Marvell has expanded its tie-up with GlobalFoundries to boost silicon germanium wafer output for optical transceivers, feeding an AI datacentre optics market Nvidia's Jensen Huang once tipped to make Marvell a trillion-dollar company. For UK buyers, the deal raises a harder question: does riding this optics wave mean tying infrastructure to one vendor's roadmap?
View the data behind this chart
| Current generation | Next generation | |
|---|---|---|
| Transceiver speed | Tbps1.6 | Tbps3.2 |
What Marvell and GlobalFoundries actually announced
Marvell and GlobalFoundries confirmed an expanded partnership on Thursday to grow silicon germanium (SiGe) wafer production at GloFo's Burlington, Vermont plant. SiGe is commonly used in optical transceivers, which convert electrical signals into light and back, and for laser modules used in switches from Nvidia and others.
GlobalFoundries says the additional capacity will support near- and co-packaged optics (NPO/CPO) alongside "next-gen" pluggable transceivers. Crucially, its SiGe process has already been validated to 200 Gbps per lane — sufficient for 1.6 Tbps transceivers, with a roadmap pointing towards 3.2 Tbps as lane speeds climb further.
Why optics is suddenly unavoidable in AI infrastructure
The driver is physics, not fashion. Marvell CEO Matt Murphy explained at Computex that as lane speeds rose from 25 Gbps through 50, 100 and 200 Gbps, copper interconnects kept pace. But at 400 Gbps per lane, copper effectively hits a wall: cables can't carry a usable signal beyond 1.25 metres, ruling copper out even within a single rack.
That's why Nvidia already uses co-packaged optics in some Spectrum Ethernet and Quantum InfiniBand switches, and why near-packaged optics is expected to feature in large multirack systems, including Huawei's Ascend 960DT-based SuperPods and Nvidia's own future multi-rack designs. Buyers scoping AI back-end fabric options should treat this lane-speed ceiling as a planning constant, not a future problem.
Marvell's optics ambitions go well beyond the rack
Marvell isn't stopping at rack-scale links. The company has said it's developing coherent optics and long-haul networking capable of linking data centres thousands of kilometres apart. Its Colorz 1600 coherent product, a 1.6 Tb/s solution built on a 2nm DSP, is aimed at inter-data-centre connectivity and due to sample later in 2026.
Alongside that sits the Ara 1.6 Tb/s interconnect family using 3nm DSPs and the Teralynx T100 switch, which supports 512 ports at 200 Gb/s or 64 ports at 1.6 Tb/s. Marvell's March 2026 partnership with Nvidia explicitly folds silicon photonics into NVLink Fusion, with Nvidia stating Marvell's optical DSP and photonics work would help customers build scalable AI infrastructure. Recent analysis also notes optics, not custom silicon, still drives the bulk of Marvell's AI-related revenue.
The lock-in question UK buyers need to ask
GlobalFoundries is positioning itself as a merchant co-packaged optics provider rather than an AI-chip foundry — supplying the optical layer to multiple customers rather than owning the whole stack. Its SCALE platform supports CWDM and DWDM with qualified 50 Gbps and 100 Gbps modulators, integrated photodiodes and coupled-ring resonators, and has demonstrated up to 16 DWDM lanes per fibre bidirectionally, implying up to 1.6 Tb/s per direction in theory.
One notable counterpoint: analysts say SCALE doesn't necessarily tie customers to GlobalFoundries' own silicon or packaging, and can be customised while staying compatible with OCI-MSA requirements — an open specification. That matters, because Nvidia's proprietary CPO switches are a different model, baking optics into a single vendor's hardware. Procurement teams weighing GPU accelerators and switching together should map which parts of their stack sit on open standards versus closed integration.
Supply chain, policy and cost signals to track
The optics buildout is being underwritten well beyond Marvell and GlobalFoundries. Nvidia has put roughly $2 billion each into Lumentum and Coherent, and Coherent has since announced a fourfold increase in indium phosphide wafer production at its Sherman, Texas fab, feeding lasers, photodetectors and modulators. GlobalFoundries' own silicon-photonics push has been backed by a US government package worth $300 million under the CHIPS Act, alongside a 1% stake arrangement, framed explicitly around AI datacentre networking needs.
For UK buyers, that concentration of US industrial policy and capital around a handful of suppliers is a genuine supply and pricing risk factor, not just a technical one. Marvell has also been acquiring capability — including PCIe and CXL switching silicon through its purchase of XConn Technologies — which extends its interconnect footprint further into the rack. Tracking AI server market share alongside these supply moves gives a clearer read on where pricing power is heading.
View the data behind this chart
| Merchant optics (GloFo) | Nvidia CPO switches | Pluggable modules | |
|---|---|---|---|
| Vendor lock-in risk | Lower, OCI-MSA compatible | Higher, integrated design | Lowest, swappable parts |
| Power efficiency | High, co-packaged design | Highest, fully integrated | Lower, more overhead |
| Upgrade flexibility | Good across suppliers | Tied to switch refresh | Easiest to replace |
| Supply diversification | Multiple customers served | Single-vendor roadmap | Broad market availability |
What this means for procurement decisions now
None of this requires an immediate rip-and-replace, but it does change the calculus for anyone building or refreshing AI clusters. Buyers should distinguish between optics embedded in proprietary switch platforms and merchant optics built to open interconnect specifications, since the latter preserves more negotiating leverage at renewal. Reviewing AI networking adoption trends alongside vendor roadmaps helps separate genuine capability upgrades from marketing timed to roadmap announcements.
Given the pace of change — 1.6 Tbps transceivers now, 3.2 Tbps on the horizon, coherent long-haul links sampling later this year — locking into long depreciation cycles on single-vendor optics is a real risk. Using third-party maintenance to avoid vendor lock-in and getting support through IT procurement services for multi-vendor optics contracts are practical ways to keep options open while the technology, and the economics, keep moving.
- 01The Register — Marvell pushes GlobalFoundries to light up wafer production · 18 September 2026
- 02Tom's Hardware — Marvell details vision of optically interconnected data centers · 1 January 2026
- 03The Register — Uncle Sam sees the light, offers GlobalFoundries $300m · 29 July 2026
- 04The Register — The tech that could make Marvell the next trillion-dollar company · 3 June 2026
- 05NVIDIA Investor Relations — NVIDIA AI Ecosystem Expands as Marvell Joins Forces Through NVLink Fusion · 1 March 2026
- 06Tom's Hardware — Co-packaged optics (CPO) foundry roadmaps: TSMC, Intel, Samsung and GlobalFoundries · 1 January 2026
- 07The Next Platform — Optics still driving Marvell's AI business more than custom chips · 2 September 2026
- 08DataCenterDynamics — Marvell acquires PCIe and CXL switch provider XConn Technologies for $540m · 1 January 2026