Network / Fabric
▶ See it in 3D The Memory Stack region
How racks of GPUs actually talk to each other. Inside a rack that's NVIDIA's NVLink today; the rest of the industry is pushing open alternatives. This is also where optics (light) is taking over from copper — the fastest-growing slice of the whole build-out.
What it is
From 400/800G InfiniBand & Ethernet to the NVLink switch fabric inside a GB200 NVL72 rack (130 TB/s). The point where "the rack" becomes the unit of compute — and where optics takes over from copper.
The business
The interconnect battleground: NVIDIA (NVLink moat) vs the merchant camp Broadcom + Marvell + Astera (UALink/Ethernet). Optics is the fastest-growing slice. Open the Photonics region.
Bottleneck severity
The cables aren't the chokepoint — three specific things are. (1) NVLink lock-in: inside a rack, GPUs talk over NVIDIA's proprietary fabric, so rival accelerators can't join the cluster. (2) Lasers: between racks, data travels as light, and every optical port needs an indium-phosphide laser — fewer than five firms make them at volume, sold out into 2027 after NVIDIA's ~$4B lockup. (3) Switch silicon: a near-duopoly (Broadcom, NVIDIA). Miss any one and the cluster stops scaling.
Who makes it
NVIDIANVDA
Buys the most HBM, NVLink and optics on earth and pre-buys supply to lock rivals out: ~60% of CoWoS, $2B each into Coherent and Lumentum (Mar 2026), $2B into Marvell. B300 (288 GB HBM3E) is the shipping reference GPU; Rubin (288 GB HBM4, 22 TB/s) samples Q4 2026, volume Q1 2027.
BroadcomAVGO
Tomahawk switching + CPO (Davisson 102.4 Tbps) + custom XPUs for hyperscalers. Anchor of the UALink/merchant-silicon camp.
MarvellMRVL
~70% optical-DSP share + custom AI silicon + Celestial AI photonic memory fabric ($3.25B). Most complete optical stack in public equities.
Astera LabsALAB
The only large-cap pure-play on AI connectivity — Aries retimers, Leo CXL controllers, Scorpio fabric switches. Q2 2026 revenue $392M (+104% YoY) as Scorpio X entered volume; Q3 guided $540–560M. Building NVLink Fusion hybrid racks for 2027. Risk: one customer is most of revenue.
CoherentCOHR
~25% transceiver share + vertically integrated InP lasers, with all EML capacity locked under long-term agreements through 2027. NVIDIA invested $2B (Mar 2026) explicitly to expand indium-phosphide capacity and pre-allocate it.
LumentumLITE
The only vendor shipping 200G/lane EML lasers at volume — the exact part 1.6T optics needs — with all EML capacity under long-term agreements through 2027. NVIDIA invested $2B (Mar 2026) to lock it. Purest laser-chokepoint play.
FabrinetFN
Contract assembler for Coherent, Lumentum, Ciena. The cleanest pluggable-demand barometer — and Thailand is now where InnoLight and Eoptolink build US-bound 1.6T volume too.
InnoLight300308.SZ
World's largest transceiver maker: ~35% of global 800G share, Q1 2026 revenue +207% YoY on the 1.6T ramp; ~60% of NVIDIA's 800G volume with Eoptolink. Depends on US DSPs + JP/US lasers; Thailand plants dodge tariffs.
CienaCIEN
Coherent long-haul transport (WaveLogic) — the pipes between data centers. Indirect AI beneficiary via data-center interconnect, not intra-cluster optics.
Metrics that move this layer
- Optical content $/GPUrising every generation
- NVLink vs UALinkthe standards war decides the TAM split
- switch shareBroadcom vs NVIDIA vs Marvell