Marvell.
More silicon per AI system.

What was pulled, and what changed

I reviewed the full August 27 Q2 FY2027 and May 27 Q1 FY2027 earnings calls, selected March 5 Q4 FY2026 passages, and the latest 10-Q’s financial statements, notes, MD&A and key risk disclosures. The August call exactly matches the refreshed FMP transcript after whitespace normalization. The latest 10-Q covers August 1, 2026 and was filed August 28; the filing search through October 1 found no newer 10-Q. Public issuer and SEC references are linked below. [1] [2] [3] [4]

The change since the September 29 screening: MRVL’s #4 position was a research-priority rank, not a valuation-adjusted buy ranking. This report adds a full earnings bridge, filing analysis, current valuation and stock scenarios. It strengthens the operating thesis while quantifying a more modest base-case return at today’s price.

Repeated increases matter more than a single earnings beat. However, the quarterly beat itself was small: Q2 revenue exceeded the $2.7B midpoint by $39.3M, about 1.5%, and adjusted EPS beat the midpoint by one cent. The stronger signal was the higher multi-year outlook. [2] [3] [4]

Where Marvell earns money

Marvell is a fabless semiconductor designer. It owns and develops complex chip designs and intellectual property, integrates customer designs where needed, and purchases manufacturing and packaging services. TSMC is its sole advanced-node wafer supplier. It is neither a hyperscaler selling cloud capacity nor a semiconductor foundry. [1]

There are two economic models inside the business. Merchant products are reusable products sold across customers: optical DSPs, switches, controllers and analog components. Custom silicon is built around a specific customer’s architecture; Marvell combines design services, its own IP, advanced packaging and manufacturing execution, then earns product revenue as the program ships. A large custom program can create years of revenue, but switching customers to that same custom chip is difficult if the original buyer cancels.

These are analytical categories, not separately audited operating segments. Marvell reports one operating segment and two end-market revenue categories. Custom memory products belong in the custom bucket in our model; they are not added again under memory or connectivity. [1] [2] [3]

The agent-demand thesis has a direct connection here. The May call links agents to repeated model calls, more routing, larger memory requirements and additional CPU-associated NICs and PCIe connectivity. This makes MRVL a way to participate in the infrastructure surrounding compute. It does not establish a fixed amount of MRVL revenue per agent or per CPU. Better software, workload placement and utilization can reduce hardware intensity per task. [3]

The quarter shows real operating leverage

The next-quarter guide is $3.15B ±5%, or about 15% sequential growth at the midpoint. Data center is expected to grow more than 20% sequentially, while communications/other declines in the low-to-mid teens. Using the $12B full-year outlook, first-half revenue of $5.1571B and the Q3 midpoint implies Q4 revenue around $3.693B, another 17.2% sequential increase. That is arithmetic from rounded guidance, not separate Q4 guidance. [1] [2]

Watch operating margin, not gross margin in isolation. Q3 adjusted gross margin is guided to 57.5–58.5%, below Q2’s 58.9%, because custom silicon is becoming a larger part of sales. That can coexist with improving operating margin when the engineering expense base grows more slowly than revenue. The CFO expects 38–40% adjusted operating margin as FY2027 ends and the upper end during FY2028. He also expects the FY2028 adjusted tax rate to rise from 11% to 13%. [2]

With $18B revenue, a 58% gross margin and FY2028 operating expenses of roughly $3.19B—25% above the $2.55B FY2027 outlook—the arithmetic yields about 40.3% adjusted operating margin. This reconciles the growth story with the stated operating model; it does not assume expanding gross margins.

2027 delivers scale; 2028 tests the next platforms

Calendar 2027 / approximately FY2028

The core outlook is already substantial: about $18B total revenue, data-center growth above 60%, and custom revenue more than doubling. Current optical DSP demand, 1.6T adoption, scale-out switching and XPU-attach programs do most of the heavy lifting. Our base model is $18.5B revenue and $6.78 adjusted EPS. It is close to the current vendor consensus of $18.21B / $6.78. [2] [7]

Three custom drivers were identified in May: growth in existing programs, more than ten attach programs reaching higher volumes, and a new Tier-1 XPU entering volume production. Management described the contributions to growth as roughly one-third each at that time; that is a dated management framework, not a permanent mix forecast. In August the company reiterated progress without publicly identifying every customer or providing audited per-program revenue.

Scale-up optics adds a new source of sales. May’s broad category was roughly $300M, including a ~$150M Celestial component. August described a meaningful increase without giving a replacement dollar figure. Our $600M base assumption for FY2028 is an explicit estimate, not management guidance. NPO and CPO can coexist, as can copper and optics. A single architecture does not have to win everywhere for Marvell to grow. [2] [3]

Calendar 2028 / approximately FY2029

This is where the custom programs and optical scale-up products could become disproportionately important. Management previously framed custom revenue above $10B in FY2029, and the August call indicated additional upside associated with the expanded Google relationship and other programs. A revised total-company FY2029 forecast was not provided on that call. Our base assumes $10.5B custom revenue within $25.5B total revenue. [2]

The investment question changes from “is AI demand strong?” to “how many of these programs become qualified, manufacturable, high-volume products on schedule?” Design activity is an early indicator. Tape-outs, qualification, manufacturing allocations and purchase orders are stronger evidence. Shipped revenue, margin and cash collections are the strongest evidence.

The September optical demonstrations support the next-generation roadmap—2nm products, 400G per lane and 102.4T CPO—but do not by themselves establish FY2029 sales. The August memory portfolio launch directly connects Bravera storage controllers, Structera CXL products and optical memory to inference workloads. These announcements support product relevance; the financial model still requires customer conversion. [8] [9]

What could make Marvell harder to replace?

The potential advantage is the combination of high-speed SerDes, packaging, memory interfaces, optical technology and large switch design, with the ability to integrate those elements into a customer’s system. A multi-product engagement can be more valuable than a standalone component win. However, Broadcom, NVIDIA, Credo, Astera Labs, specialist optical companies and hyperscalers’ internal teams all compete for parts of this architecture. NVIDIA is both a partner and a powerful platform owner. Its investment is evidence of alignment, not a guarantee of Marvell’s future share.

Variant thesis: revenue per AI system can rise as connectivity and memory become limiting factors, allowing Marvell to grow faster than aggregate capex. What is already consensus: strong AI growth and a large FY2029 custom ramp. The genuinely differentiated question is whether new attachment opportunities and scale-up optics exceed current expectations while protecting earnings per share. The AI label alone is not an edge.

Google: a large opportunity with conditional economics

The August 19 8-K identifies Google and a July 29 commercial agreement covering products that attach to the TPU ecosystem. Google received warrants for up to 58.97M shares at $206.58. About 1.36M vest over time; the remainder vest in 240 tranches tied to $500M of qualifying purchases each. Purchases are discretionary. [5] [6]

Management says revenue covered by the arrangement through FY2028 is already included in its outlook. The larger incremental opportunity is FY2029 and beyond. Buying the stock on “$18B guidance plus $120B Google” counts the same opportunity incorrectly and ignores its timing. A win around the TPU ecosystem also does not prove Marvell replaces the designer of the primary TPU. [2] [5]

The maximum warrant face share count is about 6.7% of August common shares, but that is a conditional lifetime amount, not immediate dilution. Diluted EPS depends on vesting, the stock price and treasury-stock accounting; exercise also brings cash. Our share-count scenarios make an aggregate allowance for dilution rather than assuming all warrants vest now. The final fair value and revenue-accounting effects of this subsequent event were not quantified in the August 1 statements. [1] [6]

The 10-Q: the economics behind the growth

1. Cash generation is improving, but capacity consumes cash

First-half CFO was $1.2443B, up from $794.5M. After $282.4M of PP&E purchases, conventional FCF was $961.9M, versus $628.2M a year earlier. Growth was therefore funded partly through internally generated cash. The company still expects about $1B of supplier capacity prepayments during FY2027. These are applied against future purchases, but they are real current cash uses; adding them back as if they cost nothing would exaggerate shareholder cash generation. [1] [2]

The filing also reports $5M of technology-license purchases in investing and $56.6M of technology-license obligation payments in financing. Deducting both gives a stricter first-half cash measure of $900.3M. This is our analytical measure, distinct from standard CFO-minus-capex FCF. Non-current capacity prepayments increased from $278.8M to $487M; that balance change is not the same as gross prepayments made during the period.

2. Adjusted earnings are useful, but compensation is an economic cost

Q2 adjusted net income was $865.9M versus $308M GAAP. The reconciliation includes $326.2M of stock compensation, $214.9M of acquired-intangible amortization, and $52M net acquisition-contingent-consideration/hedge expense, along with smaller items and tax effects. Stock compensation more than doubled year over year and was about 11.9% of revenue. [1] [10]

First-half stock compensation was $533.8M. FCF minus that amount is $428.1M, an economic-cost sensitivity—not GAAP cash flow and not an additional cash payment. Do not subtract both an assumed full compensation cost and the full associated future dilution as if they were independent expenses. Our equity model uses adjusted earnings and explicit diluted shares, with this cash cross-check to avoid treating compensation as free.

3. Share growth changes the payoff

Q2 diluted shares were 921.2M, versus 870.4M a year earlier, an increase of 5.8%. That denominator already includes NVIDIA’s convertible preferred investment on an as-converted basis. NVIDIA invested $2B; the preferred can convert into approximately 21.8M common shares. Adding those shares to 921.2M again would double-count them. [1]

Celestial’s initial consideration included 24.5M common shares. Its contingent arrangement can require up to roughly $233M cash and 22.4M additional shares if milestones are achieved. The recorded contingent liability was $749.5M, but that is a fair-value accounting balance, not the maximum settlement. Rising stock prices and changing milestone probabilities can create GAAP expense even while operational prospects improve. Our base share assumption increases to 940M in FY2028 and 955M in FY2029; actual dilution remains uncertain.

4. The balance sheet can fund growth; supplier obligations still matter

Cash was $3.9328B against $4.9629B of carrying-value debt, or about $1.03B net debt. There were no contractual debt maturities through FY2028, and a $1.5B revolver was undrawn. Near-term leverage is not the central concern. But the increase in cash was aided by the NVIDIA investment and new borrowing, not solely earnings. [1]

The filing lists $8.5189B of unconditional foundry/test/assembly purchase commitments, $666.3M of technology/service commitments and $351.6M of capex commitments. These cover multiple years; they are not all due next quarter and are not interchangeable with debt. The economic asymmetry matters: upstream capacity can be committed while downstream purchase orders can be cancelled or rescheduled. A forecast miss could leave Marvell paying for supply it cannot monetize.

5. Concentration is real, but the categories must be read correctly

One distributor accounted for 44% of Q2 revenue and one direct customer for 16%; four customers represented 72% of gross receivables. The distributor percentage is not proof that one hyperscaler accounts for 44% of sales. The filing does not provide complete ultimate-end-customer allocation. Likewise, 42% of Q2 shipments went to China, but the company says most China shipments relate to non-China customers’ manufacturing operations. Shipping destination is not end-demand nationality. [1]

Receivables factoring was $412M in the first half, down from $526.8M a year earlier. It remains a cash-timing consideration, but the improvement in CFO cannot simply be attributed to greater factoring volume. Subtracting all factored sales from CFO would also be wrong because the counterfactual collection timing is unknown.

6. Acquisition value is still awaiting commercialization

Celestial and XConn revenue and earnings were not material in the reported period. Goodwill and acquired intangibles totaled about $16.22B. Investors are paying for future platform economics, not an acquired stream of already-large profits. Celestial’s booked IPR&D is not yet amortized; amortization begins when projects reach the relevant production stage. Future GAAP/adjusted differences can therefore change as products launch. [1]

The 2027–2028 earnings model

Every number in this table is an analyst scenario assumption or its calculated result. The company’s ~$18B FY2028 outlook anchors the base case. It does not disclose the detailed revenue buckets or our FY2029 totals. Amounts in billions, except shares and EPS.

¹ Other data-center products include scale-out/scale-across connectivity, switching, storage and related products, excluding all revenue allocated to custom and scale-up optics. The standalone scale-up optics row includes NPO/CPO-related silicon and analog content; it is not Celestial revenue alone. Custom CXL/NIC sales appear only under custom. There is no separate addition for a “Google deal.”

For the FY2029 base case: [($25.5B × 57%) − $3.95B − $0.15B] × 87% ÷ 955M shares = $9.51 EPS. Gross margin falls as custom becomes larger, but expense discipline increases operating margin to 41.5%. That exceeds the currently stated 38–40% long-term band and requires additional operating leverage; it is not yet company guidance.

The bear case still grows from $16.5B to $21B. The stock can lose money because program timing, product economics and the valuation multiple disappoint, even without an absolute revenue collapse. The bull case requires faster-than-guided FY2028 growth, ~$13B FY2029 custom sales, much larger scale-up optics and a 44.3% adjusted operating margin. It is a demanding upside case.

The base model is near consensus in FY2028 but about 8.4% below consensus FY2029 EPS. Vendor consensus is context, not proof. The feed does not fully document the EPS adjustment basis, and its EBIT/EBITDA series does not reconcile cleanly with management’s operating-margin outlook. I used the raw filing and CFO guidance to build margins instead of copying those fields. FMP’s January 31 dates are normalized fiscal-year labels, not exact fiscal closing dates. [7]

What the stock could be worth in 12 months

At $268.08, current vendor consensus implies 63.7× FY2027, 39.5× FY2028 and 25.8× FY2029 EPS. The decline in those multiples assumes the earnings arrive. The last figure is not today’s ordinary one-year-forward P/E; FY2029 is a much later earnings period. The vendor’s analyst target consensus is $278.38, or only 3.8% above the price, with a wide $155–400 range. Neither a consensus average nor an analyst high is a promised return. [7] [11]

Valuation date: October 1, 2027. At that point, the next twelve months span roughly four months of FY2028 and eight of FY2029. I use one-third FY2028 EPS plus two-thirds FY2029 EPS as a next-twelve-month proxy. Annual pro-rating ignores quarter-by-quarter ramp timing; a back-loaded ramp would lower the relevant earnings. This approach prevents automatically treating all of FY2029 as earnings available within the forecast horizon.

The 35× base multiple assumes a still-premium business but some valuation normalization as Marvell scales. The 28× bear multiple reflects growth disappointment and greater customer/mix risk. The 40× bull multiple requires broad program success and confidence in growth beyond FY2029. These are judgmental multiples, not outputs from a historical regression or a proven peer valuation floor.

For perspective, today’s consensus-based NTM proxy is about $5.92, implying ~45.3×. Our base case increases the later NTM proxy to $8.60, but reducing the multiple to 35× leaves only 12.3% stock upside. Fast earnings growth and modest stock returns can coexist.

Required earnings: at 35×, today’s price needs $7.66 of forward EPS. A $400 stock needs $11.43 at 35× or $10 at 40×. Reclaiming the $329.88 high requires $9.43 at 35×. These are the earnings hurdles against which new forecasts should be measured.

What moves the model most: a five-point change in the base P/E changes value by about $43 per share. An extra one percentage point of FY2029 operating margin adds about $0.23 to FY2029 EPS and roughly $5.42 to the main 12-month value, holding everything else fixed. A 5% increase in both forecast share counts reduces modeled value by 4.8%. This is why a precise-looking price target should be read as a conditional range of outcomes.

Risk/reward: the bull gain is $151.08 per share against a $97.67 bear loss, or 1.55×. The base gain versus bear loss is only 0.34×. An illustrative 25% bear / 50% base / 25% bull weighting gives $297.87, or +11.1%; these probabilities are subjective, not statistically estimated. The scenarios are not bounds—an extreme macro, customer or supply shock can produce worse outcomes.

Cash-flow cross-check: trailing conventional FCF is $1.7301B, less than 1% of the $234.8B vendor common-equity capitalization. That does not disprove the growth thesis; it shows how much future cash generation is being paid for. At the base target, 955M modeled shares imply roughly $287B equity value. Even eventual $6–8B annual FCF would be only a 2.1–2.8% yield. The base valuation therefore still requires durable growth. Preferred and warrant dilution mean vendor common market cap is not a fully diluted equity value. [1] [7]

The chart confirms momentum, not a cheap entry

At the October 1 close, RSI(14) is 63.4; the stock is 18.9% above its calculated 50-day average ($225.47), 61.5% above its 200-day average ($165.97), and 18.7% below its $329.88 yearly high. The 52-week low is $70.69. A price below the high can still be far above trend.

Recent price references: $248–257 is the late-September pullback/closing cluster; $236–245 is the earlier breakout area; roughly $220–230 contains multiple September pivots and the 50-day average. The immediate overhead reference is $267–270, followed by the prior $329.88 high. These are observed zones for monitoring, not guaranteed support, a complete resistance map or trade orders.

Under the current literal four-path framework, there is no fresh-entry pass: Path A fails the 200-day extension test; Path B fails the ≥20% consolidated sequential revenue requirement (Q2 was 13.3%, not the data-center growth rate); Path C fails the ≥35% drawdown condition; Path D fails both moving-average extension limits. Disclosed communications lumpiness warrants attention as well. This is an execution constraint, not a finding that the business thesis has broken.

The earnings reaction illustrates expectations risk: the stock closed at $241.45 on August 27 and $216.62 on August 28, a 10.3% decline despite the higher outlook. Price history alone does not establish the cause. Subsequent recovery also means that the old post-earnings entry price is no longer available. [7]

Computed from 273 daily close observations starting September 2, 2025; chart window October 2, 2025–October 1, 2026. Wilder smoothing for RSI, arithmetic rolling means for SMAs. The 200-day line starts only after 200 observations; no earlier data was invented. Vendor quote fields give slightly older averages ($224.33/$165.05); this report uses averages recomputed through the October 1 close. Small intraday open/volume differences between vendor endpoints do not affect the closing-price calculations.

The evidence that changes the valuation

Do not pre-count an Investor Day raise. October 6 is confirmed; its financial conclusions remain unknown. New company forecasts should replace this model where supported. I have not scheduled monitoring or a follow-up.

Thesis-break evidence

• Named or clearly attributable major customer/program loss, a material ramp delay, or reduced FY2028/FY2029 demand expectations.
• Gross-margin pressure beyond the planned mix effect, accompanied by failure to expand operating margins.
• Supplier commitments, inventory and receivables growing without corresponding revenue and cash collection.
• Warrant, compensation or acquisition dilution materially outpacing the modeled 940–970M share counts.
• Customer insourcing, competitor displacement, or an architectural shift that reduces Marvell content faster than new products replace it.

Macro and policy transmission: higher discount rates can lower the P/E while AI capex remains strong. Hyperscaler spending cuts, power/permit constraints or restrictions on advanced chips can delay shipments. TSMC dependence is a real production concentration. The October 1 VIX close was 16.39; that does not eliminate company-specific drawdown risk. No unverified news-feed claim was converted into a confirmed supplier order or a causal explanation for the day’s share move. [1] [7]

Sources, checks and remaining uncertainty

1. [1] Latest 10-Q · quarter ended August 1, filed August 28 Primary page

   Statements pp. 2–7; notes pp. 8–27; MD&A pp. 28–37; demand/customer/order/supply risks pp. 41–49. Key sections reviewed; complete filing available from the SEC link.

2. [2] Q2 FY2027 full call · August 27 Primary page

   Full prepared remarks and Q&A reviewed; refreshed FMP body matched local copy.

3. [3] Q1 FY2027 full call · May 27 Primary page

   Full prepared remarks and Q&A reviewed.

4. [4] Q4 FY2026 call · March 5 Primary page

   Selected historical guidance, product and financial passages reviewed.

5. [5] Google commercial agreement 8-K · August 19 Primary page

   Parties, covered products and discretionary purchase-based vesting.

6. [6] Google warrant agreement Primary page

   Detailed contract reference; vesting is conditional.

7. [7] FMP snapshots · October 1

   Analysis uses October 1 vendor snapshots for quotes, daily prices, estimates, statements, transcripts, news, filings, the earnings calendar and VIX. The raw research archive is retained privately. Vendor fields are not all independently audited.

8. [8] 2nm optical demonstrations · September 21 Primary page

   Company-primary product roadmap; demonstrations do not prove volume shipments.

9. [9] Agentic inference memory portfolio · August 4 Primary page

   Company-primary product context.

10. [10] Q2 earnings release and reconciliation Primary page

    GAAP/non-GAAP reconciliation and Q3 guidance.

11. [11] FMP analyst price-target snapshot

    $278.38 mean, $275 median, $155 low, $400 high. Aggregation may include targets with different publication dates.

12. [12] October 6 Investor Day · official calendar Primary page

    Confirmed upcoming event as of this report; future guidance unknown.

Download scenario model and research metrics

Quality checks: read-only DuckDB manifest agrees with the call dates; original lake files were preserved. The raw filing’s guessed fiscal label was rejected. Current quarter vendor statements were checked against the filing. The latest call was compared with a fresh transcript. Model revenue buckets sum once, EPS includes taxes and diluted shares, and the price horizon uses an explicit earnings convention.

Excluded interpretations: $120B is not backlog; the same Google revenue is not added twice; $1B run-rate targets are not treated as $1B full-year sales; distributor and shipment-destination percentages are not end-customer exposure; supplier commitments are not all current debt; no automatically constructed vendor margin forecast is treated as company guidance. The news feed included an unrelated Canadian Marvel Biosciences ticker collision, which was excluded.

Open questions: the updated FY2029 roadmap; customer-level economics and program shares; Google warrant valuation/revenue effects; Celestial/NPO production yields; longer-term dilution; exact quarter-by-quarter ramp timing; and the issuer-confirmed next earnings date. These can materially change the scenarios. No validated statistical alpha or probability model is claimed.