Optical Computing
A real moat in AEC reliability and an n-1 cost edge, riding a verticalizing AI-interconnect ramp — but ~84% of revenue sits in three hyperscalers and ~41x forward earnings already prices in the optical pivot working. Bullish business, expensive stock; conviction gated on customer diversification proving out.
Research
The verdict
A real moat in AEC reliability and an n-1 cost edge, riding a verticalizing AI-interconnect ramp — but ~84% of revenue sits in three hyperscalers and ~41x forward earnings already prices in the optical pivot working. Bullish business, expensive stock; conviction gated on customer diversification proving out.
Credo Technology Group Holding Ltd is a Cayman-domiciled, fabless connectivity-semiconductor company that sells the silicon and cables that move data inside AI data centers at high speed and low power. Its mission, in its own words, is "to transform connectivity at scale through fast, reliable and energy-efficient system solutions". The portfolio: ZeroFlap (ZF) Active Electrical Cables (AECs), ZF optical transceivers, OmniConnect memory solutions, and a suite of retimers and DSPs for optical and copper Ethernet and PCIe — all built on Credo's own SerDes and DSP IP and wrapped in its PILOT diagnostic/analytics software. Products span 32G→200G per lane and target Ethernet, PCIe, and the emerging UALink/ESUN/SUE scale-up fabrics.
The business model is product-first with an IP-licensing tail (SerDes IP). It is fabless — it contracts manufacturing to third parties (almost entirely TSMC for ICs) — letting it concentrate spend on design and hold fixed costs/capex low. Credo runs a two-pronged sales motion: sell directly to the end user (the hyperscaler) and to the end user's suppliers (OEMs/ODMs/CMs/optical-module makers), so it often becomes a designated-source vendor the supply chain is told to use.
Customers are hyperscalers and NeoClouds. The signature design win: Credo partnered with Oracle to develop ZeroFlap Optics to kill "Link Flap" — the reliability failure that plagues commodity optics in AI clusters and delays time-to-first-revenue on a new GPU pod. By end-customer identity, the three ramped AEC hyperscalers are Amazon, Microsoft and xAI, with Meta also a >10% Q4 customer. Contract structure is standard purchase orders (cancellable within limited notice), 12-month warranties, limited distributor return rights — i.e. no take-or-pay; demand is backlog-and-PO driven, not contracted recurring.
Map: TSMC (wafers) → 3rd-party assembly/test (Taiwan) → Credo (design/IP/DSP) → cable assemblers & optical-module makers / OEMs / ODMs / CMs → hyperscaler & NeoCloud buyers (end customers).
Named, sourced stakeholders:
Chokepoints: (1) TSMC single-source — geopolitical Taiwan risk is explicitly disclosed (cross-strait tension, "a substantial amount of our revenue is derived from products manufactured in Taiwan"); (2) Hong Kong exposure — 28% of FY26 revenue shipped through Hong Kong, with the company flagging Hong Kong National Security Law / sanctions risk.
Credo's moat is narrower than a foundry's but real, and it is concentrated in three things:
Reliability as a productized moat ("ZeroFlap"). Link Flap — intermittent optical-link failure — is the thing that delays an AI cluster's turn-on and burns hyperscaler capex. Credo turned reliability into a branded SKU and co-developed it with Oracle. In AI infrastructure, a cable that doesn't flap is worth a premium because the GPUs behind it cost far more than the cable. This is a switching-cost moat: once a hyperscaler qualifies Credo AECs into a pod design and they just work, re-qualifying a rival to save a few dollars per link is not worth the risk to a multi-billion-dollar buildout.
n-1 cost/availability edge. By matching leading-edge performance on a trailing node, Credo undercuts on cost and dodges the wafer-allocation queue. That shows up directly in a 68.0% FY26 gross margin — fat for a hardware/cable vendor.
Neutrality. Because Credo does not make switch ASICs, it is a Switzerland of interconnect: a hyperscaler can buy a Broadcom or Astera switch and still choose Credo cables to save cost/power. Vertically integrated rivals (Broadcom, Marvell) carry channel conflict Credo doesn't.
Bargaining power is asymmetric and unfavorable on the demand side: with ~84% of revenue in three customers (below), the hyperscalers hold the whip hand on price. Credo's counterweight is being designed-in and hard to second-source on reliability-critical links — but that is a moat that must be continually re-won each product generation. On the supply side Credo has limited leverage over TSMC (it is a small n-1 customer), partly offset by paying refundable deposits to reserve capacity ($71.3M of cash went into "refundable deposits to the suppliers in exchange for reserved manufacturing production capacity" in FY26).
Credo reports one operating and one reportable segment — the CEO (CODM) reviews consolidated financials. So the only breakouts are product family (qualitative) and geography (quantitative).
By product: revenue "primarily consist[s] of shipments of our AEC and ICs products". The FY26 surge was overwhelmingly AECs — "the ramp-up of our AEC solutions at our hyperscale data center customers during fiscal 2026 … contributed over 99% of the increase in revenue". The forward growth narrative shifts toward optical: management guided the optical portfolio to >$600M in FY27, i.e. the next leg is DSP/optical transceivers, not just AECs.
By geography (destination of shipment, $000s):
| Region | FY2026 | FY2025 | FY2024 |
|---|---|---|---|
| United States | 768,051 | 65,097 | 49,569 |
| Hong Kong | 378,230 | 243,727 | 70,162 |
| Mainland China | 80,924 | 80,055 | 28,264 |
| Taiwan | 22,727 | 3,624 | 21,286 |
| Rest of World | 85,184 | 44,272 | 23,689 |
| Total | 1,335,116 | 436,775 | 192,970 |
The trend is decisive: US-destination revenue went from $65.1M to $768.1M in one year as the North American hyperscaler ramp landed (management notes 58% of FY26 revenue was from North America customers vs 15% in FY25). Shipment-geography lags customer-geography because of the Asia contract-manufacturing handoff (28% still ships through Hong Kong).
This is a hyper-growth print, not a normal one. FY2026 (year ended May 2, 2026):
| Metric | FY2026 | FY2025 | FY2024 |
|---|---|---|---|
| Revenue ($000) | 1,335,116 | 436,775 | 192,970 |
| YoY growth | +205.7% | +126.3% | — |
| Gross profit | 908,349 | 282,909 | 119,431 |
| Gross margin | 68.0% | 64.8% | 61.9% |
| Operating income (loss) | 445,005 | 37,124 | (37,058) |
| Operating margin | 33.3% | 8.5% | (19.2%) |
| Net income (loss) | 472,279 | 52,183 | (28,369) |
| Diluted EPS | $2.51 | $0.29 | $(0.18) |
| Operating cash flow | 464,292 | 65,083 | — |
Drivers: AECs at hyperscalers drove >99% of the revenue increase. Gross margin expanded +3.2pts on scale economics. The operating-leverage story is the headline — R&D fell from 33.6%→20.9% of revenue and SG&A from 22.6%→13.8% even as both grew sharply in absolute dollars (R&D +90% to $279.4M, SG&A +86% to $184.0M), so operating margin quadrupled to 33.3%.
Quarterly trajectory confirms a clean sequential ramp:
Balance sheet — fortress: cash & equivalents $1,165.0M + short-term investments $278.3M = $1.44B liquidity; total liabilities $232.0M (no funded debt — liabilities are payables + leases); equity $2,063.6M; total assets $2,295.6M. Working capital $1.8B (vs $605.8M a year ago). Funded partly by an ATM equity raise: $736.3M net proceeds, 4.8M shares, via Goldman Sachs.
Flags vs its own history: (1) Inventory built faster than revenue — inventories +178% to $250.8M ($90M→$250.8M) while revenue grew 206%, "to support unfulfilled backlog and related new product ramps"; defensible in a ramp but a leading indicator to watch if demand wobbles. (2) Receivables +44% to $233.4M — slower than revenue, which is healthy. (3) Stock-based comp is large: $182.6M (13.7% of revenue), the single biggest non-cash add-back to OCF — GAAP and non-GAAP diverge meaningfully here (see Lens 10). Market reaction to the Q4 print was negative despite the beat — the stock fell ~4–14% intraday because guidance was merely in-line with elevated expectations.
No transcripts in the research layer (transcripts/ empty), so this is ``. The arc across the last several quarters:
Peer set: pure-play AI-connectivity and SerDes/DSP names.
| Company | Ticker | Mkt cap | Fwd P/E | EV/Sales | Notes |
|---|---|---|---|---|---|
| Credo Technology | CRDO | ~$46.3B | ~41.0 | n/a | Trailing P/E ~90 |
| Astera Labs | ALAB | n/a | n/a | n/a | Closest direct pure-play peer |
| Marvell | MRVL | n/a | n/a | n/a | Diversified; entering AEC DSP |
| Broadcom | AVGO | n/a | n/a | n/a | Switch ASIC + CPO incumbent |
| Amphenol | APH | n/a | n/a | n/a | Cable-assembly leader, Marvell partner |
| Coherent | COHR | n/a | n/a | n/a | Optical components |
5-year average ROE: n/a for the peer set (and not meaningful for CRDO, which was loss-making as recently as FY24; FY26 ROE on average equity ≈ 34% ).
Read: CRDO at ~41x forward earnings is priced as the premium pure-play in the group. I could not source ALAB/MRVL/AVGO forward multiples from the searches run, so I will not fabricate them — the honest statement is that CRDO trades at a clear growth premium to diversified semis (AVGO/MRVL trade at materially lower forward multiples in general market knowledge, but I have no dated source here, so: n/a). The comparison that matters is ALAB, and I do not have its current multiple to hand.
``. The pattern is unusually clean and tells you what the market trades on:
Takeaway: the market reacts to (1) the forward guide relative to a ratcheting bar, (2) the customer-diversification story, and (3) sector optical/CPO sentiment — far more than to the reported quarter.
for biography/ownership; for the comp grant and 10b5-1 facts.
| Tranche | Revenue hurdle | Stock-price hurdle | PSUs |
|---|---|---|---|
| 1 | $2.5B | $244.70 | 239,500 |
| 2 | $3.5B | $293.64 | 239,500 |
| 3 | $4.5B | $342.58 | 239,500 |
| 4 | $5.5B | $391.52 | 239,500 |
| 5 | $6.5B | $440.46 | 239,500 |
| 6 | $7.5B | $489.40 | 239,500 |
This is management telling you its own internal ladder: $2.5B (≈ FY27/28 territory) up to $7.5B revenue, with the stock roughly doubling to nearly tripling. Strongly shareholder-aligned, but also a clear signal the board expects multi-year hyper-growth — and the comp expense behind it will keep SBC elevated.
Acting forensically over the FY26 statements:
Regulatory findings (required). Read regulatory/regulatory-findings.md (fetched 2026-06-18):
Built bottom-up from FY26 actuals + guidance. FY26 diluted EPS = $2.51 on $1,335.1M revenue, 188.2M diluted shares. Anchors: company guided >80% FY27 revenue growth, optical >$600M; Street FY27 revenue consensus $2.154B; Q1 FY27 guide $465–475M (Street $470.4M), Q1 EPS ~$1.16.
Assumptions held constant unless noted: gross margin ~67% (mild mix dilution as lower-margin optical/DustPhotonics scales, partly offset by scale), operating margin drifting 34%→38% on continued leverage, SBC stays ~12–13% of revenue (so GAAP < non-GAAP), tax low-single-digit %, diluted share count +~3%/yr (ATM done but RSUs/PSUs dilute), other income ~$30M+ on the cash pile.
| Path | FY27 rev | FY28 rev | FY29 rev | FY27 GAAP dil. EPS | FY28 | FY29 |
|---|---|---|---|---|---|---|
| Base | $2.15B (Street; +61%) | $3.1B (+44%) | $4.0B (+29%) | ~$4.40 | ~$6.60 | ~$9.00 |
| Bull | $2.45B (+83%, mgmt high end) | $3.7B | $5.0B | ~$5.30 | ~$8.50 | ~$12.00 |
| Bear | $1.75B (concentration shock / DAC-back) | $1.9B | $2.2B | ~$3.10 | ~$3.60 | ~$4.50 |
for all EPS: e.g. Base FY27 ≈ $2.15B × 67% GM − ~29% opex ratio → ~36% op margin → ~$774M op income; + ~$35M other income; ~3% tax → ~$785M net; / ~194M diluted shares ≈ $4.05–4.45. (The Special PSU revenue ladder — tranche 1 at $2.5B — implies management sees FY27/early-FY28 clearing $2.5B, consistent with the bull path.)
At ~$251 the stock is ~57x base FY27 / ~38x base FY28 GAAP EPS. On non-GAAP the multiple is lower but still a premium. The valuation only "works" if the FY28–29 base-or-bull path lands; the bear path (a customer pulls AEC volume back toward DAC/optics, as Microsoft once did per SemiAnalysis) would compress both EPS and the multiple — a double hit.
(Per --watchlist rules: no forecast.ts create logged in this unattended sweep. Base call to log if promoted: "CRDO FY27 (ending ~Apr-2027) non-GAAP diluted EPS ≥ $5.00", p≈0.55, resolves 2027-05-01.)
Bull case. Credo is the designed-in reliability standard for AI-cluster interconnect at exactly the moment hyperscaler capex is inflecting and every undelivered GPU pod is lost revenue. The moat (ZeroFlap reliability + n-1 cost + switch-ASIC neutrality) is durable enough to defend a 68% gross margin, and the company is widening it: from 1 to 3-to-5 ramped hyperscalers (de-risking concentration), and from copper AECs into optical DSP + silicon photonics (DustPhotonics) — a TAM expansion from AECs into the much larger optical-transceiver market, with optical guided >$600M in FY27 alone. Operating leverage is spectacular (op margin 8.5%→33.3% in one year) and the balance sheet is a $1.44B-cash fortress with no debt. Management is Marvell-pedigree, 21% insider-owned, and just signed a CEO comp plan that only pays out on a $2.5B→$7.5B revenue ladder. The CPO threat — the main long-term displacer — was just pushed to 2028–2029, extending the AEC/pluggable runway. Earnings surprise lever: a fourth/fifth hyperscaler ramping faster than modeled.
Bear case. Three risks that could permanently impair the thesis: (1) Customer concentration — ~84% of revenue in three buyers (Amazon ~49%, plus two more), and these are the most sophisticated, most price-aggressive, most vertically-integrating customers on earth (they design their own silicon and can second-source or move to cheaper DAC/optics, as Microsoft historically did with AECs per SemiAnalysis). One design-out is a cliff, not a dip. (2) Architectural displacement — AECs are a copper bridge; the long arc is optical (LPO/NPO now, CPO later). Credo is buying its way into optics (DustPhotonics) precisely because it knows copper's days at the longest reaches are numbered; if the optical pivot is late or sub-scale, the AEC cash cow erodes faster than optical replaces it. (3) Valuation/expectations — at ~41x forward / ~90x trailing with the stock already punishing in-line guidance after a beat, the bar is set where even excellent execution can disappoint.
Pre-mortem (18 months out, thesis broke): Most likely failure — a single large hyperscaler (Amazon) re-architects a generation toward DAC + commodity optics or qualifies a second AEC source (Marvell/Amphenol), AEC revenue growth stalls just as the inventory build ($250.8M) turns to write-downs, optical revenue ramps slower than the >$600M guide, and a stock priced for 40%+ growth de-rates to a market multiple — a 50%+ drawdown from compressed EPS × compressed multiple.
Multiples too high? On an absolute basis yes (90x trailing); on a PEG basis defensible if the base/bull growth lands. The honest answer: the price already pays for the optical pivot succeeding.
Contrarian view (what the market refuses to see): The bull crowd treats the optical/DustPhotonics pivot as obvious upside; the under-appreciated risk is that it is defensive — Credo is paying $770M to avoid being a one-product copper-cable company as copper's reach shrinks. The contrarian read is that the AEC franchise is closer to a peak-share moment than the consensus growth curve implies, and the real question is whether optical can carry the load before AEC concentration bites.
I am short CRDO. Here is the dismantling:
A $2.5B market cap on $682K of FY25 revenue — QUBT is a $1.5B treasury wrapped in a photonics R&D lab, sold as a quantum-computing story; the balance sheet is real, the revenue is not, and a securities-fraud class action over the exact gap between the two is unresolved.
The arms-dealer of the AI optics build-out — Lumentum owns ~50-60% of the 200G/lane EML laser chip that every 1.6T transceiver needs, NVIDIA just bought $2B of preferred to lock its capacity, and revenue is compounding ~90% YoY off a real telecom trough; but at ~52x forward earnings with two customers = ~40% of revenue and a $3.8B convertible stack now in-the-money, the price already discounts flawless execution.
A 35-year science project that just turned the corner from lab to foundry PDK — credible polymer-photonics platform now inside Tower & GlobalFoundries flows, but $237K of revenue against a $1.5B cap means you are buying a 2027-28 design-win option, financed by perpetual dilution, with a 17.7% short base betting it stays a promise.