Marvell Technology
A chip designer whose parts link the servers in AI data centers, and which designs custom AI chips for the largest cloud companies.
The scores
Growth3/4 ↑
Revenue grew 42.1% in FY26 and 37% in Q2 FY27, but FY24 fell, FY25 grew 4.7% and ten customers are 82% of sales.
Business quality3/4 →
Rivals name it and an outside firm ranks it a leader in optics, but custom work is re-won each generation and its share may fall.
Financial position3/4 →
About $1.07B of net debt, nothing due before 2028 and cash from operations covering the plan; but $8.52B is owed to suppliers through FY2030.
Execution record3/4 ↑
Beat the guide in seven of eight quarters and raised FY28 from $13B to $18B, but FY24 and FY25 revenue fell or stalled.
Per share capture2/4 ↓
Stock pay doubled to 11.9% of revenue, and shares owed to Google, NVIDIA and Celestial's sellers equal about 11% of the company.
Potential2/4 ↑
A large prize, but FY28 and FY29 rest on custom programs nobody is obliged to buy and on cloud spending Marvell does not control.
Risk profile2/4 →
One distributor is 44% of revenue, 42% ships to China under export rules, and each custom program is won or lost in one step.
Valuation2/4 ↓
About 40.3 times FY28 consensus, 49 with stock pay, pays for FY28's $18B and much of FY29; little is left if FY29 half happens.
Evidence quality3/4 →
Results, outlook and warrant are filed and both markets measured outside; the custom pipeline and the data center split rest on management's word.
The Verdict
We rate Marvell Technology unattractive. It is a mixed business growing very fast: data center sales rose 46% in its latest quarter and management expects revenue to reach about $18B next year. The price is full. At about 40 times next year's expected profit, buyers are paying for that plan and much of the year after, while Broadcom holds most custom chip work and every new chip generation must be won again. Part of the success is also promised away, in shares owed to Google, NVIDIA and the sellers of a start-up it bought. The growth is real, but the price leaves little room if it only half arrives.
The long caseCloud companies are building their own AI chips and faster links between them, and Marvell designs both; it must keep winning each new generation against Broadcom.
The evidence under this - 6 claims
The FY28 figure is an outlook given on a call, not a formal guide, and it has been raised three times in nine months, so it moves with the order book rather than being fixed by contracts the filings show.
Broadcom, the largest custom chip designer, trades at 45.0 times trailing earnings against Marvell's 89.0 in the pack's filed cohort (C16), which suggests the market prices Marvell's growth well above Broadcom's.
Marvell says custom will more than double in FY28 (C05), and NVIDIA and Google have since signed with it (C11, C06), after Counterpoint's January estimate.
The company
An AI data center ties thousands of chips together, and the links between them decide how fast the whole cluster runs. Marvell sells the chips inside those links: signal processors that clean up data sent over fiber at very high speeds, and chips for copper cables and switches.
Its second business is designing custom AI chips for cloud companies that want an alternative to buying NVIDIA's. Optical links are roughly half of data center sales and custom chips about a quarter.
The evidence under this - 3 claims
Neither share is in the financial statements, which report data center as one line, so the split cannot be checked against a filed figure.
How it makes money
Marvell is paid for each chip it ships. It designs them and pays TSMC and other factories to make them, so most of its spending is on engineers.
On the official accounts it kept 53 cents of each sales dollar as gross profit in Q2 FY27 and 17 cents as operating profit. Excluding stock pay and the cost of past acquisitions those figures are 59 and 37 cents. A handful of cloud customers and one distributor buy most of it.
The evidence under this - 4 claims
Where the money comes from, stream by stream
| Stream | Share of sales (FY 2026) | How it earns it |
|---|---|---|
| Optical interconnect | about 37% | Signal processors, amplifiers and drivers for the optical modules that link servers and data centers. |
| Custom silicon | about 19% | AI accelerators and attach chips designed to one cloud company's specification. |
| Other data center | about 19% | Switch chips, copper cable chips, retimers, memory expansion and storage controllers. |
| Communications and other | 25.6% | Chips for carrier networks, enterprise networking and consumer devices. |
Who runs it
Matt Murphy has run Marvell since 2016 and turned it from a storage and networking chip maker into a data center company. Dan Durn, formerly Adobe's finance chief, became chief financial officer this year.
Directors and officers own less than 1% of the company, and insiders sold shares and bought none in the past year, mostly under plans set in advance. The yearly bonus pays on revenue and adjusted margins.
The evidence under this - 3 claims
Why it wins, and who could beat it
Why customers come back
Few companies can design the very fast circuits that move data between chips, and Marvell reuses that work in every new product. Rivals Credo and Astera Labs both name it as a main competitor, and an outside firm ranks it a leader in optics linking data centers.
In 2026 NVIDIA invested in it and Google tied a deal to its custom chips. But a cloud company re-picks its partner each chip generation, and one firm expects Marvell's share of custom design work to fall to 8% in 2027.
The evidence under this - 5 claims
Marvell says custom will more than double in FY28 (C05), and NVIDIA and Google have since signed with it (C11, C06), after Counterpoint's January estimate.
The threat
Broadcom does the same custom work and holds about 60% of it, including Google's future chips through 2031. Taiwanese designers are winning programs too, among them Amazon's next Trainium chip.
In optics, newer designs drive the laser directly and need no signal processor, which outside firms expect to cut into Marvell's largest product from 2027. And cloud companies can do more of the design themselves.
The evidence under this - 4 claims
Marvell says custom will more than double in FY28 (C05), and NVIDIA and Google have since signed with it (C11, C06), after Counterpoint's January estimate.
Dell'Oro expects pluggable modules to coexist with co-packaged optics through 2030 (C23), and Marvell sells linear-drive parts and co-packaged optics itself through Celestial AI (C10).
The rivals: AVGO, NVDA, CRDO, ALAB, CIEN, NOK, MTSI, SMTC, RMBS, MCHP, CSCO, and 2 more that are private, in-house or listed abroad.
Who else does this, rival by rival
| Ticker | Rival | Where it competes | How far along | Priced against it |
|---|---|---|---|---|
| AVGO | BroadcomPriced against it: Yes | Data center: connectivity and custom silicon | On sale | Yes |
| Alchip, Global Unichip and MediaTekPriced against it: Listed abroad | Data center: connectivity and custom silicon | On sale | Listed abroad | |
| In-house | The cloud companies' own chip teamsPriced against it: In-house | Data center: connectivity and custom silicon | On sale | In-house |
| NVDA · AVGO | Linear optics and co-packaged optics, led by NVIDIA and BroadcomPriced against it: Yes | Data center: connectivity and custom silicon | Announced | Yes |
| CRDO · ALAB | Credo and Astera LabsPriced against it: Yes | Data center: connectivity and custom silicon | On sale | Yes |
| CIEN · NOK | Ciena and NokiaPriced against it: One segment only | Data center: connectivity and custom silicon | On sale | One segment only |
| MTSI | MACOM Technology Solutions HoldingsPriced against it: Yes | - | On sale | Yes |
| SMTC | SemtechPriced against it: Shown, not counted | - | On sale | Shown, not counted |
| RMBS | RambusPriced against it: One segment only | - | On sale | One segment only |
| MCHP | Microchip TechnologyPriced against it: One segment only | - | On sale | One segment only |
| CSCO | Cisco SystemsPriced against it: One segment only | - | On sale | One segment only |
How big it could get
How big this could get
Both markets are growing several times over. An outside firm expects shipments of custom AI chips to triple from 2024 to 2027 and pass 15 million a year in 2028. Almost 20 million fast optical modules, each needing one signal processor, shipped in Q1 2026 alone.
The money behind it is cloud companies' building budgets, which TrendForce puts above $886.7B this year. No outside source prices the chips, so the market cannot be put in dollars.
The evidence under this - 4 claims
Marvell says custom will more than double in FY28 (C05), and NVIDIA and Google have since signed with it (C11, C06), after Counterpoint's January estimate.
LightCounting expects linear optics to cut into processor sales from 2027 (C20).
NVLink dominates scale-up today, and NVIDIA leads co-packaged optics, by Dell'Oro's own account.
TrendForce notes part of the 2026 increase is rising component prices rather than more equipment.
Who has signed
Google holds a warrant that hands it Marvell shares as it buys up to about $120B of custom chips through January 2033, but it is not obliged to buy any. NVIDIA invested $2.0B and agreed to work together, with no volume named. No customer has committed to the $18B FY28 plan.
The evidence under this - 3 claims
Counterpoint expects Marvell's share of custom AI chip design to fall to 8% in 2027 as Alchip enters AWS's Trainium supply chain (C17), and Broadcom holds a long-term agreement for Google's future TPUs through up to 2031 (C18).
Where the growth is coming from, business by business
Where the growth is coming from
| Business | Share of revenue | Growth: last quarter | Last half-year | Last year | Share in 3 years at its half-year rate |
|---|---|---|---|---|---|
| Data center: connectivity and custom silicon | 76% | +27.2% | +31.8% | +48.2% | 76% |
| Rest of the company | 24% | +28.7% | +31.3% | +31.5% | 24% |
| Whole company | 100% | +27.6% | +31.7% | +43.6% | 100% |
On the filed split through 2026-05-02, if each business keeps its last half-year's growth, the company grows about 32% next year and 32% in year 3, against 30.6% over the last twelve months. A scenario, not a forecast: no business holds its rate.
How we read it: The score believes the data center business, 79% of revenue and growing 46%: it is the case, and the slower communications line is a fifth of sales and falling.
The results
The numbers
Sales grew 31% over the twelve months to Aug 2026, slower than the 42% in the year to Jan 2026.
Operating margin is 17%, up from −10% in the year to Feb 2024.
Running costs take 36¢ of each sales dollar, down from 54¢ in the year to Jan 2022.
Shareholders' money earned 17% a year, up from −3.5% in the year to Jan 2022.
It turned 18% of sales into free cash, $1.73B over the last twelve months.
It owes $1.03B more than the cash it holds, about 0.5 years of the cash its operations bring in.
The share count rose 16% in 4.5 years, and pay in shares costs 8.8% of sales.
All the numbers
Filed figures from the company's 10-K and 10-Q, read 2026-10-05. Fiscal years end in January.
| Jan 2022 | Jan 2023 | Feb 2024 | Feb 2025 | Jan 2026 | 12 mo to Aug 2026 | |
|---|---|---|---|---|---|---|
| Size and growth | ||||||
| Sales | $4.46B | $5.92B | $5.51B | $5.77B | $8.19B | $9.45B |
| Sales growth | +50% | +33% | −7.0% | +4.7% | +42% | +31% |
| Profit | ||||||
| Gross margin | 46% | 50% | 42% | 41% | 51% | 52% |
| Operating margin | −7.8% | 4.0% | −10% | −12% | 16% | 17% |
| Net margin | −9.4% | −2.8% | −17% | −15% | 33% | 28% |
| What growth costs, per sales dollar | ||||||
| Research and development | 32% | 30% | 34% | 34% | 25% | 26% |
| Sales and marketing | 21% | 14% | - | - | - | - |
| Pay in shares | 10% | 9.3% | 11% | 10% | 7.2% | 8.8% |
| Capital spending | 3.8% | 3.5% | 6.1% | 4.9% | 4.3% | 5.0% |
| Cash | ||||||
| Cash from operations | $819M | $1.29B | $1.37B | $1.68B | $1.75B | $2.20B |
| Free cash flow | $650M | $1.08B | $1.03B | $1.40B | $1.40B | $1.73B |
| Balance sheet | ||||||
| Cash and short-term investments | $614M | $911M | $951M | $948M | $2.64B | $3.93B |
| Debt | $4.61B | $5.08B | $4.17B | $4.06B | $4.97B | $4.96B |
| Net cash | −$4.00B | −$4.17B | −$3.22B | −$3.12B | −$2.33B | −$1.03B |
| Returns and shares | ||||||
| Return on capital | - | - | - | - | 6.3% | - |
| Shares (diluted) | 797M | 851M | 861M | 866M | 870M | 921M |
What growth costs
Growth is widening the margin: costs rise more slowly than sales, and management expects that to continue in FY28. But stock pay doubled in a year to 11.9% of revenue, so part of the improvement is paid in shares rather than cash. Marvell is also paying suppliers about $1B ahead this year to reserve factory space.
The evidence under this - 3 claims
Stock pay rose faster than revenue, 112% against 37%, so part of the widening margin management reports is paid in shares rather than cash.
The $8.52B of supplier commitments (C09) are owed whether or not the demand arrives.
The last report (27 Aug)
Quarter ended August 2026, filed figures.
Revenue in Q2 FY27 rose 37% from a year earlier to $2.74B, and data center sales rose 46%. Management guided the next quarter above $3.15B, up more than 50%, and raised its FY28 outlook to about $18B.
The shares still fell 10.3% the next day. Analysts' questions turned on timing: most of the revenue tied to Google's warrant arrives after FY28, and the company left the detail for its Investor Day.
What it did to the thesisIt proved the business is accelerating and that management keeps raising its plan. It left open the part the price depends on: whether custom revenue after next year arrives on time, at the scale the Google deal suggests.
The quarter in detail
Sales growth from the previous quarter, filed figures.
Earnings per share $0.94 against $0.93 expected; $0.33 as filed, the gap being what the company adds back.
The evidence under this - 3 claims
The FY28 figure is an outlook given on a call, not a formal guide, and it has been raised three times in nine months, so it moves with the order book rather than being fixed by contracts the filings show.
What management did with the money
Marvell sold its car networking business and spent about $4B on two start-ups, Celestial AI for optics and XConn for memory links, partly in shares. It took $2.0B from NVIDIA and bought back $400M of its own stock in the first half of FY27.
Insiders in 2026Q3: 18 purchases and 43 sales, from their Form 4 filings.
The evidence under this - 3 claims
The price, and where we disagree
In short. Full. The price pays for next year's plan and much of the year after, leaving little if growth only half arrives.
The shares more than tripled in a year. Sales per share grew about a quarter of that rise, and the rest came from buyers paying more for each dollar of sales.
The next test comes at once. Marvell holds an Investor Day on 6 October 2026 to set its long-term plan and size the custom business to the end of the decade. The Q3 FY27 results follow around 1 December.
The evidence under this - 4 claims
The FY28 figure is an outlook given on a call, not a formal guide, and it has been raised three times in nine months, so it moves with the order book rather than being fixed by contracts the filings show.
Counterpoint expects Marvell's share of custom AI chip design to fall to 8% in 2027 as Alchip enters AWS's Trainium supply chain (C17), and Broadcom holds a long-term agreement for Google's future TPUs through up to 2031 (C18).
How we read the multiple
A buyer pays for profit that has not been earned yet.
1. Analysts expect earnings of $6.76 a share in FY28, excluding stock pay and acquisition costs. 2. Today's price is about 40 times that. 3. Stock pay is a real cost, about $1.23 a share at this year's rate; counting it, the price is about 49 times. 4. Even on FY29's forecast, 44% higher again, the price is about 26 times.
So FY28 must arrive in full and FY29 must grow fast too before the price looks undemanding.
Against its own history. Buyers pay far more than the company's own normal for each dollar of sales, because they expect growth faster than it has ever kept up.
The evidence under this - 4 claims
Broadcom, the largest custom chip designer, trades at 45.0 times trailing earnings against Marvell's 89.0 in the pack's filed cohort (C16), which suggests the market prices Marvell's growth well above Broadcom's.
Priced against its rivals
The competitors its own annual report names, each confirmed against the rival's filing. Measured on price for each $1 of yearly profit, cheapest first:
| Ticker | Company | price for each $1 of yearly profit |
|---|---|---|
| AVGO | Broadcom | 45.0× |
| CRDO | Credo Technology Group Holding | 72.0× |
| MRVL | Marvell Technology | 89.0× |
| MTSI | MACOM Technology Solutions Holdings | 102.6× |
| ALAB | Astera Labs | 167.0× |
| SMTC | Semtech not in the median | 116.3× |
The middle of the other 4 is 87.3. Semtech Corporation competes with it in one product line inside a much bigger company, so its price says little about this business and it is not counted in the middle. Marvell Technology, Inc. is 3rd cheapest of 5.
Rivals' figures come from a data provider, so read the spread of the group rather than any single pair.
It sits in the middle of its listed rivals, which all trade high together, while Broadcom, the largest custom designer, costs about half as much.
Also checked and set aside: Advanced Micro Devices (a substitute, not the same product); Intel (a substitute, not the same product).
Private or foreign rivals that cannot be priced: Alchip Technologies, Global Unichip Corporation, MediaTek Inc., Realtek, Phison Electronics, Montage Technology, Socionext Inc., Ayar Labs, Lightmatter, Ranovus, NXP Semiconductors, Qualcomm, Silicon Motion, customers' in-house silicon teams.
Where we disagree with the market
An AI winner whose custom chip business is just starting, with a Google deal that could add a Marvell-sized business every year by FY29.
Next year is close to priced in, and the year after rests on programs re-won each generation against Broadcom, with part of the upside owed to Google.
The Google warrant pays Google in shares only when it buys, and obliges it to buy nothing, while Broadcom has a signed agreement for Google's next flagship chips. Each custom program is won one generation at a time, so FY29 is a hope, not signed orders.
Against us: the plan has risen from $13B to $18B in nine months, the quarters keep beating, and NVIDIA and Google have both put money behind Marvell's custom work.
What analysts say, and the evidence
The evidence under this - 6 claims
Marvell says custom will more than double in FY28 (C05), and NVIDIA and Google have since signed with it (C11, C06), after Counterpoint's January estimate.
The FY28 figure is an outlook given on a call, not a formal guide, and it has been raised three times in nine months, so it moves with the order book rather than being fixed by contracts the filings show.
What could go wrong, and how we would know
The risks
- A lost chip generation A cloud customer can move its next custom chip to Broadcom or a Taiwanese designer, as an outside firm says Amazon is doing.
- Optics without processors Newer optical designs drive the laser directly; outside firms expect them to cut into Marvell's largest product from 2027.
- Ordering ahead of need Optical demand ran 30% above supply in April; past cycles saw sales drop once buyers stopped ordering twice.
- Few buyers, much China One distributor is 44% of revenue and 42% ships to China, where export rules already restrict some products.
- Promised to suppliers Marvell owes factories $8.52B through FY2030 whether or not the demand arrives.
- Shares given away Google, NVIDIA and Celestial's sellers can receive about 11% more shares as the business succeeds.
The evidence under this - 7 claims
Marvell says custom will more than double in FY28 (C05), and NVIDIA and Google have since signed with it (C11, C06), after Counterpoint's January estimate.
Dell'Oro expects pluggable modules to coexist with co-packaged optics through 2030 (C23), and Marvell sells linear-drive parts and co-packaged optics itself through Celestial AI (C10).
Cignal AI raised its 1.6T forecast twice after January (C19), so demand had not slowed by June.
What we still do not know
- Who designs Amazon's next Trainium chip, and how much of it Marvell keeps.
- How much of the revenue tied to Google's warrant is new, rather than already inside the FY28 plan.
- Marvell's share of the newest optical signal processors against Broadcom's, which no free source publishes.
The evidence under this - 4 claims
Marvell says custom will more than double in FY28 (C05), and NVIDIA and Google have since signed with it (C11, C06), after Counterpoint's January estimate.
Counterpoint expects Marvell's share of custom AI chip design to fall to 8% in 2027 as Alchip enters AWS's Trainium supply chain (C17), and Broadcom holds a long-term agreement for Google's future TPUs through up to 2031 (C18).
LightCounting expects linear optics to cut into processor sales from 2027 (C20).
What would change our view
- Unattractive→NeutralA cheaper price on an unchanged plan
It needs:
- ○The shares fall to 28 times analysts' FY28 earnings forecast, from about 40 today.about $188.46 today, 31% below the last close
- Unattractive→NeutralThe next ramp is named and dated
It needs all of:
- ○Marvell sets an FY29 revenue target above analysts' $26.25B, naming the custom programs behind it.watched as it happens
- ○The FY28 revenue outlook of about $18B holds or rises after the Q3 FY27 results.checked at the next results, due by 6 Oct
We would also be wrong if
- Marvell cuts its FY28 revenue outlook below $16.5B, the level it gave in May.
- Custom chip revenue fails to more than double in FY28.
- A cloud customer moves the next generation of a Marvell custom chip to another designer.
What is coming
- 6 October 2026 Investor Day: The long-term model and the custom business's path through the end of the decade: whether FY29 is sized above the $26.25B consensus.
Every dated event ahead
- 6 October 2026 Investor Day: The long-term model and the custom business's path through the end of the decade: whether FY29 is sized above the $26.25B consensus.
- 1 December 2026 Q3 FY27 results: Revenue against the $3.15B guide, and whether the FY28 outlook of about $18B is held or raised.
- 5 March 2027 Q4 FY27 results: Whether the margin excluding stock pay enters the 38% to 40% range, and the first FY28 guide.
Since we wrote this
- 8 OctNewsMarvell's investor day lifted its FY28 revenue outlook to about $20B, from $18B. View heldInvestor Day, 6 October 2026.
- 5 OctThesis writtenUnattractive, medium convictionWritten the day before Investor Day, with the shares up 224.6% in a year and 17% below their June high.
Also checked 4 times with nothing new to act on.
Sources 25
Sources - 25
Every claim above cites one of these, or states the reasoning that derived it. The quote is frozen at the date it was read: a link can rot, and an agent fetch returns a reading of a page rather than the page. Backs counts the whole claim ledger, so it can name a claim this page does not surface.
- S01primarycompany filing2026-03-11products, end markets, the named competitors, customer concentration, the risk factors on in-house silicon and export rules, the auto Ethernet sale, FY26 statements and headcountbacksC01C02C08C30C35C36
- S02primarycompany filing2026-09-03Q2 FY27 statements, revenue by end market and by destination, customer and receivables concentration, debt, purchase commitments, the Celestial AI and XConn purchases, the NVIDIA preferred stock, stock pay and buybacksbacksC01C03C08C09C10C11C12C16
- S03primarycompany release2026-08-27Q2 FY27 results, the GAAP to non-GAAP reconciliation and Q3 FY27 guidancebacksC03C04C12
- S04primarytranscript2026-08-27Marvell Q2 fiscal 2027 earnings call transcript (FMP), held 27 August 2026the FY27 and FY28 revenue outlook, custom doubling in FY28, the Google warrant's timing, capacity prepayments, the margin target, scale-up optics and the Investor Day date
we now expect fiscal 2028 revenue of approximately $18 billion, up $1.5 billion from the $16.5 billion outlook we provided just 1 quarter ago
backsC04C05C27C29C31 - S05primarycompany filing2026-08-18the warrant's share count, exercise price, expiry and the revenue-based vesting schedulebacksC06
- S06primarycompany filing2026-05-13the optical and custom shares of data center revenue, directors and officers, ownership, the bonus and long-term pay designbacksC07C26C27
- S07primarycompany release2026-03-05FY26 revenue, GAAP and non-GAAP EPS, the record design wins and Q1 FY27 guidancebacksC02
- S08secondaryvendor data2026-10-05Research pack (FMP market data, FINRA, SEC filed figures), frozen as research-pack.jsonprice and range, the price move split, analyst estimates and the rating series, price target drift, short interest, insider trades, the earnings reactions, headcount and the cohort multiplesbacksC12C13C14C16C28C29C30C36
- S09primaryindustry data2026-01-26Counterpoint Research: AI server compute ASIC shipments to triple by 2027 Counterpoint Researchcustom AI chip shipments, design partner shares, and Alchip's entry into the AWS Trainium supply chain
Marvell’s share is estimated to dip to 8% in 2027 as competition intensifies, but we still see the shipments doubling between 2024 and 2027.
backsC17 - S10primaryindustry data2026-06-18high-speed optical module units and the 1.6T forecast
Almost 20 million modules shipped in the quarter as AI build-outs accelerated.
backsC19 - S11primaryindustry data2026-02-26LightCounting: AI capex flows down the supply chain to DSP vendors (PAM4 and coherent DSPs, February 2026) LightCountingoptical DSP chipset shipments and sales, and the linear-drive threat to the DSP
We expect that growth in sales of PAM4 chipsets will moderate in 2027-2031, as volume deployments of linear-drive solutions (LPO and CPO) will negatively impact DSP sales.
backsC19C20 - S12primaryindustry data2026-04-15LightCounting April 2026 market forecast: demand for optical connectivity continues to surprise LightCountingtransceiver market growth, the supply shortage and the double-ordering risk
The current demand is exceeding supply by 30%, but the shortages should be going away by the end of 2026.
backsC21 - S13primaryindustry data2026-04-20the optical transceiver market's size and the move toward linear optics
technology roadmaps are accelerating toward low-power linear pluggable optics (LPO) and silicon photonics integration, aiming to replace traditional high-power DSP-based architectures
backsC20 - S14primaryindustry data2026-08-03TrendForce: AI server shipments forecast raised to nearly 31% in 2026 as CSP capex surges TrendForcecloud capital spending for 2026 and 2027, AI server growth and AWS's use of its own chips
Looking to 2027, TrendForce estimates that the combined CapEx of the top nine CSPs will hit about $1.3 trillion, showing nearly 50% year-over-year growth.
backsC22 - S15primaryindustry data2026-07-28Dell'Oro Group: AI back-end switch sales to approach $1 trillion over the next five years Dell'Oro GroupAI network switch spending, the scale-up share, and Ethernet against NVLink
Scale-up is projected to account for more than half of the market by 2030.
backsC23 - S16primarycompany release2026-03-31the partnership and NVIDIA's $2 billion investment, in the counterparty's words
Marvell will provide custom XPUs and NVLink Fusion-compatible scale-up networking, while NVIDIA will provide the supporting technologies
backsC11 - S17primarycompany filing2026-04-06Broadcom's agreement to design and supply Google's future TPUs and rack networking
a Supply Assurance Agreement for Broadcom to supply networking and other components to be used in Google’s next-generation AI racks through up to 2031
backsC18 - S18primarycompany filing2026-06-15Credo Technology Form 10-K for the fiscal year ended 2 May 2026 Credo Technology Groupa rival naming Marvell among its three principal competitors
Our principal competitors with respect to our products include Broadcom Ltd. (Broadcom), Marvell Technology, Inc. (Marvell) and Astera Labs, Inc. (Astera)
backsC25 - S19primarycompany filing2026-02-20a rival in retimers and switches naming Marvell
Our principal competitors include Broadcom, Inc., Credo Technology Group Holding Ltd., Marvell Technology, Inc., Microchip Technology Inc., Montage Technology
backsC25 - S20primaryother2025-10-13Open Compute Project: Introducing ESUN, advancing Ethernet for scale-up AI infrastructure Open Compute Project Foundationthe founders of the Ethernet scale-up standards effort
With the initiation of this workstream by AMD, Arista, ARM, Broadcom, Cisco, HPE Networking, Marvell, Meta, Microsoft, NVIDIA, OpenAI and Oracle
backsC24 - S21primaryindustry data2026-01-20the share of AI servers built on custom chips against GPUs
The share of ASIC-based AI servers is expected to reach 27.8% by 2026, the highest since 2023, as North American firms like Google and Meta expand their own ASIC efforts.
backsC33 - S22primaryindustry data2026-01-07the leaders in data center interconnect optics and the rivals arriving at the next speed
Marvell and Acacia remain the leaders in 400ZRx with movement likely as Ciena and Nokia begin shipments of 800ZRx.
backsC34 - S23tertiarynews2025-08-28the objection stated when the shares fell after the Q2 FY26 resultsbacksC29
- S24primarytranscript2026-03-05Marvell Q4 fiscal 2026 earnings call transcript (FMP), held 5 March 2026the FY28 outlook as given in March, and the December figure it replaced
reaching approximately $15 billion, roughly $2 billion higher than the outlook we provided in our December earnings call
backsC04 - S25primarytranscript2026-05-27Marvell Q1 fiscal 2027 earnings call transcript (FMP), held 27 May 2026the FY28 outlook as given in May
we now expect Marvell's fiscal 2028 revenue to reach approximately $16.5 billion, roughly $1.5 billion higher than the outlook we provided on our earnings call last quarter
backsC04