AI Networking Model
Device-level tracking of the networks behind AI clusters: switches, transceivers, cables, and AECs across scale-up, scale-across, scale-out, front-end, and out-of-band fabrics. Used by hyperscalers, networking vendors, and institutional investors to size units, spend, share, and power as deployments ramp.
- Coverage
- 2023–2026
- Config panels
- 80+
- Vendors tracked
- 25+
- Speeds covered
- 200G–1.6T
The network behind every AI cluster, device by device.
The SemiAnalysis AI Networking Model decodes the networking layer of AI infrastructure: what every hyperscaler and neocloud actually deploys around each accelerator, what it costs, and who supplies it. Coverage runs from switches and optical modules to fiber, backplane, AECs, ACCs, ALCs, and DACs, across scale-up, scale-across, scale-out, front-end, and out-of-band networks.
Every network is built bottom-up from cluster configurations, so units, spend, and power trace back to the specific accelerator and topology they originate from. A top-down market view reconciles the build against total volumes and vendor share.
Who uses it, and what it decides.
The buyers of this model, and the calls they make with it.
- Hyperscalers & neoclouds
- Benchmarking cluster network designs and costs against every other major buyer: configurations, attach ratios, pricing, and power per accelerator type.
- Networking & component vendors
- Wallet share by customer and SKU: how each hyperscaler allocates switch, optics, and cable dollars, and where that share is shifting.
- Public & private market investors
- Revenue models for switch, optics, and interconnect names, built from unit volumes and pricing rather than vendor guidance.
Four modules, one network build.
Each module is built from primary research and reconciled against the others, so configurations, spend, and market share stay consistent.
Cluster configurations
Configuration panels covering how Microsoft, Google, Meta, Amazon, Oracle, X.AI, and the neoclouds build AI cluster networks for each accelerator type, across five network layers:
- Scale-up
- Scale-across
- Scale-out backend
- Front-end
- Out-of-band
Each panel carries the switch, optical module, fiber, AEC, ACC, ALC, and DAC quantities in use, their attach ratios to accelerators, and the pricing and power consumption of every device involved.
Bottom-up builds & spend
Networking units and spend built configuration by configuration for every deployed accelerator, from H100 and H200 through B200, GB200 NVL36 and NVL72, B300, GB300 NVL, VR200, MI300, MI325, MI350, and custom ASICs. Vendor coverage spans the full device stack:
- SwitchesNvidia, Arista, Celestica, Cisco, Accton, Juniper, Nexthop, Nokia, Huawei, Broadcom, MarvellSwitching
- Optical modules & CPONvidia, Zhongji Innolight, Coherent, Eoptolink, Fabrinet, TFC Optical, Lumentum/Cloudlight, AAOI, Accelink, Source Photonics, Broadcom, Cisco, MaxLinearOptics
- AECs, ACCs, DACs, ALCs, backplane & fiberNvidia, Credo, Astera Labs, Amphenol, TE Connectivity, Molex, Luxshare, Broadex, Corning, US ConecInterconnect
- TelecomCiena, NokiaTelecom
Includes market-wide volume forecasts for 200G, 400G, 800G, and 1.6T transceivers, unit shipment forecasts for Broadcom, Arista, and Nvidia switches, and per-SKU procurement detail traceable back to the configuration it originates from.
Top-down market & share
Total market conditions and vendor share, reconciled against the bottom-up build:
Volumes by speed
Optical module volumes at 1.6T, 800G, and 400G, quarterly history with forecasts through 2026
Vendor market share
Share by vendor for each speed class, history and forecast
Wallet share
Per-hyperscaler allocation across optics, black-box and white-box switches, and AECs
Master pricing & power
A master pricing table for the devices and components in AI datacenter networks, priced for hyperscalers, neocloud giants, and emerging neoclouds, with power budgets alongside:
- Switches
- Optical modules
- CPO components
- AECs, ACCs, DACs & ALCs
- Fibers
- Backplane
- NICs
- Cooling components
- Connectors
How the model is built.
Every device is counted where it is deployed, then checked against the market.
-
Configuration build
Each buyer’s clusters are mapped panel by panel: topology, device SKUs, attach ratios, pricing, and power per accelerator type.
-
Spend roll-up
Configurations roll up to by-customer and market-wide units and spend, traceable back to the cluster they came from.
-
Market reconciliation
Top-down volumes and vendor share are reconciled against the bottom-up build, by speed class and by SKU.
Research that ships with the model.
Model subscribers receive the update notes, webinars, and analysis published against each release. A sample of recent coverage:
Common questions.
Anything not covered here, ask the team directly through the form below.
What does the AI Networking Model include?
Four reconciled modules: cluster network configurations for every major buyer, bottom-up networking builds and spend, top-down market volumes and vendor share, and a master pricing and power table. Data runs quarterly with forecasts through 2026.
How is the model delivered?
As an Excel workbook with dashboard access for your team.
Is it part of the SemiAnalysis newsletter subscription?
No. Industry models are separate institutional offerings and are not included with the annual newsletter membership.
Can the data estimate vendor revenue?
Yes. Unit volumes, pricing, and wallet share by customer mean revenue for switch, optics, and interconnect vendors can be estimated directly from the build.
Models that pair with this one.
Pairing depends on the decision in front of you. The configurations here attach to the accelerator build and feed the cost models downstream.
Accelerator & HBM Model
Demand forecastThe SKU-level accelerator build these network configurations attach to: shipments, pricing, and HBM supply.
View model
AI Cloud TCO Model
Cost modelRental economics and total cost of ownership once clusters, networks included, are deployed.
View model
Datacenter Industry Model
Capacity modelThe power and datacenter capacity these networks are built into, tracked site by site.
View modelGet the AI Networking Model.
Start with the sales team. They come back with scoping, licensing, and pricing for your mandate.
- Scoped to your use case
- Dashboard and API access for your team
- Custom research engagements available
