Which companies make scale-up fabric for AI data centers?

This catalog lists 6 companies on Scale-up fabric, 5 of them with listed equity: Alphabet (Google), AMD, Astera Labs, Broadcom, Nvidia, UALink consortium.

How reliable are these scale-up fabric supplier placements?

The 6 placements on this part break down as 4 disclosed, 2 inferred. "disclosed" means the company or a filing says so; "reported" comes from trade press or a reference-design list; "inferred" is a taxonomy lead with no citation yet.

3.01

Who makes scale-up fabric for AI data centers

Scale-up fabric (GPU-to-GPU inside the rack/pod)

The fabric linking GPUs to each other inside a single rack or pod, separate from the switching fabric that connects racks to each other.

Scale-up fabric is the shortest, highest-bandwidth network in the data center: it lets GPUs inside the same rack or pod address each other's memory directly, rather than going through the scale-out switch fabric (3.02, 3.03). It sits upstream of the copper spine and backplane cartridges (3.10) that physically wire it together and, at higher speeds, of the optics and co-packaged approaches (3.04, 3.08) being evaluated as a copper replacement. AI training and inference workloads that split a single model across many accelerators depend on this fabric scaling in lockstep with GPU count, since it sets the ceiling on how large a tightly coupled GPU domain can get. An investor should watch whether a proprietary approach keeps its lead over open alternatives being pushed by a wider industry consortium, since that split determines how concentrated the supplier base for switch chips, cables, and connectors stays as pod sizes grow.

AI delta: Larger AI models split across more GPUs per training or inference job, which raises the bandwidth and reach demanded of the scale-up fabric linking them inside a rack or pod.

Appears at:Large colo / hyperscale building AI campus GW-class campus· cooling:any — Present wherever multi-GPU racks or pods exist; the fabric's reach and topology (single rack vs. multi-rack pod) scale with cluster size rather than being an enterprise feature.

Companies on this part (6)

Show:

Company	Role	Ownership	Ticker	Price	1d	Confidence	Note
Alphabet (Google)	manufacturer	public	GOOGL	320 USD	-1.3%	disclosed	ICI · source
AMD	manufacturer	public	AMD	517 USD	+2.2%	inferred	Infinity Fabric
Astera Labs	manufacturer	public	ALAB	257 USD	+0.8%	disclosed	Scorpio X · source
Broadcom	manufacturer	public	AVGO	322 USD	-0.3%	inferred	scale-up Ethernet (Tomahawk Ultra)
Nvidia	manufacturer	public	NVDA	211 USD	-0.7%	disclosed	NVLink/NVSwitch · source
UALink consortium	manufacturer	cooperative				disclosed	members: AMD, Broadcom, Astera, Marvell, Cisco, hyperscalers · source

Who competes here

The same table read the other way: every company above competes for scale-up fabric sockets. Confidence tags matter — an inferred placement is a lead, not a confirmed rivalry.

Seen in real machines

scale-up fabric is named in 7 of the rack teardowns and supercomputer breakdowns on this site. Each line below is a piece of that machine. The companies shown are the ones named on that line, not necessarily the maker of this part.

Nvidia GB300 NVL72 · rack teardown
- 9× NVLink switch trays (2× NVSwitch each) — Nvidia
AMD Helios (MI455X) · rack teardown
- Scale-up switch tray (2x Broadcom Tomahawk 6 ASICs per tray, UALink-over-Ethernet, 21.6 TB/s bidirectional per switch) — Broadcom, Celestica
Google TPU7x (Ironwood) · rack teardown
- Optical circuit switch fabric ("Apollo" — MEMS mirrors, connects cubes into pods and superpods) — Lumentum, Celestica
Aurora · supercomputer
- Xe Link GPU-to-GPU fabric (all-to-all among the 6 GPUs in a node, 28 GB/s per link) — Intel, TSMC
Eagle (Microsoft Azure) · supercomputer
- In-node GPU fabric (NVSwitch and NVLink 4.0, 3.6 TB/s bisectional bandwidth between the 8 GPUs of each VM) — Nvidia
Eos · supercomputer
- H100 GPU and in-system NVLink/NVSwitch (900 GB/s between any two GPUs in a DGX H100) — Nvidia
MareNostrum 5 · supercomputer
- In-node GPU fabric (NVLink between the four H100 GPUs, 300 GB/s bidirectional between pairs; x16 PCIe Gen5 to the host CPUs) — Nvidia

Nearby parts

Next in layer: Scale-out switch ASICs
Layer: Layer 3 — Networking & interconnect

From the blog

Part of the day: Scale-up fabric 2026-09-04

Watch

What Exactly Does NVLink do for Machine Learning (featuring Exxact Workstation w/dual 3090s)

Jeff Heaton — explains NVLink's role as GPU-to-GPU scale-up interconnect

Click to load from YouTube · open on YouTube ↗

Independent videos — no affiliation; nothing loads from YouTube until you click.

Latest feed

Jensen Huang warns Europe risks falling further behind US, China in AI — DigiTimes, 2026-10-23
Nvidia's CEO warned that Europe faces risks of falling behind the US and China in AI advancement, calling for faster adoption and eased regulations.
[AI Infrastructure Insight] How AI infrastructure will be designed for the future — newsroom:sk-hynix, 2026-10-22
Effective real-world AI application requires improvements across faster compute, wider memory bandwidth, higher-performance networking, better storage, and more stable power infrastructure.
Why Chip Design AI Benchmarks Must Go Beyond Functional Correctness — EE Times, 2026-10-21
The increasing need for specialized architectures across various domains like physical AI and edge devices is driving a demand for custom silicon optimized for specific performance, memory bandwidth, and energy targets.
SpaceX pushes AI scaling into industrial realm as power equipment becomes limiting factor — DigiTimes, 2026-10-21
The rapid scaling of AI infrastructure is being constrained by the availability of electrical capacity, cooling solutions, and necessary compute hardware.
Interview: SEMI CEO sees US$3 trillion chip industry by 2035, as trusted partners reshape supply chain — DigiTimes, 2026-10-16
SEMI's CEO predicts the semiconductor industry will achieve $2 trillion in revenue by 2035, driven by AI and quantum computing growth.
AI’s $9-Trillion Question — HPCwire, 2026-10-15
The rapid growth of the AI boom is significantly inflating costs across the supply chain, including GPUs, CPUs, memory, networking, real estate, and power infrastructure.
AI chip competition shifts to data-center efficiency as memory, interconnect limits bite — DigiTimes, 2026-10-14
The competition in AI chip development is pivoting toward data-center efficiency to overcome scaling limitations in high-bandwidth memory and interconnect speeds.
Google data center strategic sourcing exec heads to OpenAI — Data Center Dynamics, 2026-10-13
Google's strategic sourcing executive is joining OpenAI to focus on high-speed interconnects and rack power infrastructure.
Growing pains: how distributed AI training changes the network between datacenters — The Register, 2026-10-09
The article discusses how distributed AI training requirements are necessitating more robust and low-latency network architectures between data centers.
CScale Exits Stealth with $145M to Build the Interconnect for Gigawatt-Scale AI — HPCwire, 2026-09-30
CScale has raised $145 million in funding to develop high-density interconnect infrastructure specifically designed for gigawatt-scale AI data centers.

Questions about scale-up fabric

What is scale-up fabric?: The fabric linking GPUs to each other inside a single rack or pod, separate from the switching fabric that connects racks to each other.
Which companies make scale-up fabric for AI data centers?: This catalog lists 6 companies on Scale-up fabric, 5 of them with listed equity: Alphabet (Google), AMD, Astera Labs, Broadcom, Nvidia, UALink consortium.
How does AI change demand for scale-up fabric?: Larger AI models split across more GPUs per training or inference job, which raises the bandwidth and reach demanded of the scale-up fabric linking them inside a rack or pod.
How reliable are these scale-up fabric supplier placements?: The 6 placements on this part break down as 4 disclosed, 2 inferred. "disclosed" means the company or a filing says so; "reported" comes from trade press or a reference-design list; "inferred" is a taxonomy lead with no citation yet.

Sources: placements without a source URL come from the taxonomy seed and are tagged accordingly. How this data is built.· Markdown · JSON