NVLink Fusion Ecosystem: Why Did NVIDIA Introduce Marvell? A New AI Infrastructure Architecture Is Reshaping

On June 2, 2026, during the Taipei COMPUTEX exhibition, NVIDIA CEO Huang Renxun and Marvell CEO Matt Murphy appeared together for a dialogue. Huang Renxun said, “The next trillion-dollar company, ladies and gentlemen,” in a remark that anticipated Marvell would become the next chip company to break through a market value of $1 trillion. On the day of the remarks, Marvell’s closing price was $290.79, up more than 32% for the day, and its market capitalization rose to approximately $254.4 billion.

This was not an off-the-cuff statement. Three months earlier, NVIDIA had injected $2 billion into Marvell, and the two sides had established a deep strategic cooperation relationship based on the NVLink Fusion platform. In the dialogue, Huang Renxun clearly pointed out that Marvell’s data center switching chips play an “Essential” role in processing AI workloads—when compute tasks are broken up and distributed across thousands of interconnected chips, “connectivity is everything.”

Evolution of AI Infrastructure Bottlenecks: From “Compute Scarcity” to “Interconnect Scarcity”

In his COMPUTEX speech, Matt Murphy said that the bottlenecks in AI infrastructure are undergoing a structural shift. The first phase of compute scarcity has been alleviated through large-scale GPU deployment; the current core bottlenecks are inter-chip interconnect bandwidth and latency; the bottlenecks in the next phase point to system-level optical interconnects and heterogeneous integration efficiency.

This assessment is rooted in the exponential expansion of AI cluster scale. When training AI clusters of 10,000-card scale—even 100,000-card scale—the efficiency of data exchange between compute chips directly determines the system’s actual utilization. Huang Renxun described this precisely during the dialogue: “When we break down computing problems and distribute them across the entire data center, it’s the connection that truly makes everything run.”

Marvell’s technical advantages align precisely with this bottleneck. According to Marvell’s FY2027 Q1 financial report, Marvell achieved operating revenue of $2.418 billion, up 28% year over year. Of this, the data center business contributed $1.83 billion, accounting for 76% of total revenue, up 27% year over year. Shipments of the 800G PAM4 product line continued to ramp up, while the 1.6T solutions had entered a phase of rapid scaling. The data center interconnect module (DCI) business is expected to reach an annualized revenue scale of $1 billion in FY2028. Operating cash flow in the quarter reached a record $638.8 million, providing ample liquidity to support subsequent capacity locking and M&A integration.

NVLink Fusion: NVIDIA’s “Open-Closed” Strategy and Marvell’s Position in the Ecosystem

In May 2025, NVIDIA launched the NVLink Fusion platform, opening its NVLink interconnect technology to customized XPU chips. The core idea of this platform is to connect customized chips from partners such as Marvell to NVIDIA’s rack-scale systems and end-to-end networking platforms, helping hyperscale cloud providers gain flexibility in deploying customized chips while maintaining compatibility with NVIDIA’s ecosystem.

In March 2026, the two sides upgraded this technical collaboration to a strategic-level partnership. NVIDIA injected $2 billion into Marvell. Marvell will provide customized XPU and Scale-Up interconnect solutions compatible with NVLink Fusion, while NVIDIA provides supporting technologies and the ecosystem. The scope of cooperation also extends into silicon photonics and AI-RAN telecommunications infrastructure.

From an architectural design perspective, the technical cornerstone of NVLink Fusion is a small chip that provides 1.8TB/s bidirectional bandwidth. With it, cloud providers can scale their own XPU deployment to millions of chips. Marvell’s value positioning in this ecosystem is that of an “interconnect infrastructure provider”—it does not directly compete with NVIDIA’s GPUs. Instead, it provides customized XPU solutions and high-speed switching solutions for the interconnect layer of AI data centers.

This strategy contrasts with Broadcom’s path. Broadcom is also in a leading position in the field of customized AI chips. Its 3.5D XDSiP packaging platform has been specifically optimized for AI XPUs, but its cooperation depth with NVIDIA’s ecosystem is weaker. Relying on NVLink Fusion, Marvell gained direct access to NVIDIA’s technology and supply-chain ecosystem, becoming an irreplaceable link in the interconnect layer structure of AI data centers.

Marvell’s Growth Path: Key Milestones for FY2027–FY2029

From the perspective of financial guidance, Marvell’s growth trajectory has taken shape more clearly:

FY2027 full-year guidance: total revenue of approximately $11.5 billion, up approximately 40%. The company has prepared about $1 billion in prepaid funds to secure additional capacity, with the first payments to begin in the second quarter.

FY2028 full-year guidance: total revenue of approximately $16.5 billion, up approximately 45%, which is $1.5 billion higher than the guidance for the previous quarter. The data center business is expected to grow by about 50% for the full year.

FY2029 custom chip revenue guidance: the target for revenue from custom chips on a single project exceeds $10 billion. Mizuho analyst William Kerwin noted that this means that between FY2028 and FY2029, the custom chip business alone could bring incremental revenue of $5 billion.

As of now, Marvell’s market capitalization is approximately $254.4 billion (as of the June 2, 2026 close), leaving roughly a 4x growth runway to reach a $1 trillion market capitalization. Under static valuation assumptions, with a P/E multiple of 35–40x, a $1 trillion market cap would require net profit in the range of $25–$28.5 billion. Based on FY2027 revenue of about $11.5 billion and an estimated non-GAAP net profit margin of about 25%–30%, the current net profit scale still has a significant gap to the above target. Therefore, whether Marvell can achieve sustained high-speed growth—rather than a one-time valuation jump—will be a key variable in assessing whether it can reach its $1 trillion goal.

Strategic M&A: Celestial AI and XConn Complete the Interconnect Puzzle

Between the end of 2025 and the beginning of 2026, Marvell completed two key acquisitions, filling gaps in its interconnect strategy from both technical capability and market coverage dimensions:

Acquisition of Celestial AI (signed in December 2025, completed in February 2026). Celestial AI has a breakthrough photonic transmission technology platform that can support optical I/O interconnects at the chip level, system level, and rack level. Interconnect for AI data centers is shifting from electrical interconnects between racks to optical interconnects at the chip packaging layer. This technology route is a necessary condition to enable large-scale, millions-of-cards-level interconnect deployment. The acquisition was completed with approximately $1 billion in cash, giving Marvell differentiated technological barriers in optical interconnects.

Acquisition of XConn Technologies (signed in January 2026). XConn has an industry-leading product line of PCIe 5/6 and CXL switching chips. PCIe 6 and CXL 3.1 switch chips have entered the sample-submission stage. The acquisition deal is valued at approximately $540 million. The value of CXL technology in high-performance computing lies in enabling memory pooling and resource sharing. XConn’s CXL switching chips will complement Marvell’s existing CXL memory expansion controllers.

Both acquisitions center on “Scale-Up interconnect,” complementing Marvell’s existing Scale-Out Ethernet switching business. Revenue contributions from the two acquisitions are expected to gradually materialize in FY2028–FY2029.

Marvell vs. Broadcom: A Final Comparison of Two Customized AI Chip Routes

Marvell and Broadcom are often regarded by investors as the “two-headed” leaders in the customized AI chip space, but the two companies’ ecosystem positioning differs fundamentally:

Broadcom’s route: centered on in-house ASIC customization capabilities, focusing on providing “anti-NVIDIA” customized AI chip solutions to cloud service providers such as AWS, Google, and Meta. Its 3.5D XDSiP packaging platform emphasizes performance and efficiency, but it is not deeply coupled with the NVIDIA ecosystem. In FY2025 Q1, the semiconductor segment (including customized silicon solutions) grew 11% year over year, but its growth rate is clearly lower than Marvell’s data center growth (+27%).

Marvell’s route: anchored on interconnect technology, delivering customized XPU and interconnect solutions within the NVIDIA ecosystem. The NVLink Fusion cooperation enables Marvell to participate in NVIDIA’s AI factories and AI-RAN ecosystem, while also maintaining coverage of external customers through comprehensive customized manufacturing cooperation with all major U.S. hyperscale cloud providers. CEO Matt Murphy explicitly stated in the earnings call: “We work with all major U.S. hyperscale cloud providers on comprehensive customized business collaborations.”

There is no simple determination of which of the two routes is superior, but Marvell’s support from the NVIDIA ecosystem gives it clearer near-term growth certainty during the current AI infrastructure expansion cycle. On the other hand, whether NVIDIA will limit Marvell to the interconnect layer rather than the core compute layer over the long term, and whether Marvell’s deep binding to NVIDIA will restrict its expansion into other GPU ecosystems, are variables that investors need to continuously monitor.

Conclusion

Huang Renxun’s “trillion-dollar” endorsement has brought substantial market attention to Marvell, but returning to fundamental analysis, achieving this goal requires all four core conditions to be satisfied in parallel:

First, AI data center capital expenditure must continue to maintain high growth—U.S. tech giants’ AI infrastructure spending was approximately $400 billion in 2025, and is expected to exceed $700 billion in 2026. This serves as the industry’s fundamental anchor for favorable conditions.

Second, Marvell’s customized chips and interconnect products must continue to win procurement share from hyperscale cloud providers, driving FY2028–FY2029 revenue guidance to be realized on schedule.

Third, commercialization validation at large-scale data centers must be achieved for the optical interconnect and CXL switching technologies from Celestial AI and XConn, providing new momentum for continued growth beyond FY2030.

Fourth, the NVIDIA–Marvell cooperation relationship must evolve from technical collaboration into deeper supply-chain and capacity commitments, sustaining a high degree of alignment in interests between both sides.

As of June 2026, Marvell’s customized chip product lineup includes 18 XPU projects, and is expected to accelerate the release of revenue contribution within FY2027. Q2 FY2027 revenue is expected to be approximately $2.7 billion, with a clearly defined sequential growth cadence quarter over quarter. From the perspectives of financial fundamentals, technological layout, and ecosystem positioning, Marvell already has the structural conditions to move from a $200 billion market cap toward $1 trillion. However, ahead remain tests such as capacity constraints, competitive pressure, and uncertainty in technology routes. While investors focus on the “trillion-dollar” halo, they should pay closer attention to the actual pace of FY2027 revenue realization and the evolution trend of gross margins for customized chips.