Wiring the AI Factory: An Investor’s Guide to Networking
Primer, July 22, 2026, updated July 26, 2026. The AI buildout seen through the network, not the GPU. Written for investors who don’t speak fluent semiconductor. Brought to you by SignalDeck.live.
Why this matters: A GPU that can’t talk to the other 99,999 GPUs next to it is just an expensive space heater. As AI data centers scale from thousands to hundreds of thousands of chips, the wires connecting them stop being an afterthought and become one of the biggest line items in the budget. This is the “other AI trade”: companies that get paid no matter which AI model wins, and no matter which chip wins, because everything still needs to be wired together.
The one idea to remember: there are three “boundaries” in an AI data center, and almost every company below sits at exactly one of them.
Inside one cabinet of computers → copper wire (cheap, cramped, blazing fast).
Cabinet to cabinet, within one building → fiber-optic cable (light, not electricity).
Building to building, across a campus or a city → higher-grade fiber-optic cable (“coherent optics”).
Everything in this primer is a variation on those three lines.
Plain-English cheat sheet
Skip this if you already know the jargon. Otherwise, read this once and the rest of the primer will make sense.
GPU: the AI chip (NVIDIA’s specialty). A cluster wires tens of thousands of them together.
Rack: a cabinet holding dozens of GPUs, cabled together so software treats them as one giant chip.
Scale-up: the wiring inside one rack. Shortest distance, highest speed, runs on copper.
Scale-out: the wiring rack-to-rack, stitching thousands of racks into one training cluster. This is the big commercial battleground.
Data-center interconnect (DCI): the wiring building-to-building or site-to-site, needed once one building can’t physically fit enough power.
Transceiver: the small plug on the end of a fiber-optic cable that converts an electrical signal into light and back. Every GPU cluster needs a lot of these.
EML laser: the actual light bulb inside a transceiver. The scarce, hard-to-manufacture part where the real bottleneck lives.
Optical DSP: a chip inside the transceiver that cleans up the light signal so it arrives readable.
Retimer: a small chip that “boosts” a copper signal so it can travel a little farther before it needs to switch to fiber.
InfiniBand: NVIDIA’s own proprietary networking technology. One vendor, one company benefits.
Ethernet: the open, industry-standard networking technology that (almost) everyone who isn’t NVIDIA is rallying around.
NVLink: NVIDIA’s proprietary in-rack wiring. The one layer the open-standards crowd hasn’t been able to crack.
Co-packaged optics (CPO): a next-generation design that builds the light-conversion hardware directly onto the switch chip, skipping the separate transceiver plug. Still early.
Live Snapshot (as of market close 7/24/26)
What each company actually does, in one line, plus where it sits on the map. Worth knowing going in: the whole cohort sold off sharply on July 24, the last trading session before this primer. Astera Labs -11%, Credo -10%, Coherent -10%, Lumentum -8%, Fabrinet -8%, Celestica -9%, Marvell -7%, all in one day, while Cisco (+1%) and Arista (-1%) barely moved. Reporting on the move points to profit-taking and a crowded trade unwinding, not any company-specific bad news, guidance cut, or lost customer. We’re not going to tell you where it goes from here; we don’t do that. But notice the pullback landed almost exactly along the “torque vs. durability” fault line this primer draws in Part 8 below: the highest-beta names fell hardest, the steadiest names barely blinked.
Layer 1, inside the rack (copper):
Credo (CRDO): $213, $40B cap, fwd P/E 35. Makes the copper cables that link GPUs inside a rack. Revenue growth has been explosive (+206% this fiscal year), but ~90% of it comes from its top 10 customers.
Astera Labs (ALAB): $292, $50B cap, fwd P/E 86. Makes the “signal booster” chips that let copper cables run a little farther. Richest valuation in this whole primer. Still trades above the average Wall Street price target (~$243-270).
Amphenol (APH): $153, $188B cap, fwd P/E 30. The boring, diversified giant that makes connectors and cable for practically everything, not just AI. Cheapest valuation of the copper trio.
Layer 2, rack to rack, inside one building (fiber optics + switches):
Broadcom (AVGO): $382, $1.82T cap, fwd P/E 24. Makes the merchant switch chips that move data between racks, sold to nearly everyone who isn’t building their own.
NVIDIA (NVDA): $207, $5.01T cap. Makes the GPU and competes in networking with its own switches and cables. The only company that plays at every layer of this map.
Arista (ANET): $174, $219B cap, fwd P/E 46. Builds the branded switch boxes (using Broadcom’s chips, mostly) that plug into a data center.
Cisco (CSCO): $114, $450B cap, fwd P/E 24. The old-guard networking company, now competing for AI switch sales. Cheapest valuation in the entire primer.
Marvell (MRVL): $194, $170B cap, fwd P/E 43. Makes the chip inside a transceiver that cleans up the light signal (see “optical DSP” above), plus custom AI chips.
Coherent (COHR): $282, $55B cap, fwd P/E 38. Makes the transceivers themselves and, critically, owns its own laser factory.
Lumentum (LITE): $763, $59B cap, fwd P/E 48. Also makes the lasers and chips inside transceivers. This year’s biggest single stock move in the group: +130% at its peak before the recent pullback.
Fabrinet (FN): $476, $17B cap, fwd P/E 29. Assembles transceivers under contract for the bigger brands (NVIDIA, Cisco, Coherent). Cheapest multiple among the optics names.
Celestica (CLS): $305, $35B cap, fwd P/E 27. Builds AI hardware for hyperscalers under contract (“white-box” manufacturing), a lower-margin business than the branded names above it.
China-listed peers, for context only (not easily accessible to most US retail investors): Innolight (~$175B cap) and Eoptolink (~$119B cap) are the two largest transceiver makers in the world by unit volume.
Layer 3, building to building (coherent optics):
Ciena (CIEN): $391, $55B cap, fwd P/E 50. The clearest pure-play on wiring data centers together across a city or region.
1. An AI data center is three separate networks, not one
Picture a big AI data center. You probably picture one giant network. It’s actually three physically separate ones, layered on top of each other, and a different set of companies wins at each layer.
Layer 1, inside the rack. The fastest, most fused layer. The goal is to make a whole cabinet of GPUs behave like one enormous chip. NVIDIA’s technology here, NVLink, moves 1.8 terabytes per second per GPU, roughly a thousand times faster than a typical home internet connection. A full rack of 72 GPUs pushes 130 terabytes per second between them. This is NVIDIA’s deepest moat: nobody else is close, and it runs on plain copper wire (more on why in Part 5).
Layer 2, rack to rack. Connects thousands of racks into one training cluster. This is the big commercial fight: NVIDIA’s own InfiniBand technology versus the open-standard Ethernet everyone else is building around. The entire merchant supply chain (every company that isn’t NVIDIA) lives or dies here.
Layer 3, the plumbing. Storage, data loading, day-to-day management traffic. Plain old Ethernet, not AI-specific, not where the money or the fight is. We won’t mention it again.
2. The central fight: NVIDIA’s InfiniBand vs. open-standard Ethernet
InfiniBand is NVIDIA’s own technology: one vendor, purpose-built, extremely fast, extremely reliable. It won the early years of the AI boom because, for training runs, it simply worked, and NVIDIA has been the only company selling it since it bought the company that made it (Mellanox) back in 2019.
Ethernet is the open alternative. Practically every other major tech company (AMD, Arista, Broadcom, Cisco, Meta, Microsoft, Oracle, and more) has been re-engineering standard Ethernet specifically to handle AI traffic. A newer, lighter version called MRC was unveiled in May 2026 by OpenAI, Microsoft, Broadcom, AMD, and, notably, NVIDIA itself.
The single data point that tells the whole story: in the first quarter of 2026, sales of AI-grade Ethernet switches rose 61% year-over-year to more than $10 billion, and the #1 vendor was NVIDIA, at $2.1 billion (nearly 3x the year before), ahead of Arista and Cisco. When the company that invented the proprietary alternative is also the biggest seller of the open alternative, that tells you which way the industry is leaning.
The current split: roughly two-thirds Ethernet, one-third InfiniBand for this rack-to-rack layer, though InfiniBand itself grew more than 3x year-over-year in the same quarter. This is a slow grind between the two, not a rout of either side.
Why Ethernet is gaining ground: it isn’t locked to one vendor, and it’s a more efficient network design that needs roughly a third fewer of the expensive transceiver plugs to reach the same cluster size. Bluntly, every large tech company also wants an alternative to depending entirely on NVIDIA.
3. The one layer NVIDIA isn’t losing: inside the rack
Going back to Layer 1: this is the hardest layer for the open-standards camp to crack, because it’s the highest-speed, most tightly-integrated connection in the building.
NVLink (NVIDIA’s proprietary version) is shipping today, dominant, and roughly twice as fast as what the open competitor has even specified on paper. NVIDIA also made a smart move called “NVLink Fusion”: instead of losing customers who want to use non-NVIDIA chips, it now lets those chips plug into NVIDIA’s own fabric. It’s absorbing the defectors rather than fighting them.
UALink is the open answer, backed by AMD, Broadcom, Google, Intel, Meta, Microsoft, Apple and Amazon, with version 2.0 published in April 2026. It’s real and broadly supported, but it’s still an early-stage specification chasing a moving target, while NVLink is already shipping and roughly twice as fast.
4. Fiber optics: the layer with the most direct AI exposure
If you want the single most AI-levered piece of this whole map, it’s the fiber-optic layer connecting racks together. Two reasons:
Every GPU needs several of these transceiver plugs on the rack-to-rack network. Estimates run anywhere from about 2.5 to 9 plugs per GPU depending on the design, so think of it as a wide range, not a fixed number. Whatever the true number, it multiplies straight through with GPU count.
The industry is mid-upgrade. “800G” transceivers are today’s standard; a faster, roughly-double-speed “1.6T” version is ramping through 2026. That’s a price increase stacked directly on top of unit growth, about the best combination a components business can ask for.
The part most headlines miss: the bottleneck isn’t the transceiver plug itself. It’s the tiny laser inside it. Specifically, a next-generation laser component needed for the faster 1.6T speed. Building the plug around it is a commoditizing, thin-margin business. Manufacturing that specific laser is not: capacity sits with a short list of companies, Coherent, Lumentum, and Japan’s Sumitomo and Mitsubishi. That’s where the durable profit margin actually lives.
Coherent is the leading Western transceiver maker, and its edge is that it owns its own laser factory, the best-positioned company on securing 1.6T supply.
Lumentum is more focused on the laser and chip itself, and has been this year’s biggest single move in the group (+130% at its peak, before the recent pullback described above).
Fabrinet assembles transceivers under contract for the bigger brands. Cheapest valuation in the group, but its revenue leans heavily on a small handful of customers.
On raw unit volume, two Chinese companies, Innolight and Eoptolink, actually lead the world (not easily investable for most US retail accounts).
There’s also a chip inside every transceiver called an “optical DSP” (it cleans up the light signal). Marvell is #1 in this chip, Broadcom #2, and together they own most of that market. Marvell shipped the first 1.6T version of this chip, and in March 2026 NVIDIA took a roughly $2 billion stake in Marvell, effectively buying a seat inside the leading maker of this component.
The one real technology risk here: a newer approach called “linear-drive optics” removes that DSP chip entirely, cutting power use by 30-50%. It’s a genuine threat to DSP demand, but the consensus view is it only works well on shorter cable runs (the physics gets hard at the faster 1.6T speed), so it likely takes a slice of the older, slower market rather than replacing the DSP wholesale. The DSP makers sell other components too, so they’re hedged either way.
5. Why the fastest wire in the building is still just copper
Here’s the question from the top of this primer: why does the fastest connection in an AI data center, the one inside the rack, run on plain old copper wire instead of fiber optics?
Physics. The rule is simple: every time you double the signaling speed, copper’s usable range cuts in half. At today’s speeds, a basic copper cable reaches about one meter; an “active” copper cable (with a small signal-boosting chip built into the connector) reaches about two-and-a-half to three meters. That one-to-three-meter window happens to be almost exactly the inside of one rack. Go any farther, and you have to switch to fiber optics.
That short-range cutoff isn’t a footnote. It’s the entire business model for several companies. NVIDIA’s own flagship rack contains roughly 5,000 copper cables (about two miles of wire) and zero fiber optics inside the rack. That’s a deliberate design choice: running fiber optics inside the rack at these speeds would be less reliable, use more power, and add tens of kilowatts of extra heat.
Credo is the pure-play on these “active” copper cables. Growth has been explosive (+206% revenue this fiscal year), but roughly two-thirds of its revenue comes from a single customer, and about 90% from its top ten. One paused program at one customer is a real risk to the stock.
Astera Labs makes the signal-boosting chips (and related switches) that let copper cables run farther, plus a broader platform spanning several competing standards, meaning it collects a small toll regardless of which specific in-rack technology eventually wins. That hedge is part of why it carries the richest valuation in this primer.
Amphenol is the diversified giant: connectors, cable, and backplanes for practically every industry, not just AI. Its AI/data-center segment is growing over 80% a year organically, but AI is a smaller slice of a much larger, steadier business, which is why it carries the lowest single-company risk of the three.
6. The next battleground: building the transceiver into the chip
What it is: instead of a separate plug-in transceiver, move the light-conversion hardware directly onto the switch chip itself, to save power. NVIDIA and Broadcom both have versions of this in development.
The realistic timeline: this is a 2027-and-later story, not a 2026 one. 2026 is just the first, small-scale deployments. The traditional plug-in transceiver stays in more than 90% of ports for the foreseeable near-term. This technology adds to the market before it replaces anything in it. That’s the key timing risk worth watching for Coherent, Lumentum, and Fabrinet, but it’s further out than the bears on those stocks tend to imply.
Meanwhile, there’s already a land grab happening around the next frontier: putting this same optical technology to work inside the rack (not just between racks). Marvell is acquiring a company called Celestial AI for about $3.25 billion to get into this space; two private companies, Ayar Labs and Lightmatter, are the other names to know here (not publicly investable yet).
7. Building-to-building wiring: the least-known layer with real growth behind it
The new demand driver: individual data-center buildings are running into power limits. You simply can’t get more electricity to one building. So operators are stitching multiple buildings into a single logical AI supercomputer across a metro area or region. This is the least well-known part of this whole map to most investors, and arguably the most underappreciated.
Who benefits: companies making “coherent” optics, a higher-grade, longer-range version of fiber-optic technology built for exactly this job. Ciena is the clearest pure-play, with revenue up 40% year-over-year last quarter, and industry reporting suggests demand for this equipment is currently outstripping supply. Dell’Oro, a research firm that has tracked the router market for telecom companies for twenty years, now says demand for its highest-end routers is “increasingly influenced by datacenter connectivity, less by telecom.” That’s a genuinely unusual thing for that industry to be saying.
8. How to think about this whole map: torque vs. durability
Ranking these companies by how AI-exposed they are (torque) and how durable that exposure is, because torque alone is how people get hurt buying the wrong thing at the wrong price.
Tier 1 (cleanest combination of AI exposure and durability): switch chips and switch systems. Broadcom (the chips) and Arista (the branded boxes built on those chips). Why they’re the cleanest: they win money regardless of which optics technology, or even which fabric (InfiniBand vs. Ethernet), ultimately wins. They’re not betting on any single technical outcome. They’re selling the thing in the middle, no matter who’s on either end of it.
Tier 2 (biggest dollar growth, with a real bottleneck behind it): the laser makers. Coherent, Lumentum. The transceiver plug itself is becoming a commodity; the laser inside it is the actual scarce resource. Worth noting: Lumentum’s stock move this year has already priced in a lot of the good news.
Tier 3 (highest growth, but also the most fragile): copper cables and signal-boosting chips. Credo, Astera Labs. The biggest percentage revenue growth in this whole primer, but also the most extreme reliance on just a handful of customers, and the richest valuations. This is the highest-risk, highest-reward corner of the map in both directions.
Tier 4 (the steadier, lower-risk corner): Cisco (cheapest valuation here, AI is a bonus on top of an already-large business) and the building-to-building names like Ciena, which are later in the buildout cycle and more diversified.
The trap to watch for: low-margin contract manufacturers that have re-rated purely on “AI” branding without the profit structure to back it up, and any single stock trading well above where the analysts covering it think it should be. Those tend to be first in line to fall hardest if hyperscaler spending ever pauses.
The debates worth knowing about (not predictions, just the live disagreements)
Does NVIDIA’s full package just win everything? GPU, plus its own in-rack wiring, plus its own switches: that’s the most tightly bundled, highest-margin version of this whole stack. The open-standards coalition is betting that big tech companies will pay a little more, or accept a little less peak performance, in exchange for not being locked into one vendor. Which side wins the rack-to-rack layer specifically is genuinely unsettled.
Customer concentration. Several of the highest-growth names here (Credo, Arista, Fabrinet, Astera Labs, Celestica) depend heavily on just two or three giant tech companies as customers. The same concentration that gives these stocks their torque is also their biggest single vulnerability. One customer’s paused project is a real risk, not a theoretical one.
The “good enough, cheaper” threat. Big tech companies increasingly build their own switches using merchant chips (mostly Broadcom’s) and free, open-source software, skipping the branded box entirely. That specifically pressures Arista and Cisco’s most profitable business at their very largest customers.
Capex digestion. Every valuation in this primer assumes big tech company spending on AI infrastructure keeps climbing, not just holding flat. A real pause in that spending would hit the most customer-concentrated, most richly-valued names (Astera Labs, Credo, Lumentum, Fabrinet) hardest.
Bottom line
Networking is the “other AI trade”: companies that get paid no matter which AI model or which chip wins, because everything still has to be wired together. The entire map is just three boundaries: copper inside the rack, fiber optics rack-to-rack, and higher-grade fiber optics building-to-building.
The steadiest, most durable exposure is switch chips and switch systems (Broadcom, Arista). They collect revenue regardless of which specific technology path wins. The highest AI-exposure is in fiber optics, specifically the laser bottleneck (Coherent, Lumentum), where a faster/pricier upgrade cycle is stacking on top of plain unit growth.
The rack-to-rack layer is drifting toward the open Ethernet standard. Even NVIDIA’s own fastest-growing networking product is now its own Ethernet platform. But NVIDIA’s proprietary alternative isn’t dead, and it still owns the inside-the-rack layer outright.
The highest-growth names (Credo, Astera Labs) also carry the richest valuations and the least customer diversification: maximum upside, maximum fragility. The July 24 selloff is a live example of what that combination looks like when sentiment turns, even with no bad company news attached.
Three things worth watching going forward: whether the rack-to-rack layer keeps drifting from InfiniBand toward Ethernet, when “chip-integrated” optics actually arrives at volume (still tracking to 2027+), and whether big tech companies’ spending guidance keeps climbing, because every valuation in this primer rides on that last one.
Educational primer, not financial advice. For informational purposes only. Do your own research. Prices, market caps, and forward P/E ratios as of market close 7/24/26 (stockanalysis.com); re-verify before acting on anything more than a few days old.