Layer 5 of 8 · 10 companies
Networking and optics
Which links in the network have pricing power, and which are just volume?
Thousands of GPUs only work as one machine if the network between them is fast enough, so AI data centers spend heavily on switches, optics and cables. This layer grew +45% in two years, but it kept only 53% of the extra revenue as gross profit. The question is which links in the network have pricing power and which are just volume.
- combined revenue, latest four quarters
- $141B
- combined revenue, latest four quarters
- revenue change over two years
- +45%
- revenue change over two years
- gross margin (53% two years earlier)
- 53%
- gross margin (53% two years earlier)
- Credo’s revenue growth over its latest four quarters, the fastest in the layer
- +165%
- Credo’s revenue growth over its latest four quarters, the fastest in the layer
Why the network is part of the computer
Training a large model splits the work across thousands of accelerators that must constantly exchange results. If the network is slow, expensive GPUs sit idle waiting for data. So an AI data center has two networks: a front end that connects it to users and storage, and a back end that links the accelerators to each other at very high speed. The back end is new spending that barely existed before AI.
Each part of that network is a different business. Switches (Arista, Cisco, Broadcom’s chips inside them) route the traffic. Optical transceivers and lasers (Coherent, Lumentum) turn electrical signals into light for longer links. Copper cables and retimers (Credo, Astera Labs) carry signals over short distances inside and between racks. Connectors and fiber (Amphenol, Corning) are the physical plumbing. Optical transport (Ciena) links data centers to each other. Inside the optical modules sit analog driver and amplifier chips from companies such as MACOM.
Quarterly revenue: networking and optics
The layer’s 10 US-listed companies, five largest shown separately; Other is Ciena, Lumentum, Credo, Astera Labs, MACOM
- Cisco
- Amphenol
- Corning
- Arista
- Coherent
- Other
Source: company 10-K and 10-Q filings (SEC XBRL); fiscal quarters aligned to the nearest calendar quarter. Companies appear from their first reported quarter, so newly listed ones (such as CoreWeave, Sandisk or Astera Labs) add to the total partway through.
Show the numbers
| Quarter | Cisco | Amphenol | Corning | Arista | Coherent | Other |
|---|---|---|---|---|---|---|
| Q1 2022 | $12.8B | $3.0B | $3.7B | $0.9B | $0.8B | $1.5B |
| Q2 2022 | $13.1B | $3.1B | $3.6B | $1.1B | $0.9B | $1.5B |
| Q3 2022 | $13.6B | $3.3B | $3.5B | $1.2B | $1.3B | $1.7B |
| Q4 2022 | $13.6B | $3.2B | $3.4B | $1.3B | $1.4B | $1.8B |
| Q1 2023 | $14.6B | $3.0B | $3.2B | $1.4B | $1.2B | $1.7B |
| Q2 2023 | $15.2B | $3.1B | $3.2B | $1.5B | $1.2B | $1.6B |
| Q3 2023 | $14.7B | $3.2B | $3.2B | $1.5B | $1.1B | $1.7B |
| Q4 2023 | $12.8B | $3.3B | $3.0B | $1.5B | $1.1B | $1.7B |
| Q1 2024 | $12.7B | $3.3B | $3.0B | $1.6B | $1.2B | $1.6B |
| Q2 2024 | $13.6B | $3.6B | $3.3B | $1.7B | $1.3B | $1.6B |
| Q3 2024 | $13.8B | $4.0B | $3.4B | $1.8B | $1.3B | $1.8B |
| Q4 2024 | $14.0B | $4.3B | $3.5B | $1.9B | $1.4B | $2.0B |
| Q1 2025 | $14.1B | $4.8B | $3.5B | $2.0B | $1.5B | $2.1B |
| Q2 2025 | $14.7B | $5.7B | $3.9B | $2.2B | $1.5B | $2.4B |
| Q3 2025 | $14.9B | $6.2B | $4.1B | $2.3B | $1.6B | $2.6B |
| Q4 2025 | $15.3B | $6.4B | $4.2B | $2.5B | $1.7B | $3.0B |
| Q1 2026 | $15.8B | $7.6B | $4.1B | $2.7B | $1.8B | $3.4B |
| Q2 2026 | $17.3B | $8.8B | $4.5B | $3.0B | $2.0B | $3.9B |
Growth is broad; margins are not
The fastest growers over their latest four quarters were Credo (+165%), Astera Labs (+98%), Lumentum (+83%) and Amphenol (+54%). Most of them are small companies selling a component that every new AI server now needs, so their growth tracks the number of accelerators shipped.
Margins tell a different story from growth. Astera Labs (75%) and Credo (67%) earn chip-company gross margins, because they design specialized silicon. Arista earns 63% on switches, helped by software. The optical component makers earn far less: Coherent 37%, Lumentum 42%, Ciena 44%, because optics are manufactured in volume, priced competitively and bought by a small number of very large customers. Amphenol, at 38%, grew revenue +54% over its latest four quarters on the sheer quantity of connectors in each AI rack.
Across the layer, revenue grew from $97.3B to $141B in two years while the gross margin stayed at about 53%: the extra revenue came in at roughly the margin the layer already earned. Networking has benefited from AI without the margin expansion seen in chips and memory.
Revenue growth by company, latest four quarters vs the four before
Change in revenue from the year before
Source: company 10-K and 10-Q filings (SEC XBRL); fiscal quarters aligned to the nearest calendar quarter.
Show the numbers
| Company | Revenue | Change | Gross margin |
|---|---|---|---|
| Credo | $1.6B | +165% | 67% |
| Astera Labs | $1.2B | +98% | 75% |
| Lumentum | $3.0B | +83% | 42% |
| Amphenol | $29.0B | +54% | 38% |
| Arista | $10.5B | +33% | 63% |
| Ciena | $6.0B | +33% | 44% |
| MACOM | $1.2B | +28% | 57% |
| Coherent | $7.1B | +23% | 37% |
| Corning | $17.7B | +19% | 36% |
| Cisco | $63.3B | +12% | 65% |
A few enormous customers
The buyers of AI networking are the same handful of companies that buy everything else in the stack, and they buy networking in very large, lumpy orders. For a small supplier, one hyperscaler adopting its product for a new accelerator generation can double revenue within a year, and one hyperscaler choosing a different design can take much of it away. That is the main reason the small-cap networking names move so violently on earnings.
It also explains the margin pattern. A supplier selling a specialized chip that a hyperscaler has designed into its servers has some protection until the next design cycle. A supplier selling a module that two or three competitors can build to the same standard does not, and the buyer will use its size to push prices down. When you read a networking company’s 10-K, look at its customer concentration disclosure (any customer above 10% of revenue must be reported) alongside its gross margin. High concentration with rising margins means the customer needs the product; high concentration with falling margins means the customer is in charge.
The technology shifts to watch
Ethernet versus InfiniBand. Nvidia sells its own InfiniBand and Ethernet networking, which competes with Arista and Cisco for back-end networks. The industry has been moving toward Ethernet for AI clusters, which favors the independent switch makers and Broadcom’s switch chips.
Copper versus optics. Inside a rack, copper is cheaper and uses less power, which is why active electrical cables have grown so fast. As speeds rise, the distance copper can carry a signal shrinks, pushing more links to optics. Co-packaged optics, which put the optical engine next to the switch chip, could eventually reshape who captures the value in optical components. Each transition creates winners and losers among the suppliers on this page.
The companies, side by side
Latest four reported quarters for US filers; latest fiscal year for 20-F filers (marked). Underlined names have their own company page; the others link to the company’s filings on SEC EDGAR. Margins are left blank where a company doesn’t report gross profit or cost of revenue.
| Company | Revenue | vs year before | Gross margin | Operating margin | Free cash flow |
|---|---|---|---|---|---|
| Cisco CSCOSwitching, routing and silicon sold to hyperscalers | $63.3B | +12% | 65% | 24% | $12.8B |
| Amphenol APHConnectors and cable assemblies inside racks | $29.0B | +54% | 38% | 27% | $4.7B |
| Corning GLWOptical fiber and cable | $17.7B | +19% | 36% | 16% | $2.5B |
| Arista ANETEthernet switches for AI back-end and front-end networks | $10.5B | +33% | 63% | 43% | $5.2B |
| Coherent COHROptical transceivers and lasers | $7.1B | +23% | 37% | — | −$1.0B |
| Ciena CIENOptical transport between data centers | $6.0B | +33% | 44% | 12% | $814M |
| Lumentum LITELasers and optical components for transceivers | $3.0B | +83% | 42% | 17% | $300M |
| Credo CRDOActive electrical cables and SerDes connectivity | $1.6B | +165% | 67% | 32% | $439M |
| Astera Labs ALABPCIe/CXL retimers and switches inside AI servers | $1.2B | +98% | 75% | 23% | $277M |
| MACOM MTSIAnalog and optical chips for high-speed data-center links | $1.2B | +28% | 57% | 18% | $204M |
Revenue and growth. Total revenue over the four quarters, and its change from the four quarters before. Fast growth with a flat margin usually means volume; fast growth with a rising margin usually means pricing power.
Gross margin. The share of revenue left after the direct cost of what was sold. It is the best single measure of how much of each dollar a company keeps rather than passes on to its own suppliers.
Operating margin. What is left after research, sales and overhead too. The gap between gross and operating margin shows how much the company spends to stay competitive.
Free cash flow. Operating cash flow minus capital spending: the cash the business generated after paying to maintain and grow itself. Negative free cash flow during fast growth means the growth is consuming cash.
Where the money comes from and goes
This layer is paid by AI clouds and servers. Its own spending flows to chips, foundry & tools. For the dollar amounts and who keeps them, see following the money.
What to watch in the next filings
- Gross margins at the optical companies: rising margins would signal real scarcity rather than volume.
- Arista’s and Cisco’s AI back-end network wins, and Nvidia’s networking revenue.
- Customer concentration in the small-cap suppliers: a single hyperscaler can be a large share of revenue.
- Design changes in new accelerator generations that move links from copper to optics or the reverse.
- Hyperscaler capex guidance, which sets this layer’s volume.
Free · quarterly · from SEC filings
Get the AI Stack Report every quarter
After each Big Tech earnings season we update the map from the new filings and publish what changed: who spent, who got paid, who kept it and what the filings reveal about the loop. Free, in Rob May’s Investing in AI newsletter.
Form not showing? Subscribe on Substack. No spam; unsubscribe any time.
Questions about networking
Why does AI need so much networking?
Training and serving large models splits work across thousands of accelerators that must exchange data constantly. A slow network leaves expensive GPUs idle, so AI data centers build a dedicated high-speed back-end network between accelerators on top of the usual front-end network.
Which AI networking companies have the highest margins?
Those that design specialized chips or sell software with hardware: Astera Labs (75% gross margin), Credo (67%) and Arista (63%). Optical component makers earn lower margins because their products are manufactured in volume for a few large buyers.
What are optical transceivers?
Modules that convert electrical signals into light and back so data can travel over optical fiber. AI clusters use them for links longer than copper can carry at high speed, and they are bought in very large quantities.
What is an active electrical cable?
A copper cable with small chips at each end that clean up and boost the signal, letting copper carry very high-speed data a few meters between servers and switches. It is cheaper and uses less power than optics over short distances. Credo, the main supplier, grew revenue +165% over its latest four quarters.
Is Ethernet replacing InfiniBand in AI data centers?
The industry has been moving toward Ethernet for AI back-end networks, though Nvidia sells both. A shift to Ethernet tends to benefit independent switch makers such as Arista and Cisco and switch-chip suppliers such as Broadcom.
Research and education, not investment advice: nothing here is a recommendation to buy or sell any security. Figures come from companies’ SEC filings as reported; companies named were chosen because they show how the stack works, not as a view on their shares.