> ## Content Index
> Fetch the complete content index at: https://www.sotkn.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Who Makes AI Cluster Networking? The Switch, the Silicon and Who Gets Paid
- URL: https://www.sotkn.com/explained/who-makes-ai-cluster-networking/
- Published: 2026-11-02T01:19:53.000Z
- Updated: 2026-11-02T01:19:53.000Z
- Description: AI cluster networking is the switch chip, the chassis around it, and the links that let GPUs in one hall act as one machine. The chip is ordered before the chassis is accepted.
- Author: Sean
- Tags: Players, #desk-explained

EXPLAINED · Players · Last reviewed: 2 Nov 2026 · Next review: after Arista reports third-quarter results on 3 Nov 2026, and after NVIDIA reports the quarter ended October 2026

AI cluster networking is the gear that lets the GPUs in one hall act as one machine: the switch chip, the chassis around it, and the links between racks. It is not the GPU, and it is not the power equipment that feeds the building.

In short. NVIDIA and Broadcom design most of the switch silicon. Arista and Cisco sell branded chassis. Celestica builds a large share of the switches hyperscalers design themselves. The chip is ordered before the chassis is accepted. Our read: silicon is paid when the chip ships, the chassis vendor when the customer accepts it, and the stuck step is the switch chip.

Why an investor cares: extra GPUs earn nothing until they can talk. The fabric is a different cheque from the card. Our [guide to who buys AI GPUs](https://www.sotkn.com/explained/who-buys-ai-gpus/) covers the cards. This page covers who is paid for the links.

## What is AI cluster networking?

Three distances.

Scale-up is GPU-to-GPU traffic inside one rack. NVIDIA does it with NVLink. Broadcom said on its 2 September 2026 call that Tomahawk Ultra, an Ethernet switch chip, can do that job at the low loss and latency buyers used to get only from a proprietary link.

Scale-out is traffic between racks, a leaf-and-spine of switches. NVIDIA sells InfiniBand or Spectrum-X Ethernet. Broadcom sells the Tomahawk chip that goes in everyone else's chassis.

Scale-across is traffic between buildings and campuses, a third product beside the fabric inside the hall.

Our [guide to a GPU cluster](https://www.sotkn.com/explained/gpu-cluster-explained/) is the machine these links create. Our [guide to rack-scale AI](https://www.sotkn.com/explained/rack-scale-ai-explained/) is the rack. The laser in the optical port is a separate bind, in our [guide to silicon photonics](https://www.sotkn.com/explained/silicon-photonics-explained/).

## Who makes the switch silicon?

NVIDIA and Broadcom.

NVIDIA's data-center networking revenue was a record $14.8 billion in the quarter ended 26 April 2026, up 199% from a year earlier and 35% sequentially. That was 19.7% of its $75.2 billion of data-center revenue (our arithmetic). It then stopped printing the dollar. For the quarter ended 26 July 2026 it said only that networking grew 18% sequentially, and that Spectrum-X Ethernet grew 2.6 times from a year earlier. Our arithmetic: $14.8 billion times 1.18 is about $17.5 billion, about 19.6% of that quarter's $89.0 billion of data-center revenue. Treat $17.5 billion as a bridge, not a printed figure. On the same call it said a Vera Rubin rack can use InfiniBand or Ethernet for scale-out.

Broadcom sells the custom accelerator, which it calls an XPU, and the network beside it. AI semiconductor revenue was $16.7 billion in the quarter ended 2 August 2026, up 221% from a year earlier. On the call, Hock Tan said XPU shipments were 73% of AI revenue and that AI networking was up more than 2.5 times from a year earlier. Our arithmetic, treating 73% as a share of the $16.7 billion: the other 27% is about $4.5 billion. That networking line includes switch chips and optical parts, so it is wider than the chassis. Broadcom guided the next quarter's AI semiconductor revenue to $21.7 billion and said both XPUs and networking should triple from a year earlier.

The chip is Tomahawk. Broadcom said it was first with a 100-terabit Tomahawk 6, in 100-gigabit and 200-gigabit-per-lane versions, and that it had taped out Tomahawk 7, which it called the first 200-terabit-per-second Ethernet switch. Its president said Tomahawk 6 is in the hyperscalers that build custom accelerators with Broadcom, and in hyperscalers that do not.

## Who builds the chassis?

Arista, Cisco and Celestica sit on the other side of the same chip.

Arista's revenue was $3.036 billion in the quarter ended 30 June 2026, up 37.7% from a year earlier. It introduced 1.6-terabit AI fabric platforms, including liquid-cooled ones. On the call, management said Etherlink, its AI fabric, has more than 100 cumulative customers, against four or five in 2024, and named a 7060X-E7 switch at 100-terabit capacity. The chief financial officer raised the 2026 outlook to about $12.6 billion and put the AI-fabrics goal at at least $3.5 billion. Our arithmetic: that goal is 28% of the year's guide.

Cisco took $4 billion of hyperscaler AI-infrastructure orders in the quarter ended 25 July 2026 and $9.3 billion for the fiscal year. It recognised about $4 billion of revenue from that work in the year, and it expects $7.5 billion in fiscal 2027\. Our arithmetic: the year's orders were about 2.3 times revenue recognised, a gap of about $5.3 billion.

Celestica builds the box the hyperscaler designs. In the quarter ended 30 September 2026 its networking end market — renamed from Communications, same contents — was $3,263.2 million, 56% of $5,829.7 million of company revenue, against $1,942.9 million a year earlier. Our arithmetic: up 68%. Hardware-platform revenue was about $2.3 billion, up 70%. On the call, management said that growth came from 1.6-terabit and 800-gigabit switch programmes for more than one hyperscaler.

| Who       | What they sell                                                    | Latest hard number                                                                | When they are paid                         |
| --------- | ----------------------------------------------------------------- | --------------------------------------------------------------------------------- | ------------------------------------------ |
| NVIDIA    | InfiniBand, Spectrum-X Ethernet, NVLink                           | $14.8 billion networking revenue, quarter ended 26 April 2026                     | When its chip and switch ship              |
| Broadcom  | Tomahawk switch chips; optical parts share the AI networking line | AI semiconductor revenue $16.7 billion, quarter ended 2 August 2026               | When the chip ships to the chassis builder |
| Arista    | Branded Ethernet switches, including 1.6-terabit platforms        | Revenue $3.036 billion, quarter ended 30 June 2026                                | When the customer accepts the chassis      |
| Cisco     | Branded switches for hyperscaler AI builds                        | $9.3 billion of hyperscaler AI orders in fiscal 2026; about $4 billion recognised | When the order ships                       |
| Celestica | Hyperscaler-designed switches and systems                         | Networking revenue $3.26 billion, quarter ended 30 September 2026                 | When the box ships                         |

The foundry that prints the chip is paid earlier, when the wafer ships. Our [semiconductors layer](https://www.sotkn.com/stack/semiconductors/) follows that step. This page stops at the designers and the chassis.

## Why do orders run ahead of shipments?

Because the scarce object is the chip, and it is bought before the chassis can be accepted.

Arista's 10-Q for the quarter ended 30 June 2026 lists non-cancellable purchase commitments, not on the balance sheet, of $9.7 billion, of which $9.4 billion is due within 12 months. Our arithmetic: that book is 3.2 times the quarter's revenue, and 97% of it comes due inside a year. A year earlier it was $3.6 billion. The filing says the increase is for AI-network deployments and tighter supply of memory and silicon. The chief financial officer said it "mostly represents purchases for chips related to new products and AI deployments."

The same filing puts deferred revenue at $6,865.9 million. It is a mix of multi-year support and product held for acceptance clauses. Management said the majority is product, and that product deferred revenue rose about $600 million in the quarter. Our arithmetic: $6.87 billion is 2.3 quarters of revenue not yet recognised. Arista is paying for silicon now and booking some of the chassis later. Cisco's version of the same clock is the $9.3 billion order book against about $4 billion recognised.

## Who gets paid — and where is the stuck step?

Silicon is paid first, when the chip ships. The contract manufacturer is paid when the box leaves the factory. The branded vendor is paid when the customer accepts the chassis, which is why a chip commitment and a deferred-revenue balance can both be large at once. The hyperscaler pays and waits: GPUs already bought are not one cluster until the fabric is accepted. Our [guide to the hyperscalers](https://www.sotkn.com/explained/who-are-the-hyperscalers/) names who writes that cheque. A neocloud on NVIDIA's reference stack buys the fabric from NVIDIA too, which is why InfiniBand and Spectrum-X sit on the [NVIDIA Cloud Partner](https://www.sotkn.com/explained/nvidia-cloud-partner-explained/) bill of materials.

Our read: the stuck step is the switch chip, and the memory and silicon around it. A GPU order does not accept a chassis.

## What would change the view?

- NVIDIA prints a networking dollar again and it grows much more slowly than data-center revenue. The $17.5 billion bridge is only useful until then.
- Arista's purchase commitments fall while revenue holds, or its 3 November 2026 result misses the AI-fabrics goal of at least $3.5 billion.
- Cisco recognises the $7.5 billion of fiscal 2027 AI-infrastructure revenue it guided, instead of stacking more orders on unshipped gear.
- Broadcom says Tomahawk supply has caught demand. In September it said AI networking should keep growing with its custom accelerators.

## Common misconceptions

- People assume the GPU is the only scarce part in the rack. Actually the switch chip has its own clock. Arista's filing ties a $9.7 billion commitment book to AI networks and tighter silicon supply.
- People assume one company makes the network. Actually NVIDIA and Broadcom design the chip, Arista and Cisco sell a branded chassis, and Celestica builds a large share of the switches hyperscalers design themselves.
- People assume Ethernet has replaced InfiniBand. Actually NVIDIA still sells both, and it said a Vera Rubin system can use either for scale-out.
- People assume a hyperscaler order is revenue. Actually Cisco took $9.3 billion of AI-infrastructure orders in fiscal 2026 and recognised about $4 billion.

## FAQ

**What is AI cluster networking?**  
The gear that lets GPUs in one data center act as one machine: switch chips, the chassis around them, and the links between racks. It is not the GPU, and it is not the power equipment that feeds the hall.

**Who makes the switch chips for AI clusters?**  
NVIDIA and Broadcom do most of it. NVIDIA sells InfiniBand, Spectrum-X Ethernet and NVLink. Broadcom sells Tomahawk Ethernet switch chips, including a 100-terabit Tomahawk 6 and a 200-terabit-per-second Tomahawk 7 it had taped out by September 2026.

**Who builds the switch itself?**  
Arista and Cisco sell branded switches. Celestica builds a large share of the switches hyperscalers design themselves. In the quarter ended 30 September 2026, Celestica's networking end market was $3.26 billion, 56% of its revenue.

**Why do orders run ahead of revenue?**  
Because the chip is ordered before the chassis is accepted. Cisco took $9.3 billion of hyperscaler AI-infrastructure orders in fiscal 2026 and recognised about $4 billion. Arista had $9.7 billion of non-cancellable purchase commitments at 30 June 2026, mostly chips, and $6.87 billion of deferred revenue.

**Is InfiniBand the same as Ethernet?**  
No. InfiniBand is NVIDIA's own fabric. Ethernet is the open standard Broadcom, Arista and Cisco sell, and NVIDIA also sells it as Spectrum-X. A Vera Rubin rack can use either for scale-out.

**Is this a list of stocks to buy or sell?**  
No. This page explains who makes AI cluster networking and where the wait sits. It does not rank companies or give buy or sell calls. Not investment advice. Do your own research.

## Related terms

[GPU cluster](https://www.sotkn.com/explained/gpu-cluster-explained/) is the machine this fabric creates. [Rack-scale AI](https://www.sotkn.com/explained/rack-scale-ai-explained/) is the rack the scale-up link lives in. [Silicon photonics](https://www.sotkn.com/explained/silicon-photonics-explained/) is the laser in the optical port. [Who buys AI GPUs](https://www.sotkn.com/explained/who-buys-ai-gpus/) names the customers who also buy the fabric. [Who makes custom AI chips](https://www.sotkn.com/explained/who-makes-custom-ai-chips/) is Broadcom's XPU work, kept separate here from Tomahawk.

## Markets and stack doors

Layer: [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/). Also [Semiconductors](https://www.sotkn.com/stack/semiconductors/) and [Neoclouds](https://www.sotkn.com/stack/neoclouds/). Markets: [US](https://www.sotkn.com/markets/us/), [Taiwan](https://www.sotkn.com/markets/taiwan/).

How we check this. Printed dollars are from the company filings and releases linked below. Call-only remarks — NVIDIA's 18% and Spectrum-X multiple, Broadcom's 73% and Tomahawk descriptions, Arista's outlook and Etherlink count, Celestica's speed mix — are from those calls, two of them via a transcript, and are labelled in the sources. The $17.5 billion bridge, the percentage shares, the $4.5 billion slice, the 2.3 times, the $5.3 billion gap, and the 68% and 56% are our arithmetic. The $4.5 billion assumes Broadcom's 73% applies to the $16.7 billion AI semiconductor total. Who gets paid, and the stuck step, are our read. Not investment advice. Do your own research.

Last reviewed: 2 Nov 2026 · Next review: after Arista reports third-quarter results on 3 Nov 2026, and after NVIDIA reports the quarter ended October 2026

## Sources

1. [NVIDIA — CFO commentary, first quarter fiscal 2027](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000051/q1fy27cfocommentary.htm?ref=sotkn.com) — data-center networking revenue $14.8 billion, up 199% and 35% sequentially; data-center revenue $75.2 billion; quarter ended 26 April 2026.
2. [NVIDIA — Q2 fiscal 2027 earnings call (26 Aug 2026)](https://s201.q4cdn.com/141608511/files/content%5Ffiles/TRANSCRIPT%5F-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5%5F00-PM-ET.pdf?ref=sotkn.com) — networking up 18% sequentially; Spectrum-X Ethernet up 2.6 times; Vera Rubin can use InfiniBand or Ethernet; data-center revenue $89 billion.
3. [Broadcom — Q3 fiscal 2026 earnings release (2 Sep 2026)](https://www.sec.gov/Archives/edgar/data/1730168/000173016826000076/avgo-08022026x8kxex99.htm?ref=sotkn.com) — AI semiconductor revenue $16.7 billion, up 221%; next quarter guided to $21.7 billion.
4. [Broadcom — Q3 fiscal 2026 call, via The Motley Fool (2 Sep 2026)](https://www.fool.com/earnings/call-transcripts/2026/09/09/broadcom-avgo-q3-2026-earnings-call-transcript/?ref=sotkn.com) — XPUs 73% of AI revenue; AI networking up more than 2.5 times; Tomahawk 6 at 100 terabits; Tomahawk 7 taped out at 200 terabits per second.
5. [Arista — Q2 2026 results (4 Aug 2026)](https://investors.arista.com/Communications/Press-Releases-and-Events/Press-Release-Detail/2026/Arista-Networks-Inc--Reports-Second-Quarter-2026-Financial-Results/default.aspx?ref=sotkn.com) — revenue $3.036 billion, up 37.7%; 1.6-terabit AI fabric platforms.
6. [Arista — Form 10-Q, quarter ended 30 June 2026](https://www.sec.gov/Archives/edgar/data/1596532/000159653226000175/anet-20260630.htm?ref=sotkn.com) — purchase commitments $9.7 billion, of which $9.4 billion are due within 12 months; deferred revenue $6,865.9 million.
7. [Arista — Q2 2026 call, via Stock Analysis (4 Aug 2026)](https://stockanalysis.com/stocks/anet/transcripts/661010-q2-2026/?ref=sotkn.com) — 2026 outlook about $12.6 billion; AI-fabrics goal at least $3.5 billion; Etherlink above 100 customers; commitments mostly chips.
8. [Cisco — fiscal 2026 earnings release (12 Aug 2026)](https://www.sec.gov/Archives/edgar/data/858877/000085887726000106/exhibit991pressrelease-q4f.htm?ref=sotkn.com) — $4 billion of hyperscaler AI orders in Q4; $9.3 billion for the year; about $4 billion of revenue; $7.5 billion expected in fiscal 2027.
9. [Celestica — quarterly report, period ended 30 September 2026](https://www.sec.gov/Archives/edgar/data/1030894/000103089426000050/cls-20260930.htm?ref=sotkn.com) — networking revenue $3,263.2 million, 56% of $5,829.7 million; $1,942.9 million a year earlier.
10. [Celestica — Q3 2026 earnings release](https://corporate.celestica.com/news-releases/news-release-details/celestica-announces-third-quarter-2026-financial-results-and?ref=sotkn.com) — hardware-platform revenue about $2.3 billion, up 70%.
11. [Celestica — Q3 2026 call, via The Motley Fool (27 Oct 2026)](https://www.fool.com/earnings/call-transcripts/2026/10/27/celestica-cls-q3-2026-earnings-call-transcript/?ref=sotkn.com) — hardware-platform growth from 1.6-terabit and 800-gigabit switch programmes.