Marvell: the hidden chips inside Big Tech's data centres

Everyone watches Nvidia. But inside the accounts of Google, Amazon and Microsoft sits a multi-billion-dollar line item signed Marvell. It is a name almost nobody reaches for first, yet it is there, embedded in the infrastructure of the most powerful companies on the planet. Which raises one question: what exactly does this giant do, after roughly tripling on the stock market in a year?

The present, briefly

On October 21st 2026 Marvell was worth about $215bn, before the November fall. In the most recent reported quarter, closed on August 1st 2026, revenue rose 37% year on year, driven above all by one area: AI data centres, up 46% and now accounting for 79% of total revenue, against 74% a year earlier. The shares have also been pushed by the Google agreement, which could generate up to $120bn in custom-chip revenue by 2033 — a signal of the race among cloud giants to develop their own AI silicon and reduce their dependence on Nvidia. To understand why that deal matters, though, you have to go back. Thirty years back.

The 30 years that made Marvell

Sehat Sutardja grew up in Jakarta. By 13 he was a certified radio technician, taking equipment apart and building generators from spares in his father's car-parts shop. He moved to the United States, took a degree in electrical engineering at Iowa State, then a doctorate at Berkeley. It was his wife, Weili Dai, who pushed to start a company: the idea came more from her than from him. The trio was completed by Sehat's brother Pantas, also a Berkeley PhD, who brought the crucial experience gained at IBM on hard-disk technology — indispensable for the company's first product. The name came from "marvelous": they wanted to build extraordinary things, and they had noticed that the most successful technology companies of the era all ended in "-el" or "-ell". Intel, Novell, Nortel. So they called it Marvell.

Sutardja spent his whole career as the exact opposite of the cover-story chief executive: an extremely low profile, never an outsized statement, while other chip founders filled the trade press. In 2009 he and Weili gave more than $20m to Berkeley, which named an entire research building after them, Sutardja Dai Hall. When Sehat died suddenly in 2024 at 63, the building had carried his name for fifteen years.

The first product was a chip to read data off hard disks, built with a technology everyone then considered too slow to work properly. The world's largest drive makers shut the door for two years — not because the product was poor, but because entrusting so critical a component, the one that literally reads your data off the disk, to an unknown start-up with no track record was a risk no major manufacturer wanted to take. A faulty chip across millions of drives would have meant lost data worldwide. It took nearly two years of unpaid work before Weili Dai, through sheer cold-calling, reached Kenneth Burns, a scientist at Seagate. Burns was stuck with a Texas Instruments chip that could not keep pace with new drive generations, and he set Marvell a challenge that was close to impossible. Sehat spent three months refining the design. When the chip passed, Seagate did not merely sign the order: it dropped Texas Instruments entirely and moved everything to Marvell. Within a few years Marvell was supplying the chips for roughly 90% of high-end enterprise hard drives — the base that carried it to the stock market in 2000.

After the 2000 IPO Marvell did not stop. It moved into networking, signed a development agreement with Intel on Ethernet chips, and its WiFi silicon ended up inside the first iPhone in 2007. In a few years it went from hard-disk specialist to invisible supplier to half of Silicon Valley.

Then, in March 2016, an internal audit-committee investigation found that management had put heavy pressure on sales teams to hit revenue targets, leading some revenue to be booked early. The Audit Committee said it had found no fraudulent activity, but the damage to trust was done. Starboard Value, an activist fund already holding 6.5% of the company, seized the moment and pushed for wholesale change. In April both Sutardja and Dai left operational control. In July a new chief executive arrived: Matt Murphy, from Maxim Integrated. He is still in charge today, nearly ten years on — the man who turned Marvell from a hard-disk chip company into one of the central names in AI infrastructure.

How Marvell makes money

Here is the part very few people grasp properly. Marvell does not physically make chips. It has no fabs of its own and has production done by third parties such as TSMC. Its trade is design, and it sells in four main ways:

Custom silicon. This is the piece that is exploding. Google, Amazon, Microsoft and Meta do not want to depend solely on Nvidia's standard chips. They want silicon designed specifically for their data centres and tuned to their particular workloads. Marvell designs it to order. The work resembles a tailor's: it does not sell an off-the-peg suit, it cuts one to the customer's measurements, and for that the customer pays more and is willing to sign multi-year contracts.

Networking and optical interconnect. Every AI data centre needs thousands of GPUs linked at very high speed. Marvell makes the chips that manage that connection, the optical DSPs and switches. Without them, all that GPU compute would sit bottlenecked.

Storage. The legacy business: controllers for hard drives and SSDs, used both in data centres and in consumer devices. It is the most mature segment, the least spectacular, and the pillar the company grew on for twenty years.

Carrier. Chips for 5G telecoms networks. A smaller slice, but one that spreads the risk of total dependence on AI. Until August 2025 Marvell also had an automotive business, since sold to Infineon for $2.5bn; it is no longer part of the company.

These four areas are not official reporting segments — Marvell splits its accounts only into "data centre" and the rest — but they help show where the revenue really comes from. And the data centre is now the heart of the company, at 79% of total revenue. In fiscal 2026 overall revenue rose 42% to $8.2bn; growth has accelerated since, driven above all by custom silicon, which management expects to expand further in the second half of the current fiscal year.

The Investor Day forecasts: up to $90bn

On October 6th 2026, at its Investor Day, management put numbers on the long term for the first time. For fiscal 2028, which ends in late January 2028, it now expects about $20bn of revenue, up from the $18bn it gave in August. For fiscal 2031 it points to a range of $70bn to $90bn, between eight and eleven times fiscal 2026 revenue. Within that, about $30bn should come from custom chips, which brought in about $1.5bn in fiscal 2026; in the central $80bn scenario, about $37.5bn would come from interconnect, the chips that link processors together inside data centres. The interim milestone has risen too: more than $12bn of custom-chip revenue in fiscal 2029, up from more than $10bn before.

These are the company's forecasts, not results, and should be read as such: before the event analysts expected, on average, about $47bn for fiscal 2031, far below the range given. And one detail cuts the most-quoted deal down to size: according to Murphy himself, revenue from the Google contract will not weigh significantly on the accounts before fiscal 2029.

On November 2nd Marvell announced the purchase of Lumentum, a maker of lasers and optical components for data centres, entirely in shares: 4.5 Marvell shares for each Lumentum share, for an implied enterprise value of $115bn. Once the deal closes, current Marvell shareholders will own about 67% of the combined company and Lumentum shareholders 33%. Marvell estimates more than $1bn a year of cost synergies; closing is expected in the first half of 2028 and needs the approval of shareholders and regulators. The logic of the deal: Marvell makes the chips that manage the signal, Lumentum the lasers that carry it through fibre, and together they want to compete with Broadcom across the whole connection of AI data centres.

Why not just anyone can do this, and where the risk sits

Designing a custom chip for a customer like Google is not like writing software. It takes years of specialised engineering, know-how on very advanced production processes at 5 and 3 nanometres, and trust built over time, because the customer is handing you the most critical infrastructure it owns. That is Marvell's moat: very few companies in the world can do it. In practice the most direct and most formidable competitor is Broadcom, which today has higher margins and larger contracts. Smaller Taiwanese players such as Alchip and GUC are also entering the market and pressing on margins.

But this is also where the biggest risk sits for anyone looking at Marvell as an investor: heavy dependence on a handful of enormous customers. In the most recent reported quarter a single customer, identified as "Distributor A" in the filings, accounted for 44% of revenue, up from 37% across fiscal 2026 and 34% in fiscal 2025. It should be said that a distributor can conceal several end customers, so the figure does not automatically mean one hyperscaler. Even so, if just one of the large hyperscalers decided to develop more of its silicon internally and lean less on Marvell, the revenue impact could be enormous.

And here is the paradox. Hyperscalers hold enormous bargaining power, but once a custom chip has been designed, validated and built into the infrastructure, switching supplier becomes extremely expensive. That can turn Marvell from a mere vendor into a long-term strategic partner.

It is a superb business as long as Marvell stays indispensable. The deeper the customer relationship, the harder it is to replace. But the more important the customer is to Marvell, the stronger that customer's bargaining power. And today the company is heavily exposed to a single growth engine: AI. On top of that comes integration: a $115bn acquisition paid for in shares dilutes current shareholders by about a third.

The stock: a rollercoaster

No buy or sell advice here, only the facts. Over the past 52 weeks Marvell shares have traded as low as about $71, on February 5th 2026, and as high as almost $330, on June 18th. Yes, a little over four months apart. Through 2025 the stock went through a sharp correction, punished by market doubts about the durability of AI-driven growth and by quarters judged disappointing despite numbers that were still growing. Then the narrative changed: a run of record quarterly results, the acceleration in custom silicon, and finally the Google deal. But the fall is as much part of the story as the climb. After the latest quarter, despite beating expectations, the stock lost 7% in a day. In October, after the Investor Day rally, it fell again: between October 9th and 16th it lost about 17%, along with other chip stocks, on fears about how the AI race will be financed. On October 21st it stood at about $245, roughly a quarter below its June high. On November 2nd, on the announcement of the Lumentum acquisition, it closed at $171.41, down about 13%, on doubts about the price paid, approvals and the durability of AI spending, and on November 4th it stood at $159.80.

Source: Yahoo Finance, weekly closing prices. Data as of 4 November 2026.

The shares currently trade on a very high earnings multiple relative to the sector's historical average, which means the market is pricing in years of future growth. The real test for Marvell, however, is no longer proving that AI is growing. It is proving that it can turn that growth into recurring revenue and high margins.

AI made Marvell grow. The question is how long it can keep doing so.