Credo Technology Group Holding Ltd [CRDO] — Research

Tier-2 memo · framework v1.5.1 · as of 2026-07-29 · spot $192.28 (2026-07-28 close, Alpaca SIP)

This memo issues no position verdict. It scores Criteria and outputs an analysis. The book decides.

---

0. Bottom line up front

Credo is a real business with real, organic, product-volume growth — the accounting-quality work came back clean, and cleaner than any of the precedents this framework was built on (10x Genomics' settlement revenue, Applied Optoelectronics' stocking distributor, Twist's retired orders disclosure). FY2026 revenue tripled to $1,335.1m and the company states in its own MD&A that AEC volume at hyperscale data-centre customers contributed "over 99% of the increase in revenue." Days sales outstanding fell from 109.5 to 48.6. The acquisitions made during the year were explicitly declared immaterial and no pro-forma was required. There is no settlement revenue, no milestone lumpiness, no channel stuffing.

The problem is not what the growth was. It is three things about what it is, all of which are documented in primary filings:

1. The growth rate has already broken. Sequential revenue growth went +51.9% (Q3 FY26) → +7.4% (Q4 FY26), and Q1 FY27 is guided to +7.6%. Management pre-announced this itself on 2026-02-09: "Credo expects sequential revenue growth in the mid-single digits." A sustained ~7.5% sequential rate annualises to ~33.8% — not 93.5%.
2. Three end customers are 84% of revenue, the top ten are ~90%, one direct customer is 53% of receivables, and there is $31.9m of remaining performance obligation against $1,335m of annual revenue — i.e. 2.4%. Everything is on purchase orders cancellable on short notice.
3. This company has already run the downside case. In FY2023 one end customer was 55% of revenue. In FY2024 it was 26% — roughly $101m to $50m — quarterly revenue fell 40.9% sequentially in Q4 FY2023, and the full year grew 4.8%. That is not a hypothetical inventory correction. It is Credo's own three-year-old history.

The screen's +37.2pp valuation margin compares a required 56.3% five-year CAGR against a trailing 93.5% three-year CAGR. The trailing number is dominated by a single year (+205.7% in FY2026) off a $184m base, and the business is no longer running at it. Measured against the rate the company is currently compounding at, the margin is negative, not positive. That is the finding.

---

1. Verification of the screen's inputs

The brief requires every screen input be checked against primary filings. One material error found.

Source: 10-K for FY2026 (period ended 2026-05-02), accession 0001628280-26-043303, filed 2026-06-15; data.sec.gov/api/xbrl/companyfacts/CIK0001807794.json.

Scale cross-check (net income ÷ shares ≈ filed EPS): $472,279k ÷ 188,232k diluted WA shares = $2.509; filed diluted EPS $2.51. Passes. Not dual-class — single class of ordinary shares, Cayman-incorporated.

The one material discrepancy: net cash is overstated by $770m (53%)

10-K Note 16, Subsequent Events:

"In May 2026, the Company acquired 100% of the equity interest in DustPhotonics Ltd. … for a total purchase consideration of $770 million cash and approximately 0.8 million shares of the Company's ordinary shares. In addition, the Company may pay incremental contingent consideration of up to approximately 2.8 million shares and $31.6 million cash…"

The deal closed in May 2026, after the 2026-05-02 balance-sheet date. The $1,443.3m the screen used as net cash therefore includes $770m that has already left the company. Corrected net cash is approximately $673.3m before any FY2027 cash generation (Q1 FY2027 has not been reported; FY2026 operating cash flow was $464.3m against $57.3m of capex, so a quarter of free cash flow is roughly $100m — estimated, not filed).

Effect: EV rises from $34,291m to roughly $36,120m on a fully-diluted share count, EV/Sales from 25.7x to 27.1x, and the required five-year CAGR from 56.3% to 58.0% on the screen's own other parameters. Reported; not silently adopted either way. Both figures appear throughout this memo.

Secondary note on share count. The screen used the 186.5m cover-page figure, which is correct as outstanding. Fully diluted is higher: FY2026 weighted-average diluted was 188.2m and Q3 FY2026 diluted was 192.0m (the gap widens with the share price under the treasury-stock method). This memo uses 192.0m as the diluted count where a valuation is being struck, and states where the 186.5m basis is used instead for consistency with the historical multiple series.

---

2. The mechanism — named, evidenced, and singular

Credo's revenue mechanism is one product family sold to a handful of hyperscalers: ZeroFlap Active Electrical Cables (AECs) for in-rack and rack-to-rack copper interconnect in AI clusters.

The company states it directly in the FY2026 MD&A:

"Revenue for fiscal 2026 increased by $898.3 million as compared to fiscal 2025 primarily due to significant increase in volume unit shipments for AEC products. The sales increase was primarily driven by the ramp-up of our AEC solutions at our hyperscale data center customers during fiscal 2026 which contributed over 99% of the increase in revenue."

That is as specific and evidenced a mechanism statement as a filing produces. It is also the concentration problem restated: 99%+ of $898m of incremental revenue is one product family at a small number of buyers.

Why AECs win where they win

An AEC is a copper cable with a Credo SerDes/DSP retimer at each end. Against passive copper it extends reach at 200G/400G/800G per port; against optics it costs less and draws less power, and — Credo's core marketing claim — it does not "flap" (drop and re-link), which in a GPU cluster stalls a synchronous training job across thousands of accelerators. The trademarked ZeroFlap brand is the whole pitch: reliability at the link layer, priced below optics. The FY2026 earnings release frames the value in exactly those terms: "accelerate cluster time-to-stability, maximize GPU utilization, improve network reliability, and reduce overall infrastructure power and operating costs."

This is a genuine engineering position, built on an in-house SerDes that Credo licensed as IP for a decade before productising it. It is not a narrative.

The stated product portfolio (FY2026 10-K, Item 1)

ZeroFlap AECs · Optical PAM4 DSPs · ZeroFlap Optical Transceivers · OmniConnect · SerDes Chiplets · PCIe Retimers · microLED solutions · PILOT software platform.

Everything other than AECs is currently immaterial to revenue by the company's own 99% statement.

The diversification attempt is real, expensive, and dated

• Hyperlume — acquired 2025-09-29 for $92.0m; microLED chip-to-chip optical interconnect (Active LED Cable).
• CoMira Solutions — acquired 2026-02-25 for $35.1m; link-layer, FEC and security IP for scale-up/scale-out.
• DustPhotonics — announced 2026-04-13, closed May 2026, $770m cash + ~0.8m shares, plus contingent consideration of up to 2.8m shares and $31.6m cash. Silicon-photonics PIC for optical transceivers, 400G/800G/ 1.6T with a 3.2T roadmap.

With DustPhotonics, management gave a quantified target in the 2026-04-13 8-K:

"the company expects its combined portfolio of ZeroFlap Optical Transceivers, Optical DSPs, and Silicon Photonics products to generate greater than $500 million in optical revenue in fiscal 2027."

That is the single most important forward disclosure in the file. If it lands, optical is ~24% of a ~$2.1bn FY2027 and the AEC/hyperscaler-three dependence genuinely falls. It is also the first time Credo's growth will carry an inorganic component — DustPhotonics ships PICs into transceivers at hyperscale AI clusters today, so part of that $500m is bought, not built. FY2027's organic growth rate will not be readable from the headline. The company has disclosed no DustPhotonics standalone revenue figure. Flagged as an open item.

The same release sizes the SiPho PIC market at "$6 billion by 2030" (LightCounting unit data published 2025-11-21, plus Credo estimates). Worth holding next to §5's implied path, which requires Credo alone to reach $8.6–13.1bn of revenue by FY2031.

---

3. Accounting quality — is the reported growth real?

This is the highest-value section of the memo and the answer is: yes, with two forward-looking caveats.

3.1 Not acquisition-driven

10-K Note 5: "Pro forma results of operations for both acquisitions have not been presented because the effect of the acquisitions was not material to the Company's financial results." Hyperlume closed four months into the year and CoMira with nine weeks left. FY2026's +205.7% is organic.

3.2 Not milestone- or licence-driven — and the disclosure that disappeared

Credo's IP-licensing line is the lumpy, non-recurring part of the business. It has been shrinking as a share of revenue and the company stopped disclosing the split in FY2026:

A retired disclosure is exactly the Twist Bioscience pattern this framework flags. But the honest reading here is that it cannot be hiding much: the last disclosed value was 3% of a $437m year, i.e. ~$13m. Even if IP licence revenue tripled in FY2026 it would be under 3% of the year. The omission is a governance demerit, not an earnings-quality finding. Recorded as such and not inflated.

Independent corroboration from the mention-frequency corpus (§6): IP license runs 16.6 mentions per 10k words across the first four quarterly releases and 0.0 in the last two, and SerDes, chiplet and licensing all go to zero in the same two quarters. Credo has rewritten its own self-description from "SerDes IP + AEC" to "system-level optical + electrical". The filings and the language agree.

3.3 No settlement, warrant or one-off revenue — the warrant runs the other way

Credo issued a warrant to Amazon.com NV Investment Holdings LLC in FY2022 whose vesting is tied to purchases; the cost is recorded as contra revenue inside product sales. Cash-flow statement:

"The contra revenue impact associated with the Customer Warrant has been fully amortized as of May 3, 2025." So FY2025 revenue was reduced by $13.2m and FY2026 was not. Gross of the warrant, FY2025 revenue would have been ~$450.0m and FY2026 growth ~196.7% rather than the reported 205.7%. ~9pp of the reported acceleration is the disappearance of a contra-revenue drag rather than incremental demand. Small, but it is a real overstatement of the acceleration and it is not visible on the revenue line.

3.4 Receivables — clean, and improving sharply

Receivables fell in absolute dollars in Q4 FY2026 while revenue rose 7.4%. This is the opposite of the Applied Optoelectronics pattern (264-day DSO at a stocking distributor). Whatever else is true of Credo, it is collecting cash. Operating cash flow of $464.3m on $472.3m of net income confirms it.

3.5 Inventory — the one thing that is building faster than revenue

Inventory rose +20.6% sequentially in Q4 against +7.4% revenue growth. Excess-and-obsolete write-downs were $15.1m in FY2026 versus $8.0m in FY2025. Inventories consumed $174.0m of operating cash in the year.

Two readings, both legitimate, and I will not pretend to know which: (a) a build ahead of the FY2027 optical and AEC ramp, which is what management would say; (b) the first quarter in which shipments lagged what was built — the classic leading indicator of a customer digestion phase. This is the single metric to watch in the Q1 FY2027 print. If days-of-inventory goes above ~180 while sequential revenue growth stays in the mid-single digits, reading (b) is winning.

3.6 Backlog — there effectively is none

10-K Note 4: "The contracted but unsatisfied performance obligation was approximately $31.9 million which the Company expects to recognize over the next fiscal year." That is 2.4% of FY2026 revenue. ContractWithCustomerLiabilityCurrent is $0 at 2026-05-02 — no deferred-revenue cushion at all.

Risk factor, verbatim: "Substantially all of our product sales to date have been made on a purchase order basis. We generally do not obtain long-term commitments with our customers or commitments for minimum purchases… In some cases, our customers may cancel purchase orders on relatively short notice to us and without penalty."

Verdict on §3: the growth is organic, recurring in form (volume shipments, not milestones), and cash-backed. It is not repeatable by contract. Repeatability rests entirely on three customers continuing to place orders.

---

4. Customer concentration — quantified, and it has already bitten once

4.1 The current picture (FY2026 10-K)

Direct customers (the contracting parties — some are ODMs/contract manufacturers ordering for hyperscalers):

Accounts receivable:

End customers (Credo's supplemental table, looking through the contract manufacturers):

Risk factor: "in fiscal 2026, sales to our top 10 customers accounted for approximately 90% of our total revenue."

In dollars, FY2026: Customer D ≈ $441m, Customer B ≈ $427m, Customer E ≈ $254m.

Credo does not name its customers. Amazon is identifiable as a customer from the FY2022 warrant issued to Amazon.com NV Investment Holdings LLC; the mapping of Amazon (or any other hyperscaler) to Customer A/B/D/E is not disclosed and is not asserted here.

4.2 The end-customer table across four years — the churn is the story

The largest single end customer has been 55%, then 26%, then 63%, then 33% — and its identity changed. This is not a stable, contracted, annuity-like customer base that happens to be small in number. It is a design-win business where each new platform generation reshuffles who the dominant buyer is.

4.3 The precedent: FY2024 is what a single customer's correction looks like

Quarterly revenue, derived from the cumulative XBRL periods:

Customer Z went from 55% of FY2023 (~$101m) to 26% of FY2024 (~$50m) — roughly a 50% cut in one year. Full- year FY2024 revenue grew 4.8%. It took five quarters to regain the Q3 FY2023 revenue level.

This is the named cause for the downside case, and it is not a hypothesis. Apply the same 50% cut to today's book: Customer D at $441m or Customer B at $427m halving removes ~$215m, ~16% of annual revenue, from a business the market is pricing for a 37–58% five-year CAGR.

4.4 Geography — the mix flipped in one year

Revenue by ship-to destination (FY2026 10-K):

US-destination shipments went from $65m to $768m in a single year — an 11.8x increase — consistent with direct shipment to US hyperscalers rather than through Asian ODMs. Mainland China has been flat at ~$80m for two years and is now 6.1% of revenue; it is neither the growth engine nor, at that size, a first-order tariff/export- control exposure. Note also that Credo runs R&D in mainland China (99,088 sq ft leased) and Hong Kong, which is a regulatory rather than revenue exposure.

---

5. Is 93.5% a cycle peak? — the direct answer

Yes, on the evidence available, the peak is behind us. The company said so itself before the market did.

5.1 The sequential series, which is where the truth is

YoY has decelerated for four consecutive quarters (274% → 272% → 202% → 157% → 111% guided). Sequential growth collapsed from +51.9% to +7.4% in one quarter.

5.2 Management pre-announced the deceleration on 2026-02-09

8-K Ex-99.1, 2026-02-09, alongside a beat on Q3:

"Looking towards the end of fiscal year 2026 and into fiscal 2027, Credo expects sequential revenue growth in the mid-single digits leading to more than 200% year-over-year growth in the current fiscal year."

Q4 printed +7.4%; Q1 FY2027 is guided +7.6%. Management's own forward sequential rate has been mid-to-high single digits for two quarters and counting.

A sustained 7.5% sequential rate annualises to 33.8%. That, not 93.5%, is the rate this business is currently compounding at.

5.3 The base-effect check the brief asked for

The 93.5% three-year CAGR runs FY2023 ($184.2m) → FY2026 ($1,335.1m). Two things about that window:

• It contains FY2024, which grew 4.8% — the customer-correction year. The CAGR is not a description of steady compounding; it is one flat year followed by +126% and +205.7%.
• The base is $184m. Percentage growth off a $184m base is not evidence about the achievability of the same percentage off a $1,335m base — the FY2031 revenue required is an absolute-dollar question, and §6 of the Valuation document answers it in dollars for exactly this reason.

Single-year decomposition: FY2024 +4.8%, FY2025 +126.4%, FY2026 +205.7%. Two-year CAGR (FY2024→FY2026) 163.0%. One-year 205.7%. The "demonstrated CAGR" is a single extraordinary year wearing a three-year label.

5.4 Operating leverage is guided to reverse

Opex is guided to grow 2.5x as fast as revenue and operating margin to contract ~3.8pp sequentially. The screen's op_margin_delta_pp = +24.8 is a trailing measure; the forward guide points the other way. Some of this is DustPhotonics' cost base arriving ahead of its revenue, which is a legitimate explanation and also exactly why the trailing margin-expansion signal should not be extrapolated.

---

6. Transcript mention-frequency — required core metric

Corpus and its limits, stated first. Alpha Vantage's EARNINGS_CALL_TRANSCRIPT endpoint was unavailable: the shared 25/day quota is exhausted ("our standard API rate limit is 25 requests per day"). No speaker-tagged call transcripts were retrieved, so the prepared-remarks / Q&A split the method requires could not be produced. Per references/mention-frequency.md, the documented first-party fallback was used instead: EDGAR 8-K Exhibit 99.1 quarterly earnings releases, 17 consecutive quarters (2022-03-09 → 2026-06-01), one source throughout, no mixing. Word counts 2,159–2,403 — near-constant, so normalisation per 10k words is meaningful. Counts are word-boundary matched.

The honest caveat, up front: these releases are ~90% standing boilerplate. A term's count therefore moves mainly when Credo rewrites its own company descriptor — which is a deliberate act of repositioning and genuinely informative, but it is not a measure of prepared-remarks intensity and must not be read as one. The embedded "Management Commentary" block is only 81–131 words and is too short to carry a reliable series; its raw counts are reported below without interpretation.

Normalised mentions per 10k words (selected terms; 17 quarters)

Emerging (first-4-quarter average <1.0 → last-4 ≥2.0 per 10k words): ZeroFlap, optical, transceiver, PCIe, retimer, PILOT, OmniConnect, hyperscale, AI, cluster, 1.6T, GPU, demand.

Decaying (first-4 ≥2.0 → last-4 below 40% of it): IP license, SerDes, chiplet, licensing, 100G/200G/400G/800G.

Management-commentary-block raw counts (81–131 words per quarter, reported without interpretation): customer 2,3,0,1,0,1,1,2,3,1,1,1,1,2,0,0,1 · AI 0,0,0,0,1,0,0,0,2,1,1,0,1,0,1,1,0 · demand 0,0,0,0,1,0,1,0,1,0,1,1,2,1,0,0,0 · GPU 0×16 then 1 · ZeroFlap 0×14 then 1,1,0.

The one finding, with independent corroboration

In the two most recent quarterly releases, Credo deleted SerDes, chiplet, IP license and licensing from its own standard self-description — all four had appeared in every one of the preceding fifteen releases — and replaced them with optical, transceiver, PCIe, retimer, PILOT, OmniConnect, ZeroFlap and GPU.

Corroborated independently, in the filings rather than in the language: - IP licence revenue 17% → 15% → 3% → withdrawn across FY2023–FY2026 10-Ks (§3.2). - $897m of acquisitions in ten months, of which $770m is silicon photonics (DustPhotonics, May 2026). - A quantified optical revenue target of >$500m in FY2027 (8-K, 2026-04-13).

The mention series, the revenue disclosure and the M&A all point the same way: Credo has stopped describing itself as an IP-and-SerDes company and is spending nearly a billion dollars to become an optical systems company. That is a real strategic pivot, made from a position of strength, and it is under-discussed relative to the AEC story that carries 99% of current growth.

The disappearance of 100G/200G/400G/800G alongside the appearance of 1.6T in the descriptor is a speed-grade roadmap statement rather than a demand signal, and is read as nothing more than that.

Not claimed: that any of these terms is a "first-ever" mention. The window is the 17 quarterly releases retrieved (2022-03-09 → 2026-06-01) and nothing outside it was examined.

---

7. Quality Criteria — is this a business worth owning?

Archetype: COMPOUNDER. Already profitable ($472.3m net income, 33.3% GAAP operating margin), growing, with gross margin expanding (64.8% → 68.0%). It is not an INFLECTION name — the operating margin is not thin or negative, so the INFLECTION standard would be the wrong test.

Quality Criteria: PASS.

Two demerits recorded rather than buried: - Share-based compensation is $182.6m, 13.7% of revenue and 41% of GAAP operating income. Non-GAAP net income of $662m for FY2026 against GAAP $472.3m is very largely this. Diluted share count rose from 181.2m (FY2025 WA) to 188.2m (FY2026 WA) and was 192.0m in Q3 FY2026 alone. - The CEO's May-2026 special PSU grant (1,437,000 shares across six tranches) adds to that, and its hurdles are used in the Valuation document as a management-revealed forecast.

---

8. Liquidity Criteria

• 60-day average dollar volume $1,881m/day; 20-day $1,487m/day; 8.2m shares/day.
• $35.9bn market capitalisation; public float reported at $30.0bn as of 2025-10-31 (10-K cover).
• Any plausible position for this book is a rounding error against daily volume.
• Options chain: January 2027 is genuinely liquid (open interest 98–3,397 per strike, quoted size 60–435 contracts). June 2027 is not — open interest 1–78 across the strikes examined, and bid/ask spreads above 10% of mid. The HCA lesson applies: the longer-dated expiry is uninvestable at size and must not be proposed.

Liquidity Criteria: PASS (equity, and the January-2027 option chain specifically).

---

9. Momentum Criteria — MEASURED, blocks nothing

252-day realised volatility 90.7% (screen: 90.8% — confirmed). 52-week range $87.81 – $302.52; spot is 36.4% below the 52-week high.

Read: 12-1 momentum is emphatically top-decile cross-sectionally, but the last month is a sharp reversal and the name is in a 36% drawdown from its high. Momentum governs timing, never admission. The timing read is deteriorating, and the deterioration is coincident with the Q4 sequential-growth break — i.e. the market is already repricing the thing this memo identifies.

---

10. Short Mechanism Criteria — MEASURED, scored on every name

Required test: decelerating growth and exhausted margin runway.

• Decelerating growth: YES, unambiguously. Sequential +51.9% → +7.4% → +7.6% guided; YoY 274% → 111% guided.
• Exhausted margin runway: PARTIALLY. Operating margin is already high (33.3%) and has already expanded 24.8pp in one year, and Q1 FY2027 guidance implies ~3.8pp of sequential contraction. Against that, gross margin is guided flat-to-slightly-down but not falling, and the opex step-up is identifiably acquisition-related rather than a loss of pricing.

Both legs are at least partially present. This is the first name in the recent record where the Short Mechanism Criteria genuinely fires on a long candidate, and it is logged for the relative-value fork. Nothing on the long-only book acts on it.

---

11. Sub-sector Criteria

Not a healthcare name, so the reference taxonomy does not have a native bucket. Tagged SMID Growth / AI-infrastructure semiconductors (SIC 3674) for concentration purposes. The correlation warning for the book: CRDO's revenue driver is hyperscaler AI capex, which is the same single factor driving NBIS, MU, SNDK, AMZN, MSFT, META and GOOGL already in this corpus. A CRDO position is not diversifying against those.

---

12. Consensus Criteria — INDETERMINATE

Alpha Vantage returned the rate-limit notice; the shared 25/day quota was exhausted before this memo ran. No Street NTM revenue or EPS, no analyst count, no revision history. Per the criteria this blocks nothing and is recorded as INDETERMINATE.

What is available in its place, and is arguably better: company guidance (Q1 FY2027 revenue $465–475m, GAAP gross margin 66.9–68.9%, GAAP opex $167.6–171.6m) and the CEO PSU revenue hurdles (Valuation §4), which are a management-revealed five-year forecast and are not a consensus substitute but are a genuine external anchor.

---

13. What is unsupported, missing, or estimated

Listed explicitly so nothing here is mistaken for a filed figure.

1. Customer identities are not disclosed and are not asserted. Amazon is identifiable as a customer only via the FY2022 warrant to Amazon.com NV Investment Holdings LLC. The mapping to Customer A/B/C/D/E/Y/Z is unknown. Any published attribution of these letters to named hyperscalers is speculation and none is made here.
2. DustPhotonics' standalone revenue, margin and customer list are undisclosed. The ">$500m optical revenue in FY2027" target is a combined portfolio figure and the organic/inorganic split cannot be computed.
3. Post-year-end cash is estimated, not filed. $1,443.3m less the $770m DustPhotonics cash consideration = $673.3m; Q1 FY2027 cash generation (~$100m on FY2026 run-rate) is an estimate and is excluded from the valuation base as a conservatism.
4. No consensus estimates (§12).
5. No speaker-tagged call transcripts — the mention-frequency corpus is 8-K Ex-99.1 releases and the prepared-remarks/Q&A split required by the method was not produced (§6).
6. Peer own-history multiple percentiles were not computed for ALAB or MRVL; only point-in-time multiples are reported, and ALAB's net cash is incomplete (its MarketableSecuritiesCurrent tag is absent from companyfacts), so ALAB's EV and EV/Sales are stated as upper bounds.
7. The Q1 FY2027 earnings date is not company-confirmed — see the Catalyst Calendar, where it is flagged as estimated from a four-year pattern rather than announced.
8. The FY2026 IP-licence revenue share is unknowable from public filings; the 3% FY2025 figure is used to bound it (§3.2).
9. Terminal margin in the screen's reverse DCF was capped at 16.4% against a demonstrated 33.3%. That is a parameter choice, not a fact, and it moves the answer by ~21pp of required CAGR. Handled explicitly in the Valuation document rather than adopted silently.

---

14. Criteria summary