, very frequent communication between nearby GPUs. This communication, happening inside the same server or the same rack, is called scale-up. The other links servers or racks that sit far apart, and this is called scale-out. Scale-up covers short distances but runs at overwhelmingly higher speed and frequency. Scale-out is the opposite. Right now the center of gravity in AI infrastructure spending is tilting fast toward scale-up, because the larger the model gets, the more GPUs have to be bound into a single block. The parts that actually implement this connection come in three kinds. First, a signal weakens as it travels far, and there’s a chip that cleanly revives it midway. This is the retimer. Second, there’s the path that actually carries the signal, the cable. It could be copper or it could be optical. Third, there’s the switch that directs traffic from many GPUs and routes it to its destination. ALAB and CRDO stand in different places on this map. ALAB is lodged narrowly and deeply in the switch position, the brain that directs traffic. CRDO sits in the position rooted in copper, the side that revives signals and embeds chips into cables. This difference splits everything about the two companies. Of course, hard-walling the two into “switch” and “cable” is only a starting point. In reality, ALAB starts from retimers and is trying to climb up into switching, memory connectivity, and optical integration as a platform company. CRDO starts from the AEC copper-link system and is widening into optical DSP, silicon photonics, and retimers. Their starting points differ, and the directions they expand in differ. That difference in origin and expansion path forms the backbone of everything that follows. * * * ## 2. ALAB, from a retimer company to a fabric platform company ALAB was founded in 2017. The start was simple. There was a problem where signals of the PCIe standard couldn’t travel far on a board, and the company made and sold a retimer chip (product name Aries) that regenerated the signal midway. As this chip went into hyperscalers’ AI servers almost like a standard, the company grew fast. Today ALAB’s product lineup splits into four branches. The Aries retimer that revives signals, Taurus that sends Ethernet signals far over cable, Leo that expands memory, and the Scorpio switch that directs traffic. The core is one thing. The company’s center of gravity is shifting from simple components toward designing the nervous system of the entire rack. In the first quarter of 2026, PCIe Gen 6 revenue already exceeded a third of the total, and the company expects Scorpio to become its largest product line by year-end. What binds these four products into one is software called COSMOS. It looks at, in real time, where a signal is weakening, which cable has come loose, and which link has developed a problem, and diagnoses it. This software is the core of why ALAB gets called a connectivity platform company rather than a simple chip company. It doesn’t just sell chips. It sells the chips and the software that operates them as one bundle. So why is the newly released Scorpio X-Series an event? A single switch supports 320 lanes and processes 20 terabits per second. The previous generation was 144 lanes and 9 terabits, so it more than doubled in a single generation. More important is that a single switch can bind up to 80 accelerators. Before, 32 was the limit. Once the GPUs to be bound exceed 64, the old approach required stacking multiple switches. Every time data passes through a switch it has to transfer once more, and each transfer piles on latency. Scorpio binds up to 80 directly with a single switch, so these transfers disappear. It means the same-scale cluster can be connected with fewer switches over shorter paths. On top of this, Scorpio moves beyond the switch’s old role of merely directing traffic. Two engines go in, Hypercast and In-Network Compute. Put simply, the switch directly handles part of the repetitive work where GPUs combine and split data. The grunt work the GPU originally had to do gets absorbed by the switch, so the GPU can focus on its real job, computation. This matters especially for the increasingly popular MoE-style models, which constantly scatter tokens across hundreds of expert modules and gather them back. For these, communication efficiency translates directly into power efficiency per token. Strategically, the most important thing is that all of this sits on open standards. The scale-up domain has effectively been monopolized by NVIDIA’s NVLink. ALAB, on a PCIe basis and by supporting both NVLink Fusion and UALink at the same time, is putting forward the first serious alternative from outside NVIDIA. The numbers reveal ALAB’s weight class. First-quarter revenue of $308 million with a gross margin in the 76% range. That’s a margin rare even among chip companies. Second-quarter guidance of $355 million to $365 million also blew well past market expectations. * * * ## 3. CRDO, the company that built $1.3 billion by putting chips in cables Credo Technology (CRDO) keeps setting new revenue records every quarter. In FY26 Q3, the company posted $407M in revenue, up 201% year over year. The stock once ran as high as $213, cratered to the $80s, and has since bounced back to around $157. Throughout all of this, the market’s read on the company kept shifting. For a while, Credo was just a copper cable company. After the DustPhotonics acquisition, the narrative started to change, with investors beginning to see it as an optical expansion story. To understand CRDO you have to start with a technology called SerDes. It converts the data coming out of a chip into something that can be transmitted at ultra-high speed over a few thin wires, and CRDO is especially strong at doing this while drawing very little power. Every product in this company ultimately comes from this root. CRDO’s cash cow is a product called AEC, short for active electrical cable. The principle is this. Ordinarily, the faster you push the data rate on a copper cable, the shorter the distance the signal can reach. So CRDO embedded chips that regenerate the signal at both ends of the cable, making it possible to send the same copper farther. Cheaper than optical cable and drawing less power, it has come to dominate the segment that links within a rack or to the rack right next door. The result this simple idea produced is fearsome. CRDO’s fiscal 2026 revenue was $1.3 billion, up 206% year over year. The most recent quarter alone was $437 million, growth of 157% versus a year earlier. In two years revenue grew roughly sixfold. Gross margin in the 68% range is lower than ALAB’s, but for a company selling cable systems it’s a very high level. The next cards CRDO is now playing are three. The optical-based ZeroFlap optics, Active Line Cards that replace data center line cards, and OmniConnect, a memory connectivity solution. The company sees each of these as a new market in the billions of dollars. It came this far on the single leg of AEC, and now it’s building three more legs. And here is where the two companies collide head-on for the first time. CRDO has begun building its PCIe Gen 6 retimer business in earnest. This means entering the retimer market, ALAB’s home turf. The two companies, which until now sat in different places, overlap for the first time at this point. How big management sees the future is worth noting too. CRDO’s board granted the CEO performance-based stock awards that vest only on hitting revenue milestones ranging from $2.5 billion to $7.5 billion. It’s a signal of how high internally they see the company’s ceiling. * * * This is where the two companies stand right now. ALAB is climbing into a platform from the switch position, the brain. CRDO is stretching legs in every direction from the root of copper cable. And as we just saw, the two have started colliding for the first time at the retimer. But the variable that will truly split the fate of these two companies is something else. **It’s the moment copper ends and optics arrive.** How the two companies’ positions get rearranged ahead of that transition, and what that means from an investment standpoint, is what we look at next.