Marvell (MRVL) is a publicly traded Networking & Custom Silicon company founded in 1995 and based in Santa Clara, USA.
| Date | Stage | Amount | Valuation | Lead investors |
|---|---|---|---|---|
| Jun 27, 2000 | IPO | $90M | — | — |
Marvell Technology reported record Q2 fiscal 2027 revenue of $2.739 billion, up 37% year-over-year, driven by accelerating Data Center segment growth of 46% YoY. GAAP gross margin contracted to 53.1% (58.9% non-GAAP), while GAAP diluted EPS was $0.33 (non-GAAP $0.94). Operating cash flow reached $605.5 million. The company highlighted strong AI-related bookings and significant acceleration in its Custom business starting in H2 fiscal 2027. Management raised full-year guidance for both fiscal 2027 and 2028. For Q3, the company guides revenue of $3.150B (±5%), GAAP gross margin 52.9%–53.9%, and GAAP diluted EPS of $0.53 (±$0.05).
Q3 FY2027 revenue $3.150B (±5%); raised FY2027 and FY2028 full-year outlooks

Marvell's AI revenue is still driven more by optical interconnects than by custom ASIC chips.

Google signs a broad custom-silicon deal with Marvell extending beyond the TPU into networking and infrastructure chips.

IEEE roadmap research finds multi-die assemblies and heterogeneous integration are becoming standard at 2nm and below, driven by AI compute demands outpacing transistor scaling.

Cisco enters the supercomputing market with a full-stack silicon and system play, leveraging its networking and UCS compute to compete with Nvidia.
Marvell co-designs custom AI accelerators (XPUs) and compute ASICs for hyperscalers, taking a customer's architecture to silicon on leading-edge TSMC nodes with Marvell's IP for SerDes, packaging, and chiplets. This custom-compute business, anchored by multi-year programs with Amazon, Microsoft, and Google, is Marvell's fastest-growing segment and positions it as the primary alternative to Broadcom for bespoke data-center AI chips.
Marvell's optical-interconnect portfolio, including industry-leading PAM4 DSPs and coherent DSPs that drive the optical modules linking GPUs and switches across AI clusters. As AI training scales out, bandwidth between accelerators becomes the bottleneck, and Marvell's electro-optics, spanning 800G and 1.6T, are central to the high-speed connectivity fabric of modern AI data centers.
Marvell's high-radix data-center Ethernet switch silicon, such as the Teralynx 10 (51.2 Tbps), built for low-latency AI and cloud networking. As hyperscalers push Ethernet as an open alternative to proprietary interconnects for AI fabrics, Teralynx competes with Broadcom Tomahawk and Nvidia Spectrum to move traffic between thousands of accelerators at the top of the rack and across the spine.
368 patents on file, but none with both an extractable figure and an abstract on Google Patents yet.