This group sells the switches that connect thousands of AI chips (GPUs — graphics processing units, the chips that do AI math) to each other inside a data center, plus the switch chips (ASICs — application-specific integrated circuits, custom silicon designed for one job) inside those switches. A modern AI training cluster is not one big computer; it is tens of thousands of GPUs that must talk to each other constantly at enormous speed. The "wiring" that lets them do that is the switching fabric. There are two competing flavors: InfiniBand (a specialized, low-latency network — meaning very short delays — that NVIDIA controls) and Ethernet (the open, industry-standard network being upgraded for AI via a standard called "Ultra Ethernet"). The unit you are really buying exposure to is a high-speed switch port — one connection slot running at 400 gigabits, 800 gigabits, or soon 1.6 terabits per second. The makers are Arista (ANET), Cisco (CSCO), NVIDIA (NVDA), and Broadcom (AVGO), the last two of which supply the chips that the whole industry's switches depend on.
On the facts in the scan, demand for high-speed switch ports is growing faster than the supply of the chips that make them: every new GPU cluster needs far more switching than old cloud workloads, and larger clusters need switching that grows faster than the GPU count. The named bottleneck is not bending sheet metal into a switch — it is the supply of the switch ASIC (the custom chip), which two companies dominate and which competes for the same advanced chip-factory capacity as the GPUs themselves. In money terms, the leaders trade at high multiples of current revenue and profit (see "The price of exposure" for the arithmetic), so a large amount of future growth is already reflected in today's price. The reader can weigh the demand-supply gap against what that exposure costs today.
A data-center switch is a box, usually 1-2 rack units tall, full of ports. Cables (copper or optical) plug into those ports, and the switch's job is to move data packets from any port to any other port at "line rate" (full speed, no slowdown). In an AI cluster the switches are arranged in a layered tree (called "leaf-and-spine" or "Clos" — a standard network design): leaf switches sit near the GPU servers, spine switches sit above them and tie the leaves together. The more GPUs you add, the more layers and the more switches you need — and the count grows faster than the GPU count because every GPU needs multiple high-speed links for the cluster to stay balanced.
The single most important component inside the box is the switch ASIC — a custom chip that does the actual packet forwarding. Today's top chips push 51.2 terabits per second; the next generation roughly doubles that to 102.4. Whoever makes the best ASIC effectively sets the pace for the whole industry. Two business models exist on top of that chip:
So an owner of this group is really buying two stacked profit pools: the chip at the bottom, and the branded-system-plus-software layer on top. The same company can play in both (NVIDIA and Broadcom sell chips; NVIDIA also sells whole InfiniBand and Ethernet switch systems).
Today (known/observed, from the scan): AI clusters use "far more switches per server than traditional workloads," a 100,000-GPU cluster "can need 10,000+ switches," and "the AI switching market is growing 40%+ annually." The architecture is built so that switching demand scales faster than linearly with cluster bandwidth — doubling cluster size more than doubles the ports and switch chips required.
Why demand keeps rising (AGI lens): given that AGI is arriving, the binding input is compute, and compute is being assembled into ever-larger clusters. Each more-capable model generation trains on a bigger cluster than the last; recursive self-improvement (models helping design the next models) pushes cluster sizes from today's ~100k-GPU range toward the 1M+ GPU designs already being discussed. Because interconnect cost and complexity grow faster than linearly with cluster size (stated plainly in the scan), the share of total AI build-out spend going to networking rises as clusters get bigger. On top of training, the shift toward heavy inference (running already-trained models) and "physical AI" (robots and autonomous systems pulling from cloud models) adds a second, broader demand stream. In dollar terms the AI portion of switching is the fast-growing slice; the legacy enterprise/campus switching market underneath it is comparatively flat. forecast
Who the buyers are: a small set of very large buyers — the hyperscale cloud companies (Microsoft, Google, Amazon, Meta), the new "neocloud" GPU-rental operators (e.g. CoreWeave), large AI labs, and sovereign/enterprise AI projects. This concentration matters two ways: a handful of purchase decisions move the whole market, and several of the biggest buyers (Google, Meta, Amazon) design their own switches and buy chips directly — which is demand for Broadcom's silicon even when it is not demand for Arista's or Cisco's branded boxes.
Forecast vs fact: the "10,000+ switches per 100k-GPU cluster" and "40%+ annual growth" figures come from the provided scan and reflect its analysis, not contracted orders. Specific multi-year market-size dollar forecasts are est. and should be read as direction, not precision.
✓ VERIFIED — the following figures were confirmed from primary sources after initial publication:
Assembling the switch is not the bottleneck. The scan is explicit: switch assembly can scale, and white-box (unbranded) makers "can ramp faster than branded vendors." Sheet metal, fans, and circuit boards are not scarce.
The real constraint is the switch ASIC. The scan states the bottleneck is "switch ASIC supply (Broadcom, custom chips) … more than switch assembly." Those chips are designed by a tiny number of firms and fabricated at the same leading-edge foundries (chip factories, chiefly TSMC) that make the GPUs — so switch chips compete for the very capacity AI demand is already straining. Designing a 51.2T-and-beyond chip is a multi-year, deep-expertise effort; a new credible supplier cannot appear quickly. For InfiniBand specifically, the scan notes NVIDIA's "InfiniBand switches and cables have been in allocation alongside GPUs" — i.e. rationed — and that the open-standard Ultra Ethernet alternative is "18-24 months behind." A second-order limit on the optical-Ethernet side is the supply of high-speed optical transceivers (the parts that convert electrical signals to light for fiber) and high-density fiber connectors that plug into the ports (covered as separate constrained categories in the scan), which gates how fast ports can actually be lit up.
Market-share structure (who controls supply):
Capacity expansion is happening — more chip wafer allocations, faster product cadence (the scan notes the move to 51.2T and 102.4T is being "accelerated by AI") — but it is gated by foundry capacity and design talent, not by factory floor space. That is why supply stays tight even as everyone tries to add it.
Putting the two sides together: on the scan's evidence the product is short today — demand for high-speed switch ports is outrunning the supply of the chips that make them. The scan's direct evidence: InfiniBand gear "in allocation" (rationed), Broadcom's custom interconnect ASICs "capacity-constrained," and the AI switching market growing 40%+ a year. Lead times on the scarcest pieces (high-density connectors, top-end optics that pair with the ports) are extended.
| Signal | What it shows | Direction |
|---|---|---|
| InfiniBand allocation | NVIDIA rations switches/cables alongside GPUs | Short |
| Switch-ASIC capacity | Chip supply is the named bottleneck; shares foundry with GPUs | Short |
| AI switching growth | ~40%+ per year est. outruns supply additions | Short |
| Switch assembly / white-box | Can ramp fast — not the constraint | Ample |
| Ultra Ethernet (open standard) | ~18-24 months behind; new supply still maturing | Lagging |
When could it flip to oversupply? Three things would loosen it, and on the scan's timeline none is imminent: (1) the open Ultra Ethernet ecosystem matures and lets many vendors ship interchangeable high-end Ethernet ports (the scan puts this ~18-24 months out, so roughly 2027-2028 before it bites est.); (2) leading-edge foundry capacity catches up enough that switch ASICs stop competing with GPUs for wafers; (3) AI cluster build-out slows. Under the AGI premise, (3) looks least likely — the scan frames the demand pressure as structural, not cyclical. The more plausible loosening comes from (1) and (2), which would tend to compress pricing power on commodity ports first, while the highest-end chips and the software-locked branded systems would tend to stay tight longer. forecast
| Company | What it makes here | Exposure to switching | Rough size | Position / role |
|---|---|---|---|---|
| Arista (ANET) | High-end data-center Ethernet switches + EOS software | Pure-play — essentially all revenue is data-center networking est. | ~$15B+ rev est. | Share leader in AI/cloud back-end Ethernet (est.); software stickiness; sells heavily to hyperscalers |
| NVIDIA (NVDA) | InfiniBand (Quantum) + Ethernet (Spectrum-X) switches & the chips; NVLink/NVSwitch inside the GPU rack | Small slice of a giant — networking is a minority of NVDA revenue; GPUs dominate est. | Mega-cap (GPU-led) | Near-monopoly in InfiniBand; bundles networking with GPU sales; fast-rising in AI Ethernet |
| Broadcom (AVGO) | Merchant switch ASICs (Tomahawk/Jericho) + custom interconnect chips for hyperscalers | Important but diversified — switching is one of several large chip lines plus a big software arm est. | Mega-cap | Dominant merchant switch silicon (est.); the chip inside many rivals' boxes; gates industry supply |
| Cisco (CSCO) | Branded switches (Nexus) + NX-OS software; routers | Diversified incumbent — switching is a major segment but blended with security, software, services est. | ~$50B+ rev est. | Largest installed base; strongest in enterprise/campus; pushing into AI back-end |
| Juniper (now in HPE) / Marvell (MRVL) | Juniper: branded switches. Marvell: interconnect PHYs (physical-layer chips) / custom-silicon help | Smaller / component roles est. | Mid-to-large | Juniper a #4 system vendor (now HPE-owned); Marvell a custom-ASIC and SerDes (high-speed signaling) supplier |
Source: 500-stocks scan, sections "InfiniBand & High-Speed Interconnect" and "Network Switches & Routers" (key US-listed companies: ANET, CSCO, JNPR, AVGO, NVDA, MRVL); company sizes/shares are est. from general knowledge, not live-verified.
In plain money terms, the market pays very different prices for these names, and they have very different money-in/money-out shapes. All multiples below are est. and not live-verified — confirm against current filings.
The money-in/money-out shape of the group as a whole: this is a capital-light group. Unlike the data-center builders or chip foundries, these companies do not pour tens of billions into plant and equipment (capex = capital expenditure, money spent on long-lived physical assets); they design and outsource manufacturing. That means a high share of revenue becomes owner cash. The price reflects a trade-off: for the high-growth names, the high multiple means many years of the demand-supply gap are already reflected in today's price; for the incumbent (Cisco), the low multiple corresponds to less expected AI upside. These are descriptions of what each dollar of exposure costs and what cash it currently throws off — the reader draws any conclusion.
The most informative single comparison for an owner is ANET vs CSCO (pure-play high-growth vs diversified incumbent) layered against AVGO/NVDA as the upstream chip suppliers that both depend on.
/Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/500-stocks/03-networking-connectivity.html. This supplies the company list (ANET, CSCO, JNPR, AVGO, NVDA, MRVL), the "10,000+ switches per 100k-GPU cluster," "40%+ annual growth," the 51.2T/102.4T chip generations, the InfiniBand-allocation note, and the "ASIC is the bottleneck" framing./Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/reports/research/industries/data-center-networking.html does not exist in this worktree (the entire reports/ directory is absent), so no V1 dated numbers could be folded in. If a V1 exists elsewhere, its figures should be reconciled into this sheet later.Hard vs approximate: Hard / grounded — the company list and roles; the qualitative supply-demand notes; InfiniBand is NVIDIA-controlled; Broadcom is the leading merchant switch-silicon supplier; switch assembly is not the bottleneck; the chip is. Approximate / not live-verified (carry the est. tag) — every dollar market-size and TAM figure, the "~40%+" growth as a precise number, all market-share percentages, all per-company revenue figures, and all valuation multiples in "The price of exposure." Treat ranges as direction, not precision, and confirm against current filings before acting.
Source: 500-stocks networking scan (May 2026 vintage); company filings (well-known, not re-pulled); general knowledge (~early 2026, not live-verified).