Credo Technology Group Holding Ltd (CRDO) - Deep Dive Research Report

Report Date: May 11, 2026 Listing: Nasdaq (CRDO) Sector: Semiconductors / AI Infrastructure Connectivity

1. What the Company Does

Credo Technology Group makes the chips and cables that keep the inside of AI data centers talking to themselves reliably. That sounds mundane. It is not.

Here is the problem Credo solves. When a hyperscaler builds a training cluster of 100,000 GPUs, those GPUs must exchange data at extraordinary speeds - 400G, 800G, soon 1.6 terabits per second per link. Every GPU connects to a network switch, and every switch connects to other switches. The cables connecting them have historically been either dumb copper wires (direct attach cables, or DACs) or expensive fiber optic assemblies (active optical cables, AOCs). DACs are cheap but get unreliable above about 100G per lane and fail to reach beyond a meter or two without errors. AOCs push the signal optically, but consume twice the power, cost more, and carry their own reliability problems (particularly laser failures that cause "link flaps" - brief, silent connectivity interruptions that force the entire AI job to restart from a checkpoint, wasting hours or days of compute time).

Credo invented a third option: the Active Electrical Cable, or AEC. An AEC looks like a copper cable - thinner, lighter, and more flexible than a DAC - but embedded in each end of the connector housing is a small integrated circuit containing a SerDes (Serializer/Deserializer), a retimer that regenerates the electrical signal, and Forward Error Correction logic. The result is a cable that carries data reliably up to 7 meters at 800G, weighs 75% less than an equivalent DAC, consumes 50% less power than an AOC, and - critically - does not have a laser that can fail. Credo's ZeroFlap branding refers to the zero "link flaps" that their cables produce, which is the single metric hyperscaler network reliability engineers care about most when training large AI models.

"Maximizing network reliability and energy efficiency have been our core mandates." - CEO Bill Brennan, Q3 FY2026 concall, March 2, 2026

The business started further upstream. Credo was founded in 2008 by Bill Brennan (CEO), Lawrence Cheng (CTO, known as Cheng Chi Fung in filings), and Job Lam (COO) - all semiconductor veterans trained at Marvell and Sun Microsystems. For its first decade, Credo operated nearly invisibly, perfecting its SerDes technology and licensing it to other chipmakers and to hyperscalers building custom ASICs. This IP licensing business never made Credo rich, but it gave the company world-class signal integrity expertise and deep relationships with the buyers who would eventually become its best product customers.

The pivot came as AI workloads exploded and network demands scaled beyond what passive copper could handle. Credo took its SerDes IP and wrapped it into a complete system - the AEC. Critically, Credo controls the full stack: the SerDes IP, the IC design, the cable architecture, and the manufacturing qualification through BizLink (a dedicated AEC cable manufacturing partner). When a hyperscaler buys a Credo AEC, they are buying the output of a 15-year investment in signal integrity science that is extremely difficult to replicate quickly. Credo went public on Nasdaq in January 2022.

The company is fabless - it designs its chips and contracts manufacturing to foundries, using what it calls an "N-minus-one" strategy (deliberately staying one process node behind the bleeding edge, at 12nm, rather than cramming onto 5nm or 3nm). This keeps it out of the foundry capacity wars while maintaining sufficient performance for its use cases.

2. Business Segments

Credo reports in a single reportable segment but has several distinct product lines that operate with different competitive dynamics and growth trajectories.

2.1 Active Electrical Cables (AECs)

This is Credo's largest revenue driver and the product that defines the company to the outside world. AECs are copper cables with embedded chips, manufactured in partnership with BizLink (the sole AEC manufacturing partner), and sold to hyperscalers for connecting GPUs to switches inside data centers.

The core capability here is the SerDes and signal processing IP embedded in the cable connector. What Credo knows how to do - that took 15 years to build and would be hard to replicate quickly - is design a SerDes that runs at 112Gbps per lane with extremely low power consumption, mate it with a retimer and FEC engine in a package small enough to fit inside a standard QSFP-DD or OSFP connector housing, and qualify that assembly for reliable operation in the harsh thermal environment inside a working rack.

Credo invented this product category and claims 88% market share per 650 Group data referenced on the company website. The ZeroFlap brand name is the marketing manifestation of the key technical advantage: AECs do not produce the brief connectivity losses (link flaps) that optics-based solutions can. In an AI training cluster where a single link flap forces a job restart, this is worth paying for.

The AEC product line covers 400G (56G per lane) and 800G (112G per lane) today, with 1.6T (200G per lane) products expected in production in H1 FY2027. The cables support Ethernet, UALink, and ESUN protocols. By Q3 FY2026, the AEC line had won five hyperscaler customers, with four each representing more than 10% of quarterly revenue at various points.

2.2 IC Solutions - Retimers and Optical DSPs

This segment sells standalone semiconductor chips rather than complete cable assemblies.

Retimers - Retimers are standalone chips that sit on a PCB (inside a switch, server, or other networking device) and regenerate electrical signals that have degraded in transit across the PCB traces or backplane. Credo's PCIe retimer portfolio covers PCIe Gen 5 and the newly completed Gen 6 (its Toucan product was the first PCIe 6.0 to achieve PCI-SIG compliance at 32.0 GT/s). PCIe Gen 6 retimers are designed for AI servers connecting GPUs to host CPUs and memory controllers - they are the plumbing that makes fast interconnects work inside individual servers. Revenue from PCIe retimers is expected in FY2027 following design wins in FY2026.

The more recent and potentially larger retimer opportunity is the Blue Heron - a 200G per lane retimer for scale-up AI fabric networks (the fast interconnects between GPUs within a training pod). Blue Heron supports UALink, ESUN, and Ethernet protocols and was announced in Q3 FY2026 with Upscale AI as the first named customer.

Optical DSPs - Credo's Cardinal family of optical DSPs processes PAM4 signals at up to 224Gbps per lane, enabling 800G and 1.6T optical transceivers. Optical module manufacturers embed Credo's DSP into their modules, which are then sold to hyperscalers for longer-reach connections that copper cannot cover. This business sits in a well-established competitive market against DSPs from Marvell, Broadcom, and others - but Credo's low-power design philosophy gives it differentiation on thermal constraints in dense AI deployments.

2.3 ZeroFlap Optics

This is Credo's newest-in-production segment. Where AECs use copper to connect over short distances and optical DSPs are components embedded in other companies' transceivers, ZeroFlap Optics is Credo's own complete optical transceiver - a pluggable module that generates its own laser light.

The insight is to take Credo's reliability and telemetry expertise and apply it to optics. A standard optical transceiver can produce link flaps if the laser degrades. Credo's ZeroFlap Optics integrates a customized DSP with its PILOT platform (Predictive Integrity Link Optimization and Telemetry) and a switch-level SDK so the system can monitor link health and predict failures before they occur. The current product is an 800G 2xDR4 transceiver.

Production shipments began to TensorWave (an AMD-exclusive AI cloud provider) in Q3 FY2026. By Q3, Credo was in qualification with three additional customers including hyperscalers and neocloud operators. The production ramp was pulled forward from H2 FY2027 to Q1 FY2027 based on clear customer demand. Management expects to ramp ZeroFlap Optics across more than four customers during FY2027 and projects the combined optical portfolio (ZeroFlap Optics + optical DSPs + DustPhotonics silicon photonics) to generate more than $500M in revenue in FY2027.

2.4 Active LED Cables (ALCs)

ALCs are a future product category being developed through a partnership with Hyperloom, an Ottawa-based micro-LED specialist. The concept: use micro-LEDs (which are more reliable and cheaper than traditional laser diodes) as the light source in a thin, lightweight pluggable optical cable. This would give reach up to 30 meters at AEC-level reliability and power - filling the gap between AECs (0.5-7m copper) and full optical transceivers (hundreds of meters to kilometers).

ALCs are targeted for sampling/qualification in FY2027 and production ramp in FY2028. Management estimates the ALC total addressable market will ultimately exceed the AEC TAM by more than double, reflecting the larger fraction of data center connections that span rack-to-row distances.

2.5 OmniConnect Gearboxes

OmniConnect is Credo's bet on disaggregated compute architectures. In current AI accelerator designs, memory is tightly coupled to compute on the same package or very nearby. OmniConnect enables designers to move to commodity DDR memory while achieving up to 30x more capacity and 8x more bandwidth than traditional approaches - allowing a processor's memory to be physically separated from the compute while maintaining extremely high bandwidth.

The underlying technology is Credo's 112G VSR CertiS - a copper interconnect that achieves 1,200 CertiS units per maximum reticle die, creating 120 terabits per second of aggregate bandwidth with 10-inch reach. The first named customer announced plans for an inference-targeted XPU with 2 terabytes of memory capacity. Revenue is targeted for FY2028.

2.6 SerDes IP Licensing and Engineering Services

The smallest segment by revenue (roughly 6% of total in recent fiscal years), this business licenses Credo's core SerDes intellectual property to hyperscalers and chipmakers building their own custom ASICs. A hyperscaler designing a custom networking or AI training chip will license Credo's SerDes IP to embed high-speed connectivity into their device without having to develop the analog expertise themselves.

This business exists as a historical artifact of Credo's first decade but remains strategically valuable: it keeps Credo's technology inside hyperscaler custom chips, creates relationships at the chip architecture level, and generates royalty streams. The business has been deprioritized in recent periods as product revenue has scaled dramatically, but management has not signaled any intention to exit it.

3. Products and Business Detail

Product Catalogue:

• ZeroFlap AEC 400G: 56G per lane, QSFP-DD/QSFP56/QSFP28 form factors, reach 0.5m-7m. For intra-rack GPU-to-switch connectivity at 400G Ethernet speeds. In production with all major hyperscalers.
• ZeroFlap AEC 800G: 112G per lane, OSFP/QSFP-DD/QSFP112 form factors, reach 0.5m-7m. Current high-volume product. Production across five hyperscalers as of Q3 FY2026.
• ZeroFlap AEC 1.6T: 200G per lane. Sampling; mass production targeted for H1 FY2027. Supports Ethernet, UALink, ESUN protocols.
• ZeroFlap AEC 1.6T OSFP-XD: Purpose-built for hyperscaler spine switching (the "fat spine" of the AI cluster network, where each switch port carries maximum bandwidth). Announced ahead of Q3 FY2026.
• Toucan PCIe Gen 6 Retimer: First PCIe 6.0 PCI-SIG-compliant retimer. In design-win stage FY2026; production revenue FY2027.
• Blue Heron 200G Retimer: For scale-up AI fabric; UALink/ESUN/Ethernet; announced Q3 FY2026.
• Cardinal Optical DSP: 224Gbps per lane PAM4. Embedded in third-party optical transceivers. Supports 800G and 1.6T modules. In production.
• ZeroFlap Optics 800G: Complete 2xDR4 optical transceiver. ZeroFlap-branded for reliability. PILOT SDK integration. In production with TensorWave; qualifying with three additional customers as of Q3 FY2026.
• ZeroFlap Optics 1.6T: On roadmap via DustPhotonics silicon photonics PIC integration.
• OmniConnect Weaver: First gearbox product. Die-to-die disaggregated memory connectivity. 112G VSR CertiS. Revenue FY2028.
• PILOT Platform: Software platform for predictive link monitoring, telemetry, and SDK integration into switch management systems. Not sold separately but embedded in ZeroFlap products.
• SerDes IP (OmniConnect 112G VSR): Licensed to hyperscalers for custom ASIC integration.

Manufacturing:

Credo is fabless. Its chips are fabricated at contracted foundries using what management calls an "N-minus-one" node strategy - predominantly 12nm class processes rather than the 5nm or 3nm nodes that receive the most attention and supply pressure. Management explained this in Q2 FY2026: foundry partners recognize that connectivity chips are critical enablers of GPU shipments and prioritize Credo's orders accordingly. The strategy also insulates Credo from the worst foundry capacity constraints.

AEC cable assembly is outsourced exclusively to BizLink, a Taiwanese contract manufacturer. BizLink handles the mechanical assembly of the cable with the Credo-designed chips embedded in the end connectors. This means Credo does not own AEC manufacturing assets, but it also means BizLink is a single point of supply dependency. Management has expressed confidence in BizLink's capacity planning and noted that AEC manufacturing is fundamentally different from wafer foundry production - it does not face the same kinds of capital or lead-time constraints.

DustPhotonics Acquisition:

On April 13, 2026, Credo announced a definitive agreement to acquire DustPhotonics, an Israeli startup specializing in silicon photonics Photonic Integrated Circuit (SiPho PIC) technology. The deal is structured at $750M cash plus approximately $123M in stock upfront, with an earnout of up to 3.21 million additional shares (~$430M) contingent on DustPhotonics hitting revenue milestones. The acquisition is expected to close in Q2 of Credo's FY2027 calendar.

DustPhotonics has a portfolio of SiPho PICs for 400G, 800G, and 1.6T transceivers, with a roadmap to 3.2T. Silicon photonics integrates key optical functions onto a single chip, reducing component complexity, improving manufacturing yields, and enabling lower cost at scale versus traditional approaches using discrete lasers and waveguides. By owning the PIC technology, Credo would control all four layers of the connectivity stack in its optical products: SerDes IP, DSP, PIC, and system integration.

CoMira Acquisition:

Credo also completed the acquisition of CoMira Solutions (closed by Q3 FY2026) for its protocol IP - specifically link layer, error correction, and security semiconductor IP. CoMira's technology strengthens the ZeroFlap AEC, ZF Optics, and OmniConnect products with integrated FEC and security capabilities.

IP and Patents:

Credo filed a patent infringement complaint with the U.S. International Trade Commission (ITC) in March 2025 against Amphenol, Molex, TE Connectivity, and Volex, seeking to block importation of AEC products infringing Credo's patents. Management noted the company had invested "tens of millions of dollars" in AEC IP over seven to eight years. The result: by Q2 FY2026, Credo had entered licensing agreements with three cable companies and had "a couple more in flight." By March 2026, Credo settled with Molex and TE Connectivity under confidential terms. Additional licensing deals were struck with 3M and the Siemon Company.

4. Customers

Who buys: Credo sells almost exclusively to hyperscalers - the world's largest cloud computing and AI infrastructure operators. By Q3 FY2026, five hyperscalers had each purchased material volumes of AEC products. By Q2 FY2026, four customers each represented more than 10% of quarterly revenue simultaneously. Customer names are not publicly disclosed in Credo's filings, but industry context and analyst commentary point to Microsoft, Amazon Web Services, Google, and Meta as the most probable top-four, consistent with these being the four largest AI training cluster operators globally.

TensorWave - an AMD-focused neocloud AI infrastructure operator - was the first named customer for ZeroFlap Optics and a public design partner. TensorWave's involvement indicates Credo is beginning to sell into the emerging neocloud AI-as-a-service tier below the hyperscalers.

Concentration is the dominant financial fact about this business. In Q4 FY2025, one customer represented 61% of revenue. In Q1 FY2026, the top three customers were 35%, 33%, and 20%. In Q2 FY2026, four customers were 42%, 24%, 16%, and 11%. In Q3 FY2026, three customers were 39%, 32%, and 17%. Quarterly mix shifts dramatically as different hyperscalers ramp infrastructure deployments at different rates. The company is genuinely not diversified away from a small number of very large buyers.

Why they buy: Hyperscaler network engineers have a primary mandate that most procurement functions do not: minimize AI training job interruptions caused by link failures. A link flap in a training cluster running a 10,000 GPU job can force a restart that wastes hours of compute at cost. The savings from using cheaper DAC or AOC cables are small compared to the cost of lost training time. Credo's AECs were engineered specifically to address this. The PILOT platform adds active monitoring so network operations teams can predict failures before they disrupt jobs. The combination of hardware reliability and software observability is what generates preference over alternatives.

Switching costs: Once a hyperscaler has qualified a Credo AEC for a specific rack configuration and switch platform, switching to an alternative requires re-qualification - not just of the cable, but of the entire system behavior. Given that hyperscalers run proprietary network fabrics with custom firmware, and Credo's PILOT SDK integrates into switch management systems, the switching cost is real but not insurmountable. What is more decisive is the pace of innovation: Credo's claim is that it delivers new speed grades first, qualifies first, and ramps production first - making the incumbent position sticky through sheer velocity rather than contractual lock-in.

Contract structure: Credo does not have long-term supply contracts in the traditional sense. Hyperscalers issue purchase orders against agreed pricing and qualification parameters. Revenue is not recurring in a SaaS sense. However, deployments are ongoing - a hyperscaler building 100,000 GPU clusters in FY2026 will need AECs for every switch port in that build, and then again when they expand in FY2027. This creates a de facto recurring demand that tracks hyperscaler capex cycles.

Warrant arrangements: At least one hyperscaler (Amazon, per SemiAnalysis reporting on Trainium3) received stock warrants from Credo as part of their commercial relationship - a common structure where the hyperscaler's volume commitment entitles them to equity upside. This arrangement, while dilutive at high stock prices, reflects the strategic importance of winning hyperscaler platform deployments where Credo's product becomes the default standard.

5. Competitive Landscape

The AEC, retimer, and optical DSP markets are not empty. But they are markets where Credo has a meaningful head start and where the barriers to matching its full stack are real.

Marvell Technology is the most formidable competitor. Marvell has far greater resources, a much larger product portfolio, and strong existing relationships with hyperscalers across multiple semiconductor categories. In December 2025, Marvell launched its "Golden Cable" initiative - a validated AEC architecture and firmware platform that enables cable manufacturers (Foxconn Interconnect Technology, Luxshare) to build Marvell-chip-based AECs with accelerated time to market. This is a direct assault on Credo's AEC dominance. The difference is that Credo pioneered the technology, qualified the use case, and has years of deployment data that Marvell is now catching up to. Marvell's strength is broad optical networking solutions (it bought Inphi, Innovium); its AEC push is more ecosystem coordination than pure technology differentiation.

Broadcom is the largest semiconductor company in the networking space and competes in retimers and optical DSPs. Broadcom does not appear to compete directly in AEC cables. Its strength lies in merchant switch silicon (Tomahawk, Trident families), where it has entrenched positions. For Credo's retimer business, Broadcom is a competitor via its retimer product lines, but the PCIe Gen 6 space is early enough that Credo's Toucan certification advantage matters.

Astera Labs is the retimer pure-play most directly comparable to Credo in the PCIe/CXL space. Astera Labs went public in 2024 and focuses on PCIe and CXL connectivity for AI servers. It competes head-on with Credo's retimer business. Astera Labs tends to lead on PCIe retimers while Credo leads on AECs, but both companies are expanding into each other's territory as the product lines mature.

InnoLight, Luxshare, Amphenol, Molex, TE Connectivity, Volex: These are cable and connector manufacturers who attempted to build AEC products in competition with Credo. Three of them are now Credo licensing partners following the ITC patent action - they can manufacture AECs but pay Credo royalties. The ITC action and its resolution converted would-be competitors into revenue contributors, a strategically elegant outcome.

Barriers to entry: Building a high-speed SerDes from scratch requires 5-10 years of analog and mixed-signal circuit design expertise, a library of characterization data for silicon process corners, and hyperscaler qualification relationships that take years to develop. The PILOT software platform adds another layer that competitors would need to replicate. These are real but not absolute barriers - Marvell clearly has the resources to develop and qualify AEC alternatives, which is why Credo's patent portfolio and velocity of innovation matter as second-line defenses.

Where Credo is strong: AEC market creation and dominance, ZeroFlap Optics as a differentiated reliability-first optical transceiver, PCIe Gen 6 first-to-compliance retimer, and the PILOT software integration that is embedded in hyperscaler switch management systems.

Where Credo is exposed: Customer concentration makes any shift in a single customer's purchasing plan a company-level event. Marvell's Golden Cable initiative, if it gains traction, could erode AEC market share over time. The optical transceiver market (where ZeroFlap Optics competes) is structurally more competitive than the AEC market, with InnoLight, Coherent, and Lumentum as established players.

6. Industry

What drives demand: The AI training compute build-out is the dominant demand driver. Every Nvidia H100, H200, or Blackwell GPU deployed in a hyperscaler cluster requires multiple high-speed network connections - at minimum one to a top-of-rack switch, and in scale-up configurations additional high-bandwidth links to neighboring GPUs in the same pod. The density and speed of these connections has increased dramatically: from 4x50G (200G) to 4x100G (400G) to 4x200G (800G) in three years, with 4x400G (1.6T) coming in FY2027. Each speed transition requires new AECs, new retimers, and often new optical transceivers.

Industry size: The AEC market specifically is projected to grow from approximately $644M in 2025 to $1.4B in 2029 per 650 Group estimates cited by Marvell. More aggressive estimates project $3.6B by 2032 at roughly 22-24% CAGR. These figures reflect only AECs; the combined addressable market for Credo's five-pillar portfolio (AECs, IC solutions, ZF Optics, ALCs, OmniConnect) is estimated by the company at more than $10B (up from approximately $3B eighteen months prior).

Supply chain position: Credo sits between the silicon foundries (TSMC/Samsung for chips), cable manufacturers (BizLink for AEC assembly), and hyperscalers. It is a fabless semiconductor designer that creates uniquely embedded cable products. In the optical segment, Credo's products (Cardinal DSPs) are components inside optical modules made by other companies (InnoLight, Coherent, etc.) and then sold to hyperscalers.

Regulatory environment: No specific government approvals are required to sell AEC cables or chips. There are FCC Part 15 and CE compliance requirements for emissions, standard for all networking hardware. Credo's ITC filing against AEC infringers is effectively using the U.S. trade regulatory system as a competitive weapon, which it did successfully. The DustPhotonics acquisition (an Israeli company) may require regulatory approval in multiple jurisdictions, which management flagged as a condition to close.

Tariff exposure: Q3 FY2026 guidance was explicitly built on "current tariff regime, which remains fluid." Credo's chips are fabbed in Taiwan (TSMC), assembled into cables in Taiwan (BizLink), and sold to U.S. hyperscalers. A tariff escalation targeting Taiwan or semiconductor imports would directly affect Credo's cost structure or its customers' willingness to absorb tariff-inflated prices. Management flagged this risk and indicated it would discuss tariff implications more frequently as calendar 2026 progresses.

Cyclicality: Hyperscaler capex is not immune to cycles - periods of AI investment caution (or a significant AI model disappointment) would slow cluster builds. However, the current capex trajectory from Microsoft, Amazon, Google, and Meta is measured in hundreds of billions of dollars over the next several years, providing a demand backdrop that would need to deteriorate dramatically to materially hurt Credo's FY2027 trajectory. The more relevant near-term risk is quarterly lumpiness from individual hyperscalers pausing or accelerating builds.

Co-packaged optics (CPO) risk: The industry has debated for years whether optical components will eventually be co-packaged directly onto switch ASICs, eliminating the need for external cables and transceivers. This would, in theory, make AECs irrelevant. Credo's management position on this in Q2 FY2026: "We do not see customers moving forward in a big way" with CPO currently. Reliability, serviceability, and cost concerns with CPO persist, and hyperscalers have prioritized alternatives. However, CPO adoption by major switch vendors (Broadcom, Marvell) in the 2028-2030 timeframe remains a technology risk that the industry tracks seriously.

7. Growth Triggers

The four concall dates used in this section: Q4 FY2025 (June 2, 2025), Q1 FY2026 (September 3, 2025), Q2 FY2026 (December 1, 2025), Q3 FY2026 (March 2, 2026).

• Fifth hyperscaler AEC ramp beginning. A fifth hyperscaler began contributing initial AEC revenue in Q2 FY2026 and is expected to grow its contribution through FY2027. Management indicated this customer, given its size, has "potential" to reach 10%+ of quarterly revenue in a future period. (Q2 FY2026 concall, Dec 1, 2025; confirmed fifth AEC win noted in Q3 FY2026 concall, Mar 2, 2026)

• ZeroFlap Optics production ramp pulled forward to Q1 FY2027. Originally expected to ramp in H2 FY2027, the ZeroFlap Optics production ramp was accelerated by approximately six months following strong customer pull. Management expects to ship production volumes across more than four customers during FY2027 and projects combined optical revenue greater than $500M in FY2027 (including DustPhotonics).

  "It is very rare that somebody would say...they really do not want" the product. - CEO Bill Brennan, Q3 FY2026, March 2, 2026 (Q3 FY2026 concall, Mar 2, 2026; ZeroFlap Optics first cited as a new pillar in Q2 FY2026)

• 1.6T AEC mass production in H1 FY2027. AECs supporting 200G per lane Ethernet, UALink, and ESUN protocols are in development and slated for mass production in the first half of FY2027. This coincides with hyperscaler transitions to next-generation AI cluster topologies. (Q3 FY2026 concall, Mar 2, 2026)

• PCIe Gen 6 retimer production revenue in FY2027. Design wins for the Toucan PCIe Gen 6 retimer were targeted for FY2026. Production revenue is expected to follow in FY2027. Toucan became the first PCIe 6.0-compliant retimer at 32.0 GT/s. (Q4 FY2025 concall, Jun 2, 2025; confirmed trajectory maintained in Q3 FY2026, Mar 2, 2026)

• Blue Heron 200G per lane retimer for scale-up AI fabrics. Announced in Q3 FY2026 with Upscale AI as the first customer. Supports UALink, ESUN, and Ethernet for intra-pod GPU connectivity (scale-up networks). Revenue timing not yet specified. (Q3 FY2026 concall, Mar 2, 2026 - new trigger)

• DustPhotonics acquisition closes and begins contributing. The $750M+ acquisition of the Israeli silicon photonics PIC developer is expected to close in Q2 FY2027. Once closed, Credo will own the full optical connectivity stack (SerDes + DSP + PIC + system integration) and immediately adds DustPhotonics' existing customer relationships.

  "Greater than $500 million in optical revenue in fiscal 2027." - company guidance with DustPhotonics included. (Announced April 13, 2026)

• Active LED Cables sampling in FY2027, production ramp FY2028. The Hyperloom partnership targets ALCs for 30-meter reach with AEC-level reliability. Management estimates ALC TAM will exceed AEC TAM by more than double at full market maturity. (Q2 FY2026 concall, Dec 1, 2025; Q3 FY2026 concall, Mar 2, 2026)

• OmniConnect first production revenue in FY2028. First gearbox customer announced plans for an inference-targeted XPU with 2TB memory capacity. OmniConnect is the enabling technology for disaggregated AI compute architectures. (Q2 FY2026 concall, Dec 1, 2025; Q3 FY2026 concall, Mar 2, 2026)

• FY2027 guidance: sequential mid-single-digit growth and more than 50% year-over-year growth. This is the first formal multi-quarter FY2027 framework management has communicated publicly. It implies FY2027 revenue roughly 50%+ above FY2026's projected total. (Q3 FY2026 concall, Mar 2, 2026)

• AEC revenue expanding from intra-rack to rack-to-rack applications. Management identified the scale-up network (rack-to-rack, within a training pod) as the highest-volume remaining AEC opportunity not yet captured. (Q1 FY2026 concall, Sep 3, 2025; repeated in Q2 FY2026)

8. Key Risks

1. Customer concentration. The mechanism: Credo's top three customers have consistently represented 70-90%+ of quarterly revenue. If any single hyperscaler pauses its AI cluster buildout - due to internal project reassessment, macroeconomic pressure, or a shift in AI training architecture that reduces AEC density - the revenue impact is severe and immediate. There is no diversified revenue base to absorb the shock. In Q4 FY2025, a single customer was 61% of revenue. This is not being obscured in management guidance - it is disclosed prominently every quarter. The risk is not unknown; it is structural.

2. AEC market disruption: co-packaged optics and technology transitions. The mechanism: if hyperscalers adopt CPO (co-packaged optics, where light is generated directly at the switch chip) at scale, AECs become unnecessary for those connection points. Copper connections would be replaced by on-chip optics. Management explicitly downplays this risk and argues CPO adoption is not imminent. However, Broadcom and others are developing CPO-capable switch ASICs. A successful Broadcom CPO launch with hyperscaler adoption could begin to reduce AEC density requirements starting in the 2028-2030 timeframe. The risk is not near-term but is architecturally real.

3. Marvell's Golden Cable initiative gaining traction. The mechanism: Marvell launched its "Golden Cable" AEC ecosystem in December 2025, enabling cable manufacturers Foxconn Interconnect and Luxshare to build validated AEC products faster and with Marvell chip architecture. If Marvell successfully qualifies its AEC platform at multiple hyperscalers, Credo's claimed 88% market share will compress. Marvell has the resources to sustain a multi-year qualification campaign. The risk is not immediate - hyperscaler qualification takes 12-18+ months - but the initiative has begun.

4. Single-supplier manufacturing dependency. The mechanism: BizLink is the sole AEC cable manufacturer. If BizLink faces a production disruption (fire, labor action, geopolitical event affecting Taiwan, or simply capacity mismanagement), Credo cannot easily switch to an alternative manufacturer without extended qualification cycles. The AEC manufacturing process involves precise chip-in-connector assembly that is not easily replicated elsewhere on short notice.

5. DustPhotonics integration and earnout execution risk. The mechanism: Credo is paying up to $1.3B for an Israeli startup whose technology is not yet in high-volume production. The earnout structure (up to $430M in additional shares) is contingent on DustPhotonics hitting revenue milestones - creating dual risk. If DustPhotonics underperforms, Credo has paid $750M+ cash for technology without the planned revenue contribution. If DustPhotonics outperforms, the earnout dilutes existing shareholders significantly. Acquisitions of this scale, especially those requiring regulatory approval across multiple jurisdictions, carry integration and execution risk that is difficult to model.

6. Tariff and geopolitical risk. The mechanism: Credo's supply chain runs through Taiwan (chip fabrication at TSMC, AEC assembly at BizLink). U.S. tariffs on semiconductor imports from Taiwan, or escalating U.S.-China-Taiwan geopolitical tension, could directly affect Credo's cost structure. Management explicitly flagged that Q3 FY2026 guidance was built on the "current tariff regime, which remains fluid." This is an acknowledged known risk.

7. Operating expense scaling and cash consumption from DustPhotonics. The mechanism: the $750M cash component of the DustPhotonics deal represents a significant portion of Credo's cash. At the end of Q3 FY2026, Credo had $1.3B in cash, meaning the acquisition consumes the majority of that balance. If revenue growth slows or gross margins compress before DustPhotonics begins generating cash, Credo could find itself needing to raise additional capital at an inopportune time.

8. SemiAnalysis critical view on AEC demand (historical). SemiAnalysis published a report titled "Thick Cables, Thin Margins" arguing that Microsoft's AEC adoption was lower volume than sell-side analysts believed and that Microsoft had largely reverted to DAC and multi-mode optics. This was published in approximately 2023. Since then, Credo has reported explosive revenue growth that contradicts the bear case. However, the underlying concern - that AEC is deployed selectively for high-reliability requirements rather than universally - remains relevant when modeling market share at full saturation. Credo's own management does not assert that AECs will replace all copper connections; they target the high-reliability backend AI cluster use case specifically.

9. Walk the Talk

Concall dates used: Q4 FY2025 (June 2, 2025), Q1 FY2026 (September 3, 2025), Q2 FY2026 (December 1, 2025), Q3 FY2026 (March 2, 2026).

The most striking fact about Credo's management credibility is that they have beaten their own guidance materially and repeatedly over the four quarters covered here. Not once but consistently. The pattern is not ambiguous.

Q4 FY2025 to Q1 FY2026:

In June 2025, Credo guided Q1 FY2026 revenue at $185-195M. The midpoint implied about 12% sequential growth from Q4 FY2025's $170M. Management also guided a full-year FY2026 exceeding $800M, representing 85%+ growth over FY2025's $436.8M. These were bold numbers at the time - tripling revenue in two years.

When Q1 FY2026 reported in September 2025, revenue was $223.1M - 31% sequential growth against guidance of 12%, and $28-38M above the guidance range. Management noted that a fourth hyperscaler had generated its first material revenue in the quarter.

Q4 FY2025 guidance: $185-195M. Q1 FY2026 actual: $223.1M. Beat: approximately 17% above midpoint.

Q1 FY2026 to Q2 FY2026:

In September 2025, Credo guided Q2 FY2026 revenue at $230-240M. The actual Q2 FY2026 result was $268M - again substantially above guidance. Non-GAAP net margin of 47.7% was well above the 40% management had indicated as a target.

The Q2 call was also where management introduced three new product pillars (ZeroFlap Optics, ALCs, OmniConnect) and raised the full-year FY2026 revenue growth outlook from 120% year-over-year to more than 170% year-over-year. This was the quarter the company did an ATM equity offering, raising cash for the upcoming DustPhotonics acquisition. The dual signals - raising equity while the stock was high and introducing three new product categories - both appear to have been well-timed and strategically coherent.

Q1 FY2026 guidance for Q2: $230-240M. Q2 FY2026 actual: $268M. Beat: approximately 12% above midpoint.

Q2 FY2026 to Q3 FY2026:

In December 2025, Credo guided Q3 FY2026 at $335-345M. This was itself a huge number - implying 27% sequential growth. The actual Q3 FY2026 result was $407M, which was so far above guidance that Credo pre-released its preliminary revenue figure before the full earnings report. The 52% sequential growth was Credo's largest quarterly percentage jump in this period.

Gross margins improved to 68.6%, above the guided 64-66% range. Non-GAAP net income of $208.8M quadrupled year-over-year.

Q2 FY2026 guidance for Q3: $335-345M. Q3 FY2026 actual: $407M. Beat: approximately 19% above midpoint.

Q3 FY2026 (current guidance):

Credo guided Q4 FY2026 at $425-435M and provided formal FY2027 guidance for the first time: mid-single-digit sequential growth per quarter and more than 50% year-over-year growth for the full fiscal year.

The gross margin guidance for Q4 at 64-66% is notably below Q3's 68.6%, which management attributed to product mix shifts (optical products and early-stage products typically carry lower initial margins than mature AEC products). This is consistent with Q2's guidance for Q3 also carrying 64-66% gross margin guidance, which also came in above range.

Pattern assessment:

Bill Brennan and Daniel Fleming have consistently guided conservatively. In three consecutive quarters, the company beat revenue guidance by 12-20%. Gross margins have consistently come in above the top of the guided range. Full-year guidance has been raised at every quarterly update. This is a management team that appears to have genuine real-time visibility into customer purchasing behavior and chooses to guide below actual demand signals - either from genuine conservatism or as a deliberate investor relations strategy. There is no instance in these four concalls of a guidance miss. There are no promises made and quietly dropped.

The two areas where management has been most forthright about uncertainty are tariffs (explicitly flagged as variable in Q3 FY2026) and the DustPhotonics integration (an earnout structure acknowledging execution dependency). These are appropriate disclosures rather than deflections.

The one counterpoint is that extraordinary growth of this magnitude can invite questions about sustainability. Management's FY2027 guidance of >50% year-over-year growth is the first time they have guided a moderation in growth rate. Whether they beat that guidance as consistently as they have beaten quarterly guidance is the critical question for FY2027.

10. Shareholder Friendliness Index

Dividends: Credo has paid no dividends in any of the last three fiscal years (FY2024, FY2025, FY2026). The company is in a high-growth phase and has reinvested all operating cash into product development, mask tape-outs, and acquisitions. No dividend initiation has been signaled.

Buybacks and dilution: Credo has not conducted a share buyback program. To the contrary: in Q2 FY2026, the company executed an at-the-market (ATM) equity offering that raised a meaningful cash balance (cash increased by $333.9M in Q2, partially from operations and partially from the ATM). This was explicitly dilutive, increasing the diluted share count from approximately 190M (FY2025 full year) to approximately 194M by Q3 FY2026. The equity raise was strategically sensible - the stock was trading near multi-year highs and the capital has been deployed toward the DustPhotonics acquisition. However, it is dilution, not buyback activity. FY2026 share count is higher than FY2024 on a diluted basis, partly from stock-based compensation and partly from the ATM.

Verdict: Hoards Capital - Credo does not return capital to shareholders. The company raises equity at opportunistic prices and allocates it to acquisitions and R&D. For a company tripling revenue in two years with a $10B+ TAM ahead of it, this is the appropriate capital allocation posture, but shareholders should not expect dividends or buybacks in the near term.

11. Insider Activities

Primary Source: SEC Form 4 filings via EDGAR (CIK: 0001807794). Data sourced from secform4.com and stocktitan.net aggregating EDGAR Form 4s.

All transactions below were executed under Rule 10b5-1 trading plans adopted months in advance, which is standard for semiconductor executives with regular vesting schedules.

Buys: No open-market insider purchases were identified over the last 12 months. No insider bought shares in the open market during this period.

Sells - working out the why: Every material transaction reported was executed under a pre-established Rule 10b5-1 trading plan, a mechanism where U.S. insiders file a trading plan with specific parameters months in advance to avoid accusations of trading on material non-public information. The CEO's plan was adopted April 15, 2025, well before the FY2026 revenue acceleration was visible to the public. The CTO's plan was adopted September 5, 2025, just as Q1 FY2026 results were announced. These are routine, scheduled diversification transactions by insiders who hold very large equity stakes (the CTO alone has 6+ million shares). The scales are notable - the CTO's family trust alone is selling 27,500-55,000 shares per tranche, which at current prices represents substantial dollar amounts, but the insider's remaining position is also substantial.

The CEO's charitable gift of 7,500 shares on April 15, 2026 is not a sell signal; it is a donation to a charity and does not indicate a bearish view.

Net assessment: Credo insiders are net sellers on a cash basis - every reported transaction has been a sale. However, the mechanistic explanation (scheduled 10b5-1 plans, vesting-driven diversification) accounts for all of the activity. There are no open-market purchases and no sells outside of plan. The absence of any open-market buying by insiders is mildly notable given the stock's price volatility and the management team's obvious conviction in their growth narrative on concalls, but it is not unusual for executives who receive equity compensation regularly to manage concentration via scheduled sales. Overall: neutral signal - no red flags from plan-based selling; no bullish signal from absence of open-market buying.

12. Scenarios

Bull Case

Credo becomes the connectivity infrastructure layer for the AI infrastructure decade. In this scenario, every major AEC speed transition - from 800G to 1.6T and then to 3.2T - runs through Credo's product generations because it consistently arrives first and qualifies fastest. The fifth and sixth hyperscalers ramp to significant AEC volumes. ZeroFlap Optics ships to four or more hyperscalers by end of FY2027 and captures a meaningful fraction of the optical transceiver market from incumbents by virtue of its reliability advantage and integrated PILOT telemetry. The DustPhotonics acquisition closes cleanly, adds silicon photonics PIC technology, and DustPhotonics hits its earnout milestones - meaning combined optical revenue exceeds $500M in FY2027 and the product roadmap extends to 3.2T without requiring further major acquisitions. Active LED Cables open a new market in 2028 for row-scale AI networking where no AEC alternative exists. OmniConnect wins in the disaggregated AI compute architecture wave as XPU designs proliferate. FY2027 growth comes in above the 50%+ guide, and the market begins discounting a durable multi-year franchise rather than a single-product cyclical.

Marvell's Golden Cable initiative qualifies at hyperscalers but Credo maintains 70%+ AEC share because its PILOT software integration creates enough switching friction and because Credo's next-generation products are already qualified when Marvell's current generation is just shipping. The patent licensing program converts additional cable manufacturers into royalty payers.

Base Case

Credo delivers roughly what management has guided: FY2027 achieves 50%+ year-over-year growth, primarily driven by continued AEC demand as hyperscalers deploy 1.6T infrastructure and ZeroFlap Optics ramps across three to four customers. The DustPhotonics acquisition closes without material delay, adds engineering talent and PIC technology, but the earnout structure means DustPhotonics must execute against ambitious milestones that take 12-18 months to validate. Gross margins remain in the mid-60s range as product mix shifts toward newer, lower-initial-margin optical products. The retimer business (PCIe Gen 6, Blue Heron) adds a meaningful new revenue layer in FY2027 that partially offsets any AEC growth moderation. ALC and OmniConnect remain on the 2028 production schedule without major acceleration or delay.

Customer concentration remains a feature rather than a bug - three to four hyperscalers will drive the vast majority of revenue, which means quarterly results are volatile but the annual trajectory is more predictable given the scale of hyperscaler capex commitments. Marvell qualifies its AEC platform at one or two hyperscalers but does not displace Credo as the primary AEC supplier at any major account within the FY2027 time frame.

Bear Case

A meaningful hyperscaler capex slowdown - triggered by AI investment caution following a notable model failure, regulatory action, or a broader economic slowdown - causes one of Credo's top two customers to pause AEC deployments for a quarter or two. Given that two customers regularly represent 60-70%+ of quarterly revenue, even a 20-30% reduction in volume from the largest customer creates a significant revenue shortfall. Marvell's Golden Cable initiative qualifies at multiple hyperscalers faster than expected, and pricing pressure on AECs increases as the cable manufacturers (now with Marvell architecture and Credo licenses) compete aggressively on price for hyperscaler purchase orders.

In this scenario, ZeroFlap Optics faces delays in hyperscaler qualification (optical transceiver qualifications routinely take longer than copper qualification) and the FY2027 optical revenue ramp disappoints against the >$500M target. The DustPhotonics earnout creates a contingent liability on the balance sheet: Credo has spent most of its cash on the acquisition and may face pressure to raise additional equity at lower prices than the ATM it executed in Q2 FY2026. The retimer business (PCIe Gen 6, Blue Heron) develops more slowly than guided because hyperscaler AI server procurement cycles shift.

The bear case does not require catastrophic failure - it requires a combination of customer purchasing lumpiness, product ramp delays in optical, and competitive pressure in AECs arriving simultaneously. Given Credo's near-zero revenue diversification, these risks compound quickly.

13. Further Reading

• Thick Cables, Thin Margins - Microsoft, Amazon, and Google Demand Overstated By $CRDO Credo - SemiAnalysis [paid] - Critical bear case arguing Microsoft AEC adoption was lower-volume than sell-side models assumed; written in 2023 but the framework remains relevant for stress-testing AEC market share assumptions.

No coverage found from Stratechery or MBI Deep Dives at the time this report was generated.

Sources:

• Credo Technology Investor Relations
• Credo Q3 FY2026 Earnings Call Transcript - Motley Fool
• Credo Q4 FY2025 Earnings Call Transcript - Motley Fool
• Credo Q1 FY2026 Earnings Call Transcript - Motley Fool
• Credo Q2 FY2026 Earnings Call - Yahoo Finance
• Credo Q3 FY2026 Results - BusinessWire
• DustPhotonics Acquisition Announcement - BusinessWire
• Credo ZeroFlap AEC Product Page
• Credo About Page
• Marvell Golden Cable AEC Initiative
• CRDO Insider Trading - secform4.com
• Credo ITC AEC Patent Filing - StockTitan
• SemiAnalysis - Thick Cables, Thin Margins (CRDO)
• AEC Market Outlook 2026-2034 - Intel Market Research
• Credo & TensorWave Partnership - Nasdaq