Marvell data centre AI
Marvell: Custom Silicon and AI Interconnect
Marvell sells the silicon around the accelerator: custom XPUs, the attach chips and optical interconnect that tie AI clusters together, and Ethernet switching. Here is how its model works, how it differs from Broadcom, and what to treat as marketing.
Export controls: AI chips, servers, networking and memory can need an export licence depending on destination, end user and end use under US, UK and EU rules. We screen every project for this. See what applies.
How the Marvell model works
- XPU and XPU attach. Marvell describes two segments: the XPU, meaning custom CPUs and accelerators, and XPU attach, meaning the supporting devices such as PCIe retimers, co-processors, CXL controllers and co-packaged optics components.
- Pipeline of custom projects. At a July 2025 investor event Marvell said it had 18 active custom projects: 12 for four major hyperscalers and 6 for emerging AI customers. No customer is named, and the count is a company figure.
- Market size claims. Marvell estimates data-centre semiconductors at $94 billion by 2028, of which custom silicon is $55.4 billion, using third-party forecasts. These are projections, not results.
- How it differs from Broadcom. Both build custom silicon for hyperscalers. Marvell stresses the attach silicon and optical interconnect around the accelerator, and has a strong position in optical DSPs and coherent optics.
Products and technology
- Die-to-die and memory. Marvell claims die-to-die interfaces of up to 50 Tbps at under 0.1 picojoule per bit and custom SRAM with far higher bandwidth than standard SRAM. These are vendor claims without stated baselines.
- Co-packaged optics. A 3D silicon photonics engine of 6.4 Tbps integrates optics with the XPU package. It is described as being evaluated by customers, not yet shipping.
- Interconnect. PAM4 DSPs at 200G per lane are in production and 400G per lane has been demonstrated, according to Marvell. COLORZ pluggable coherent optics and Alaska active electrical cables are deployed at several US hyperscalers.
- Switching. Teralynx 51.2T Ethernet switches are in high volume, with 102T on the way.
Side by side
| XPUCustom accelerators | XPU attachAround the accelerator | Optical and DSPInterconnect | SwitchingEthernet | |
|---|---|---|---|---|
| What it is | Custom CPUs and accelerators designed with hyperscalers | PCIe retimers, co-processors, CXL controllers, co-packaged optics parts | PAM4 DSPs, COLORZ coherent optics, Alaska active electrical cables | Teralynx Ethernet switches |
| Marvell claim | 3 XPU and 2 emerging-customer XPU projects (July 2025) | 9 hyperscale and 4 emerging-customer attach projects (July 2025) | 200G per lane DSPs in production, 400G per lane demonstrated | 51.2T in high volume, 102T on the way |
| Status | Under design and ramping per Marvell | Under design and ramping per Marvell | Deployed at scale per Marvell | High volume |
| Customers | Four major hyperscalers, unnamed | Four major hyperscalers, unnamed | All five US hyperscalers for COLORZ, per Marvell | Not named |
Most performance figures are Marvell claims with no stated test method, and the product mix and customer counts refer to different periods. Treat them as direction.
The software and standards layer
- Silicon, not software. For custom silicon the customer owns the software. Marvell’s value is in the chip, package and interconnect, so there is no CUDA-style platform to adopt.
- Standards that matter. PCIe, CXL and Ethernet standards decide how well Marvell parts work with accelerators from NVIDIA, AMD and others.
- What it means for you. You will mainly meet Marvell inside optical modules, cables, network cards and switches, and inside cloud services built on custom accelerators.
- Keeping it secure. Firmware on optics, adapters and switches needs patching. See our agent security and defence practice.
What it means for your data centre
- Optics and power. Optical interconnect is a growing share of cluster power and cost, so the choice of optics and cabling affects both rack design and operating cost. See the rack power and cooling guide.
- Supply chain. Optical and attach components can set delivery dates as much as GPUs, so we check them early in an assessment.
- Networks. Compare Teralynx-based switches with other Ethernet options under AI networking.
Where this fits in the stack
Common questions
What is the difference between XPU and XPU attach?
The XPU is the custom accelerator or CPU. XPU attach is the supporting silicon around it, such as retimers, co-processors, controllers and optics.
Does Marvell sell AI GPUs?
No. It designs custom accelerators with large customers and sells the attach, optical and switching silicon around AI systems.
How does Marvell compare with Broadcom?
Both build custom XPUs and switching. Marvell emphasises attach silicon and optical interconnect, while Broadcom is the larger supplier of Ethernet switch silicon and custom XPUs.
Do export controls apply?
Advanced interconnect and AI silicon can need a licence depending on destination, end user and end use. We screen every project.
Sources and dating
Figures are as of 18 November 2026, from Marvell’s custom silicon blog, Marvell’s AI data centre page and Converge Digest. These are vendor and trade-press sources. Some figures are vendor claims or preliminary and some differ between sources, so we confirm against current datasheets when we specify a build. Marvell, Teralynx, COLORZ and Alaska are trademarks of Marvell Technology, Inc. Embedica is independent and is not affiliated with or endorsed by any of the companies named.
Next steps
Related: Other Chip Providers · Broadcom · NVIDIA Data Centre GPUs · AI Networking · Memory Suppliers
Planning interconnect and optics
Check supply and power early
