NVIDIA Data Center Revenue Reaches $75B as Marvell Rides Optical Networking Demand

NVIDIA Data Center Revenue Reaches $75B as Marvell Rides Optical Networking Demand

NVIDIA's position at the center of the AI infrastructure cycle remains solid, supported by trailing twelve-month revenue of $253.49 billion and a 74.1% gross margin. However, as physical bottlenecks in data center space, power, and networking deliverability intensify, NVIDIA is aggressively deploying its massive cash flows to secure and consolidate the physical networking and silicon photonics ecosystem. Concurrently, Broadcom is emerging as the dominant custom ASIC counterbalance, albeit with a major strategic shift in its packaging model.

NVIDIA's $6 Billion Silicon Photonics Consolidations

To bypass scale-out networking bottlenecks and accelerate the transition to co-packaged optics (CPO), NVIDIA has made a series of massive $2 billion strategic investments¹ in key hardware and optical suppliers:

• The Marvell $2 Billion Strategic Investment: On March 31, 2026, NVIDIA announced a $2 billion strategic investment in Marvell Technology (MRVL) to help Marvell ramp up its silicon photonics and networking technologies into volume production.
• The Lumentum and Coherent $2 Billion Investments: Earlier in March 2026, NVIDIA made two separate $2 billion investments in Lumentum and Coherent. These investments ensure that both suppliers can scale up production of lasers used in co-packaged optics (CPO) components, which NVIDIA is integrating directly into its Quantum-X InfiniBand and Spectrum-X Ethernet scale-out switches.

The Marvell Partnership and NVLink Fusion

The NVIDIA-Marvell strategic partnership centers heavily on NVLink Fusion, a rack-scale platform. Under this agreement:

• Marvell will provide custom XPUs and NVLink Fusion-compatible scale-up networking. This allows custom accelerators (such as AWS's Trainium 4, which is shepherded by Marvell and designed to support both UALink and NVLink protocols) to connect directly into NVIDIA's proprietary NVLink scale-up fabric.
• Marvell's custom silicon capabilities are heavily supported by two major acquisitions: Celestial AI (acquired in December 2025 for $3.25 billion), which developed a photonic fabric for row-scale coherent memory, and XConn Technologies (acquired in January 2026 for $540 million), expanding Marvell's leadership in high-lane PCI-Express and scale-up connectivity.
• This partnership enables hyperscalers like AWS to design custom chips for internal applications while maintaining compatibility with NVIDIA's hardware and software ecosystem.

Broadcom's Custom ASIC Dominance and "Chips Only" Pivot

Broadcom (AVGO) remains the leading counterbalance to NVIDIA's semiconductor hegemony, reporting Q2 FY2026 revenue of $22.19 billion (+47.9% YoY) and net income of $9.31 billion (+85.4% YoY). Broadcom's AI semiconductor revenue grew 143% YoY to $10.8 billion in Q2, with CEO Hock Tan expecting it to reach $16 billion in Q3 FY2026.

Key developments in Broadcom's custom ASIC business include:

• Six Core Custom Chip Customers: Broadcom is the custom silicon partner for six major players: Google (TPUs), Meta (MTIA), Anthropic, OpenAI (using Broadcom to develop its custom "Titan" XPU), Apple, and ByteDance.
• The "Chips Only" Strategy Shift: During the Q2 FY2026 earnings call on June 3, 2026, CEO Hock Tan announced a major strategy shift, stating that Broadcom will offer "chips only" instead of the complete integrated AI systems (racks/servers) it had previously planned to deliver to customers like OpenAI. This highlights the severe engineering and supply-chain complexities of rack-scale integration, forcing Broadcom to step back to pure silicon delivery.
• Physical Constraints on Deliveries: Tan noted that customer bookings are not for immediate delivery because hyperscalers "accept they need to align quite a few other things in place before they can deliver," pointing directly to the physical bottlenecks of power, cooling, and data center space.