Merchant silicon is networking silicon — switch chips, PHYs, and related semiconductors — sold openly by a third-party semiconductor vendor to any equipment maker that wants to buy it. The word “merchant” is the tell: these vendors sell on the open market, the way a grocer sells produce, rather than building chips that only their own systems can use. Broadcom and Marvell are the defining examples; as 650 Group has published, the merchant silicon market is dominated by Broadcom and Marvell. The alternative to merchant silicon is vertically integrated silicon: chips designed by a systems company for its own products, such as Cisco’s Silicon One in Cisco systems or Juniper’s custom silicon in its routing platforms.
The merchant model matters because of what it enables. Any switch maker — from Cisco and Juniper down to ODM white-box builders — can buy the same Tomahawk or Teralynx silicon and compete on the layers above the chip: software, service, price, and supply chain. That single fact reshaped the networking industry. It created the white box and OCP ecosystems, it let hyperscalers like Google, Meta, and Amazon design their own switches around merchant chips, and it forced every proprietary-silicon vendor to justify why its own silicon is better than the open market’s. Our research shows the hyperscale and cloud data center is where merchant silicon wins most often, and the AI buildout has extended that lead.
The economics are the second reason the model matters. Merchant silicon vendors amortize a single chip design across dozens of customers, which drives a cost curve that a single-vendor chip cannot match. Speed transitions show this clearly: when the industry steps from 25.6T to 51.2T switch silicon — and now toward 102.4T — the merchant vendors ship those generations across the whole ecosystem within quarters, while a proprietary design ramps inside one company’s product line. For an AI cluster builder choosing switch silicon, the merchant route offers volume, second-sourcing, and a broader software ecosystem; the vertical route offers integration and, in some cases, differentiation at the feature level.
The AI era has raised the stakes on both sides. Broadcom’s Tomahawk family anchors the Ethernet switch layer of many AI fabrics, and its scale-up silicon ambitions (including co-packaged optics) aim at the accelerator fabric itself. Marvell fields Teralynx switch silicon alongside its captive relationships. Meanwhile NVIDIA sells Spectrum-X Ethernet to the market while using it in its own systems — a hybrid posture that is neither purely merchant nor purely vertical. Cisco sells Silicon One to third parties as well, another hybrid. We think these hybrid postures are the notable structural change of the last two years: the clean line between “merchant” and “captive” has blurred, and share data — which we track quarterly — now has to account for chips that serve both their maker and the open market.
We believe merchant silicon’s role will keep expanding through the AI buildout: switch generations are stepping faster than any single systems vendor can consume them, hyperscalers want freedom of action across ODMs, and the cost curve favors the open market. We publish our merchant silicon share and forecast data here, and our data center switch research covers the systems that consume it.
Tags: #Merchant Silicon #Broadcom #Marvell #Tomahawk #Teralynx #Silicon One #Spectrum-X #White Box #OCP #AI
650 Group researches several of the technologies we covered in this explainer, including Data Center, Co-Location & Cloud and Enterprise, and more.