What Sustains Jensen Huang’s Trillion-Dollar Marvell?

In early June, when introducing Marvell CEO Matt Murphy at the Computex trade show in Taipei, Jensen Huang said, "They will be the next trillion-dollar company."

MRVL's stock price surged 32.5% that day. Over the following two weeks, fueled by speculation around the optical communications concept, its market capitalization briefly surpassed $270 billion.

Data source: Koyfin

The rest of the story is well known. As market skepticism grew over the return on investment (ROI) of AI capital expenditures, rate hike expectations, crowded trades, and leveraged liquidations occurred in succession. The US tech sector experienced a deep correction starting in late June. Having run up quickly with a high valuation, Marvell suffered a steeper correction than the broader market. By late July, it hit a low of $163, basically giving back all the gains made after Jensen Huang's remarks.

Entering August, the AI narrative warmed up again, coupled with a steady stream of positive news in the optical communications sector, bringing Marvell back into the market spotlight.

Returning to Jensen Huang's trillion-dollar prediction, his statement was not mere hype. Rather, he saw where AI infrastructure is heading next and who will become an indispensable player in it. To judge whether his words hold weight, we must first understand what Marvell does and why it is so difficult to replace.

What exactly does Marvell sell?

When discussing AI infrastructure, many people focus solely on GPUs. However, an AI cluster does not work simply by stacking tens of thousands of GPUs together; these chips must constantly exchange parameters, computational results, and memory data.

As GPU computing power continues to climb, the center of gravity for value in AI infrastructure is also shifting. In the past, the market focused more on "computing power," but in the future, the efficiency of an AI cluster will be determined not just by the GPUs themselves, but also by whether data can be exchanged efficiently between them. Consequently, connectivity has become one of the fastest-growing segments in terms of value within AI data centers.

Historically, this was mainly done using copper. Copper is cheap, mature, and reliable, long serving as the workhorse for data center interconnects over short distances. However, as port speeds evolve from 400G and 800G to 1.6T and even 3.2T, copper interconnects are beginning to approach their physical limits.

The issue is no longer cost, but rather the constraints of signal integrity. When speeds reach 1.6T and above, the attenuation of high-frequency electrical signals in copper media increases rapidly. To maintain signal quality, not only does power consumption rise significantly, but the stable transmission distance is also compressed to within a single cabinet. For AI data centers that require cross-cabinet and cross-cluster interconnects, traditional copper interconnects are increasingly falling short of meeting these needs.

As Jensen Huang said at GTC: "Copper has a physical limit on transmission distance." As electrical transmission gradually approaches its limits, replacing electricity with light has become the evolutionary direction of the industry.

This is precisely where Marvell's core competitiveness lies. Its advantage does not lie in a single optical module, but in covering the entire data transmission path: from low-level chip interfaces (SerDes) and signal conversion (DSP/Driver), to cross-server scaling (Retimer/CXL) and cluster scheduling (switch chips). Marvell has effectively positioned itself at almost every key node of data flow.

In addition to connectivity, Marvell's other major growth engine is custom ASICs. Google, Amazon, Microsoft, and Meta are all accelerating the development of self-designed XPUs to reduce their reliance on Nvidia. While cloud service providers define the chip architecture, Marvell leverages its physical design, high-speed IP, and advanced packaging capabilities to bring these designs into mass production.

Therefore, Marvell's positioning in AI data centers can be summarized into two things: Helping cloud service providers build their own XPUs and providing the high-speed interconnect capabilities that support these XPUs to work together efficiently.

In FY2026, Marvell's data center business revenue reached $6.1 billion, accounting for a whopping 74% of its total revenue (this proportion was only 40% two years ago, and reached 76% in the latest 1QFY27 quarter). Today, Marvell has transformed into an AI infrastructure company driven by the dual engines of custom ASICs and high-speed interconnects.

Custom ASICs: Grabbing a slice of the $73 billion pie, how does Marvell secure its position as the "No. 2 Player"?

Developing custom ASICs on advanced nodes is by no means easy. From physical design, HBM interface integration, and high-speed SerDes IP, to CoWoS advanced packaging and tape-out verification, the entire development cycle typically spans two to three years, requiring massive investment and presenting extremely high technical barriers. More importantly, once cloud service providers complete early-stage verification and lock in a partner, they do not easily switch suppliers throughout the product lifecycle, which also grants ASIC manufacturers high revenue visibility.

According to Bloomberg Intelligence forecasts, the global custom ASIC market size is expected to grow from approximately $10 billion in 2024 to about $73 billion in 2028, expanding more than sevenfold in four years, as the entire market undergoes rapid expansion.

Currently, Broadcom still holds an absolute leading position, while Marvell has become the most important second-source supplier in the custom ASIC market. Bloomberg Intelligence expects Broadcom to still capture about 80% to 90% of the market share by 2028, while Marvell is projected to increase its share from 5% to 10% in 2024 to 10% to 15%. For Marvell, this means it can both benefit from the rapid expansion of the market size and have the opportunity to further amplify its growth elasticity by gaining market share, enjoying the massive dividend brought by the industry's rapid expansion and gradually consolidating its position as the "No. 2 Player" in the custom ASIC market.

Source: Bloomberg Intelligence

Customer Roadmap: AWS lays the foundation, Microsoft takes the baton, Meta and Google penetrate

Based on the company's five-year cooperation agreements, product generational roadmaps, and industry supply chain tracking, Marvell's custom chip projects have a clear rollout pace:

• AWS: The current revenue cornerstone. The two parties have signed a five-year cooperation agreement covering multiple generations of products, with current revenue primarily derived from the 5nm Trainium 2. As subsequent 3nm products and Trainium 4 progress, the ASIC projects will also continue to drive sales of supporting products such as custom NICs, CXL memory expansion, and optical interconnects, continuously increasing the value per customer.
• Microsoft: The largest growth engine for FY28. Maia is expected to enter mass production in 2027 (corresponding to Marvell's FY28). Due to its use of multi-die packaging, its average selling price (ASP) per chip is higher. Management stated that the company has secured demand visibility for the entirety of FY28, with the project expected to contribute approximately one-third of that year's incremental custom ASIC revenue, serving as the core pillar to achieve the goal of doubling custom chip revenue in FY28.
• Meta and Google: Medium-to-long-term growth reserves. Meta has already contributed orders for custom networking products like the FB NIC, and continues to expand into interconnect products like CXL. The cooperation with Google started with the Axion ARM server CPU, with potential for further expansion into AI accelerators and other custom chip projects in the future.

Source: Bloomberg Intelligence

Three Hidden Risks Behind High Growth

Although the custom chip model has brought about a revenue explosion, there are non-negligible structural risks underlying its business model:

High customer concentration. In FY26, the top ten customers contributed 82% of revenue, with the largest customer accounting for approximately 14%. Because ASIC projects typically drive sales of supporting products like network interface cards (NICs), optical interconnects, and CXL, once a core customer's project is delayed, the impact is often not limited to a single product but propagates throughout the entire AI product portfolio.

Cloud service providers continue to strengthen their bargaining power. One of the core goals of self-developed chips is to reduce costs. As products gradually mature, cloud service providers generally tend to introduce a second-source supplier to improve supply chain resilience and enhance their bargaining power. Even if current products have been selected, future generational projects may still face more intense competition.

Operating leverage from advanced capacity. To secure 3nm process nodes and advanced packaging capacity, the company needs to pay deposits in advance and lock in capacity. Once a customer's project is delayed or demand is adjusted, the capacity originally secured for delivery may turn into working capital pressure, or even drag down earnings performance.

High-Speed Interconnects: Optical interconnects contribute cash flow, while switch chips and Scale-up open room for growth

Compared with custom ASICs, Marvell's networking business is more like a parallel layout of "the present and the future."

Currently, the company's largest revenue source still comes from optical interconnect-related products, including high-speed interconnect chips such as PAM4 DSP, TIA, Driver, SerDes, and DCI. Meanwhile, Marvell is also actively positioning itself in next-generation switch chips and Scale-up optical networking, hoping to capture new opportunities brought by the upgrade of AI network architectures.

As a single AI server integrates more and more GPUs, data transmission is gradually extending from between servers (Scale-out) to between chips (Scale-up), and network architectures are beginning to welcome a new round of upgrades.

Image source: Lightbits

Scale-out: 100T switch chips play catch-up with Broadcom

Marvell's Scale-out switch chip business is entering a stage of rapid volume expansion. Revenue in FY2026 was approximately $300 million, and management expects it to double to over $600 million in FY2027, with annualized revenue exceeding $1 billion in FY2028.

Currently, the company's revenue is still mainly derived from 12.8T and 51.2T products, while the key determinant of its next-generation competitiveness is the Teralynx T100 released this year. Built on a 3nm process, this product features a switching capacity of 102.4Tbps and can support 512 200G ports, with the company claiming it reduces power consumption by approximately 25% compared to competing products.

However, the challenge Marvell faces is not product performance, but customer validation. Broadcom's Tomahawk 6 has already taken the lead in entering mass production, while the Teralynx T100 is currently still in the sampling phase with Tier-1 cloud service providers. For Marvell, the significance of the T100 lies more in proving that it already possesses the product capabilities to compete in next-generation switch chips; translating this into large-scale revenue will still require waiting for customer adoption and mass production verification over the next two years.

Optical Interconnects: Upgrading from 800G to 1.6T, Content Value Continues to Rise

As total switch bandwidth upgrades from 51.2T to 102.4T, terminal network ports will also accelerate their transition from 800G to 1.6T.

Doubling the speed brings not only increased bandwidth, but also means signal processing difficulty increases exponentially. To ensure signal integrity, high-speed analog chips with higher performance, such as PAM4 DSP, TIA, and Driver, are required. This directly drives up the chip value per port (ASP), becoming the core driver for the continuous growth of Marvell's optical interconnect business.

This capability has also made Marvell an important partner in the Nvidia NVLink Fusion ecosystem. The two parties cooperate on underlying interconnect technologies such as high-speed SerDes and optical interconnect IP, helping cloud service providers integrate their self-developed ASICs and Nvidia GPUs into a unified high-speed interconnect architecture, further expanding Marvell's participation in AI networking.

In the data center interconnect (DCI) field, the company has shipped over 1 million DCI modules powered by its silicon photonics technology over the past decade. In FY2026, the DCI business contributed approximately $500 million in revenue, and management expects annualized revenue to increase to approximately $1 billion in FY2028.

In addition to DCI, the company also provides a range of high-speed interconnect chips such as PAM4 DSP, TIA, and Driver, which are widely used in AI data center internal networks. Due to the high technical barriers and high added value of these products, their overall gross margin is typically superior to custom ASICs, making this one of the company's highest-margin and most cash-flow-stable businesses. As 800G continues to scale up and 1.6T begins commercialization, optical interconnects will remain Marvell's most important profit pillar.

Scale-up: A $3.25 Billion Bet on Next-Generation Optical Interconnects

If Scale-out corresponds to today's data center network upgrades, then Scale-up bets on next-generation internal interconnects within AI servers.

Scale-up connects not different servers, but the GPU, CPU, and HBM within the same AI system. As a single AI server integrates more and more computing chips, the bottlenecks of traditional electrical interconnects in bandwidth, power consumption, and transmission distance are becoming increasingly apparent. The industry has begun exploring the use of optical signals to replace some electrical signals to achieve higher-speed and lower-power data transmission between chips.

To seize this opportunity, Marvell completed the acquisition of Celestial AI in 2026, with an upfront transaction consideration of approximately $3.25 billion. If agreed performance targets are met in the future, the total consideration will increase further.

Celestial AI's most core asset is its Photonic Fabric technology. Compared with traditional electrical interconnects, it uses optical signals to transmit data between chips, significantly reducing power consumption at higher bandwidths and breaking through the distance and bandwidth limitations of electrical interconnects. It is considered an important technical route for next-generation Scale-up networking. Currently, this technology has been adopted by a Tier-1 cloud service provider for a next-generation AI accelerator interconnect solution. Based on this, Marvell has raised its FY2028 Scale-up optics business revenue target from $150 million to over $300 million.

Three Questions Still Waiting to Be Verified

Celestial AI represents future growth potential, but what truly needs to be verified remain three questions:

• Whether the technology route can win. Photonic Fabric is still a new-generation Scale-up optical interconnect solution. Currently, the industry has not yet formed a unified standard, and uncertainty remains as to whether it can become the mainstream architecture in the future.
• Whether customer validation can translate into revenue. Although a Tier-1 cloud service provider has adopted the technology, it is still in the product validation phase, and it will take a long time to generate scaled revenue.
• Whether the heavy investment can yield returns. The upfront transaction consideration for this acquisition was approximately $32.5 billion, and the recognized acquisition consideration post-closing was approximately $35.4 billion, with the actual issuance of approximately 24.52 million shares of stock; if agreed performance targets are met in the future, an additional approximately 27.20 million shares will be issued. In other words, before the technology truly delivers commercial value, existing shareholders have already borne the cost of equity dilution upfront.

Book 'Failing Grade' vs. High Forward Valuation: A Battle of Valuation Paradigms

By traditional value-investing standards, Marvell can hardly be considered a typical 'high-quality company.' Over the past few years, the company's GAAP profitability has fluctuated widely, with its historical ROE staying at a low level for a long time and its ROIC remaining negative for several consecutive years. Meanwhile, despite the previous sharp correction, the company's current forward 12-month P/E ratio of approximately 46.5x remains significantly higher than its five-year historical average of approximately 32x.

Data Source: Koyfin

At first glance, this seems somewhat contradictory: why can a company with an undistinguished history of profitability secure a valuation higher than its historical average over the long term?

The reason is not complicated. Marvell's current valuation is built on the expectation of rapid expansion in its AI infrastructure business over the next two years. The question that truly needs to be answered is not 'how much money did it make in the past', but whether the growth of the next two years can be delivered.

Therefore, compared to historical financial data, the market will focus more on three questions in the future: whether revenue can maintain high-speed growth, whether profit margins can remain stable, and ultimately, whether this growth can truly translate into earnings per share (EPS).

High Revenue Growth Does Not Equate to Simultaneous Profit Growth

Management expects FY27 revenue to be approximately $11.5 billion, up about 40% year-over-year, and to further increase to approximately $16.5 billion in FY28, up about 43% year-over-year. Over the next two years, the company will remain in a phase where AI infrastructure demand is being rapidly unleashed.

However, revenue growth is primarily driven by custom ASICs, which happens to be a business with relatively lower gross margins for the company. Compared to optical interconnect chips, DSPs, and switch chips, ASICs must bear wafer procurement and advanced packaging costs. Therefore, as projects like Trainium 2 and Maia continue to scale up, the increasing revenue share of ASICs will put ongoing pressure on the overall gross margin.

In 1QFY27, the company's Non-GAAP gross margin was 58.9%, down 0.9 percentage points year-over-year. Therefore, the focus of future market attention is not whether the gross margin can continue to improve, but whether the company can maintain a stable overall gross margin close to 58% as the share of ASICs continues to increase.

EPS Growth Faces Share Dilution

For growth companies, revenue growth is only the first step. More importantly, whether this growth can ultimately reflect in earnings per share.

In 1QFY27, the company's GAAP net income was only $34.50 million, while Non-GAAP net income reached $718 million. The difference between the two mainly came from stock-based compensation (SBC) and amortization of intangible assets. Among these, SBC alone reached $208 million, representing approximately 9% of quarterly revenue.

Meanwhile, the company has continued to bolster its AI networking capabilities through M&A in recent years, including transactions such as Celestial AI and XConn, which have also brought continuous share expansion. Management expects the weighted average diluted share count for the next quarter to be approximately 915 million shares. If relevant performance targets are met in the future, there remains the possibility of further share issuances.

For investors, this means profit growth will not necessarily translate proportionally into EPS growth. If the share count continues to expand, the realization rate of EPS may lag behind the growth rate of profits themselves.

At 34x FY28 P/E, What Has the Market Already Priced In?

According to market consensus estimates, Marvell's FY27 Non-GAAP EPS is approximately $4.05, and FY28 is approximately $6.24. Based on the closing price on August 6, the company's current forward P/E ratios for FY27 and FY28 are approximately 52x and 34x.

A 34x FY28 Forward P/E is not unacceptable, but it implies a crucial premise: growth over the next two years must be delivered largely in line with market expectations.

This means that multiple growth engines, including AWS Trainium, Microsoft Maia, 1.6T optical interconnects, 100T switch chips, and Scale-up optical networks, all need to smoothly enter the volume ramp-up phase. If any of these core businesses experience significant delays, market earnings expectations for FY28 may be revised downward, the high valuation will lose support, and the stock price will face the risk of a simultaneous correction in both earnings and valuation.

In other words, Marvell's current valuation is essentially pricing in the performance delivery of the next two years in advance. Therefore, Marvell's greatest current risk is not a change in its growth thesis, but rather that the pace of growth realization cannot keep up with the expectations already reflected in the market.

Summary and Investment Thesis

Jensen Huang is bullish on Marvell not because it sells optical modules, but because it simultaneously controls the two most critical links in AI infrastructure: compute (custom ASICs) and connectivity (high-speed interconnects).

As cloud service providers continue to push self-developed AI chips and AI clusters continue to expand, data center demand for high-speed interconnects will also continue to rise. Marvell benefits from both industry trends simultaneously and is one of the few players globally capable of providing comprehensive connectivity solutions around XPUs, which is also its greatest competitive advantage.

However, from an investment perspective, the market has already fully recognized this thesis. Moving forward, what determines stock price performance will no longer be industry trends, but whether the company can consistently deliver on growth expectations.

Core verification points worth continuous attention in the future include:

• Whether Microsoft Maia can enter mass production as scheduled in 2027 (FY28);
• Whether the overall Non-GAAP gross margin can remain stable at around 58% after the share of ASICs increases;
• Whether Teralynx T100 and Scale-up optical interconnects can smoothly translate into scaled revenue;
• Whether share dilution from SBC and M&A will continue to impact the quality of EPS realization.

For Marvell, earnings growth over the next two years may not be the greatest challenge; the bigger challenge lies in whether this growth can exceed the expectations already priced into the stock. This is precisely the core logic behind why it enjoys a high valuation despite a 'failing grade' on paper.