Credo Technology (NASDAQ: CRDO)

01The Decision, Up Front

Credo makes the physical-layer connectivity — the wires, cable chips, and optical engines — that move data between the tens of thousands of GPUs inside an AI cluster. It is the runaway leader (~73–88% share) in Active Electrical Cables (AECs), the copper-plus-silicon link that is the cheapest, lowest-power, most reliable way to move 800G–1.6T of data over the critical 1–7 metre reach inside and between AI racks. The thesis in one line: this is a picks-and-shovels monopoly on the one connectivity problem that gets harder every GPU generation, run by a team that owns the hardest analog IP in the industry, compounding at +206% revenue with ~48% operating margins and — even after a 31% run — a still-below-0.6 PEG.

The update (18 Jun): two weeks ago this was a BUY at $207 with a +17% path to the $243 weighted target. That call has played out — CRDO is +31% to $271.83, a fresh all-time high, on zero new fundamental news (the move is the 1 Jun beat-and-raise + the 2 Jun target-hike wave + the DustPhotonics optical re-rating, finally digested). At $272 the stock trades through the consensus mean ($248), at ~45× FY27 earnings, with insiders selling into it. So the call moves to HOLD — conviction 3/5: take partial profits, hold the core. The downgrade is purely valuation; the franchise thesis below (§2–§9) is unchanged. The two structural caveats are the same as before: (1) the moat is wide but narrowing (Marvell's "Golden Cable" took Credo from ~88% → ~73% AEC share, and physics will eventually let optics eat copper, §7–8); and (2) customer concentration is extreme (top-4 ≈ 87%). China remains the least of the worries (§9).

02First Principles — The Physics of Moving One Bit

To understand why Credo exists, it helps to understand why moving data across a few metres of an AI data center is one of the hardest problems in the building. It comes down to a single physical fact: copper destroys high-frequency signals, and it destroys them worse every time you double the speed.

Bits, baud, and the move to PAM4

A serial link sends bits down one "lane." The two ways to encode them:

Why copper fails — the "reach wall"

Three physical mechanisms eat a copper signal, and all three get worse with frequency: skin effect (current crowds into a 0.42 µm surface layer at 25 GHz; loss ∝ √f), dielectric loss (the insulator turns signal into heat; loss ∝ f), and conductor roughness (surface features become significant versus the shrinking skin depth, adding 15–25%). A 2 m AWG-30 twinax cable at 112G PAM4 (28 GHz Nyquist) already loses 20–25 dB. The OIF reach budgets cap a long-reach copper channel at ≤40 dB of insertion loss — and the receiver must still hit a raw bit-error-rate ≤ 2.8×10⁻⁴ so that RS(544,514) Forward Error Correction (5.84% overhead) can scrub it to a post-FEC target of 1 error per 10¹⁵ bits.

Here is the punchline, and the entire reason Credo's market grows: when you double the lane rate, the Nyquist frequency doubles, and copper loss at that frequency rises ~40–70%. Passive copper that reaches ~2 m at 800G reaches under 1 m at 1.6T. The reach wall closes in every generation.

03The Tradeoff Triangle — Three Ways to Move a Bit

Every link in a data center is an optimisation across three axes that fight each other: reach (metres), power (watts / picojoules-per-bit), and cost ($/port). Nothing wins on all three. There are three families, and Credo now plays in all of them.

Where each lives in the rack

04What Credo Actually Makes — Down to the Silicon

Everything Credo sells is built on one foundational capability: world-class SerDes (Serializer/Deserializer) mixed-signal IP. Master that, and you can sell it as licensable IP, embed it in a retimer chip, put that chip in a cable (AEC), or put a cousin of it inside an optical module. Here is the full stack, physically.

The product families

05Why Credo Is the Market Leader

Credo was founded in 2008 by Bill Brennan, Lawrence Cheng and Job Lam — veterans of Marvell's PHY group and Sun Microsystems — and spent a decade as an invisible SerDes IP licensor before its "Big Bang": the HiWire Switch AEC with Microsoft in July 2021. That single Tier-1 design win gave it a 2–3 year qualification head-start that nobody has fully closed.

The share story (and an honest number)

650 Group — the analyst of record for this market — put Credo at ~88% of AEC unit share in 2024, declining to ~73% by Q2 2025 as Marvell's ecosystem and Astera's Taurus began shipping. Both numbers are real: 88% reflects the period when Credo was the only qualified vendor; 73% is the honest current state as second-sourcing arrives. The trajectory matters as much as the level — and this is returned to in §8.

Five real sources of leadership

• Vertical integration + an "N-1" node trick. Credo owns IP → chip → cable → firmware → PILOT telemetry. Crucially, it builds AEC chips on mature 12nm/28nm nodes (not bleeding-edge 3nm), which is cheaper and supply-secure — a key reason it holds 68% gross margins on a product that ships with a physical cable. Marvell's Alaska runs on costlier 5nm.
• Power. AECs draw ~50% less power than optical — up to 14 W saved per 800G link; at 1.6T the ratio is ~6–8× (≈2.5 W AEC vs 15–20 W optical). In a 100k-GPU cluster, that is megawatts.
• ZeroFlap reliability. ~100–1,000× fewer link disruptions than optics, validated by hyperscaler operations — not a spec-sheet claim. A single flap can crash a distributed training run.
• Qualification lock-in. Qualified at Microsoft (2021), Amazon (2023), xAI (2024), Google (2025), and a 5th hyperscaler (Q2 FY26). Each cycle is 12–18 months; the PILOT telemetry platform then embeds Credo into operations as a software dependency.
• Patent judo. Rather than fight cable houses, Credo licenses them: patent deals with Siemon (Nov 2025) and 3M (Jan 2026) turn would-be competitors into royalty streams and put legal risk on unlicensed entrants. 700+ patents issued/pending.

06Why This Is Genuinely Hard to Build

The bull case lives or dies on one question: can a well-funded competitor just spin up an AEC business? The honest answer from the physics is "yes, but slowly, and only if they already have the rarest ingredient." Here is why world-class SerDes is a structural barrier.

• Analog talent is the bottleneck, and it is tiny. A sub-100-femtosecond-jitter PLL at a 100+ GHz VCO, in 3nm CMOS, drawing <5 mW, is not a software problem — it is transistor-level analog artistry. The global pool of engineers who can do this at the frontier is perhaps a few hundred people, concentrated at Credo, Broadcom, Marvell, Alphawave and a handful of hyperscaler teams. Credo was purpose-built around this scarcity.
• Silicon iteration is slow and expensive. A 3nm SerDes tape-out costs $15–50M in mask NRE alone; a single PLL/DFE bug burns a 6–9 month cycle. Credo's Cardinal is its 7th-generation DSP architecture — ~16 years of accumulated characterisation data and methodology a 2024 entrant simply cannot buy.
• It's analog and algorithm co-design. The Blue Heron retimer runs a 30-tap FFE plus 16 reflection-cancellation taps tuned specifically to AI-backplane channel reflections — not a generic commercial SerDes. The "secret sauce" is the marriage of novel analog front-ends with proprietary digital adaptation that converges across every process/voltage/temperature corner.
• Foundry co-optimisation + design-win lock-in. Securing TSMC N3 capacity for IP that competes with TSMC's own SerDes offering takes years of relationship. And once a hyperscaler builds a custom ASIC around Credo's SerDes IP, switching mid-tape-out is impossible — a 3–5 year revenue lock.

07Every Meaningful Competitor in the World

Here is the full global competitive map, ranked by threat, with what overlaps and why.

The one that matters: Marvell's "Golden Cable"

Launched 9 December 2025, Golden Cable is not a product — it's an ecosystem play. Marvell hands 20+ cable houses validated reference designs, calibration data and firmware so they can build "good-enough" 1.6T AECs around Marvell's Alaska DSP. Foxconn completed a 1.6T AEC in two months using it. The genius (and the threat) is that Marvell doesn't need to beat Credo on a spec sheet — it just needs to give hyperscalers a credible second source to reduce single-vendor dependence and pressure price. The 88%→73% share slide already partly reflects this. Expect Marvell at 20–25% AEC share by 2027.

08Is the Moat Impenetrable? The Honest Verdict

No — it is not impenetrable. But "not impenetrable" is very different from "spinnable by anyone." Here is the balanced teardown.

09What About the Chinese?

The China question is the sharpest one for this thesis, and the answer is reassuring on competition and more nuanced on geopolitics. China is a low-to-moderate risk for Credo — and counter-intuitively, it's a supply-chain/manufacturing risk far more than a "Chinese rival steals the market" risk.

Can Chinese firms replicate it? Not at the frontier — yet.

• Montage Technology (688008.SS) is the only credible Chinese name. It's bigger than expected (~$500M FY24 rev, +59%) — but its bread-and-butter is DDR5 memory-interface chips (36.8% global share), which is not Credo's business. Its overlap is PCIe6/CXL3 retimers and a Jan-2026 PCIe-based AEC — but its self-made SerDes tops out at ~64 GT/s, roughly 7× slower per lane than Credo's 224G. And it sells into China-domestic servers, not US hyperscalers.
• No Chinese company ships production 224G SerDes. The only firms at that frontier are Broadcom, Marvell, MediaTek and Credo. Estimated catch-up gap: 3–5 years at 112G, 5+ years at 224G — and that assumes foundry access that EUV restrictions currently deny.
• Huawei's UB-Mesh (Hot Chips 2025; ~150 ns hop, scaling to 1M NPUs) proves China can build frontier interconnect when motivated — but it's in-house for Ascend only, not sold to anyone, so it doesn't compete for Credo's sockets.
• Innolight / Eoptolink dominate optical module assembly (7 of top-10 slots) — but they buy their DSPs from Marvell and Broadcom. There is no mature Chinese domestic DSP alternative. This is assembly, not the silicon Credo makes.

Credo's own China exposure — and the scar from 2023

Here is the part most investors miss. Credo's revenue is reported by manufacturing geography, not end-customer. In FY25, "Greater China" was ~75% of revenue (HK 55.8%, China 18.3%) — but that is where the cables are assembled by contract manufacturers, for US hyperscalers in Seattle and San Francisco. The paying customers are American.

• Export controls: SerDes/AEC/retimers are not currently named in BIS controls (those target GPUs, HBM, EUV). They sit in a gray zone that could be extended — but Credo barely sells to Chinese clouds anyway, so the demand hit would be small.
• TSMC concentration: 100% of leading-edge silicon (incl. N3) is at TSMC/Taiwan — a binary tail risk, but one shared by Nvidia, Broadcom, Marvell, Astera and every fabless peer.
• Tariffs: the CFO flagged the tariff regime as "fluid"; cables assembled in China/HK for US buyers carry Section-301 exposure. 68% gross margins absorb it; hyperscalers renegotiate it.

10The Tape — What the Market Is Saying Right Now

A 30-day social-sentiment sweep, two passes pulled 8 Jun 2026 (quotes dated as shown). Read in the context of the +31% move since. Citations clickable.

The 30-day sweep ran twice — once on the stock debate, once on the moat/competition question. As of 8 Jun the crowd's verdict corroborated the then-BUY call: the bull narrative (platform shift to optics, "nervous system of AI") was intact, and the post-earnings drop was read as a valuation reset on a clean beat — the asymmetry that has since played out.

"Just came out of the $CRDO earnings call. The stock fell, but the call made the bigger story clearer. Credo is moving from an AEC/copper-cable winner to a broader AI connectivity platform … Revenue grew 157% YoY to $437M. Non-GAAP operating margin 49.6%. FY27 80%+. The real standout was optics."
— @EmmaStockNotes, 1 Jun · the platform-shift thesis (top-scored item)

"$CRDO drops despite blowout results … −4.2% regular, −8% to −11% after hours. A classic case of 'Great earnings. Higher expectations.'"
— @EmmanuelInvest, 2 Jun · the consensus framing of the drop

"🧠 If $NVDA is the brain of AI, then $CRDO is becoming the company that builds the nervous system — high-speed interconnects, AECs, optical DSPs, silicon photonics, SerDes."
— @TaoAlphaTrader, 2 Jun · the secular bull frame

"Redburn launched coverage with a Buy and a $206 target … the CTO sold 27,500 shares (~$5M) while the stock pushed new highs — but still holds over 6.1M shares … the real story is 157% revenue growth and the $600M optical runway."
— The Strategic Investor on YouTube · DustPhotonics + insider-sale color

The moat debate, in their words

The second sweep targeted the moat-durability question directly — Marvell, and the copper→optical transition. The debate on X is unusually well-articulated, and it maps cleanly onto §2 (the reach wall) and §8 (the moat verdict):

"$CRDO at ATH is a textbook case of breaking out of a technical dead end. Their core business — AEC copper cables — was headed straight for the bin. At 800G and 1.6T, copper physically dies due to signal attenuation, and NVDA is aggressively shifting data centers to optics."
— @edge_of_power, 22 May · the real bear case — exactly the §2 reach wall, weaponised

"Long $CRDO for a while, and the DustPhotonics acquisition is why I'm not selling a share. The perennial bear case: 'what happens when the industry shifts from copper to optical?' With the ~$1.3B DustPhotonics deal, $CRDO [now owns SerDes + DSP + SiPho]."
— @RosannaInvests, 7 Jun · the bull rebuttal — the §4/§8 optical-pivot hedge

"$CRDO — AEC is replacing DAC inside the data center. MSFT + Amazon design wins. Years ahead … vertically integrated: SerDes + DSP + SiPho PICs + optics — positioned across pluggable, near-packaged, and co-packaged."
— @InvestmentGuru_ & @angrybear168, 3–8 Jun · the full-stack moat thesis

11The Investment Decision — Re-Marked at $271.83

Tying the physics, the moat and the tape back into Jeff's valuation frame — at the new price. At $271.83 (a fresh ATH, +31% from the 5 Jun mark), Credo trades at ~45× FY27 earnings for ~80% growth — a ~0.55 PEG. That is still cheaper per unit of growth than peer Astera (~75× / ~1.8 PEG), but the gap has narrowed and the absolute multiple now embeds a near-flawless optical ramp. The stock sits above the consensus mean ($248), with 15 of 18 analyst targets below spot.

What the move did to the math: at $207 the bear/base/bull of $135/$250/$335 implied a +17% weighted path — a clear buy. The stock has since captured the entire base case and most of the way to the bull. Re-striking the scenarios at today's higher base of validation (optical TAM more proven, FY27 >80% reaffirmed) but a higher entry, the weighted target lands at $263 — roughly flat-to-down 3% from spot. The risk/reward has inverted from "asymmetric up" to "balanced-to-rich."

Weight it independently

Note: the bull case now needs a weight ≳ 45% just to get the target back above spot — a precise statement of how much the easy money has already been made.

How to actually own it now

• For existing holders (e.g., positions built near $207): the call worked, +31%. Trim 25–40% into this strength to lock the gain, and let a core position ride the optical/1.6T story. Round-tripping the whole win while waiting for the bull case is not the recommended path.
• For new money: wait. Entry at $272 (above every target, ~45× FY27) is paying up for momentum. A starter position only makes sense on a pullback toward $235–240 (~38× FY27 / the CFO-sale zone), adding toward $200–210 (where the prior buy thesis re-arms with a real margin of safety).
• Sizing: 1.5–2.5% of a thematic AI-infrastructure sleeve at a disciplined entry; β≈3.18 and top-4 ≈ 87% concentration argue against paying up at the high. Pair against lower-beta MRVL.
• Re-upgrade to BUY if price <$215 with thesis intact, OR a clean optical-revenue inflection (>$600M FY27 tracking) de-risks the multiple. Thesis-break (exit regardless): AEC share below ~70% on an accelerating slope; a top-customer capex pause/vendor shift; two quarters of decelerating bookings.