01 The Trigger Event

On August 12, Cisco reported FY2026 Q4 results: quarterly revenue of $17.3 billion, up 18% year-over-year; GAAP net income of $3.9 billion, up 51% YoY, with EPS of $0.97 (+52%). Full-year revenue reached $63.3 billion, +12% YoY.

Two line items deserve separate attention:

  • Product orders grew +35% YoY in Q4, with networking products +40%, marking the eighth consecutive quarter of double-digit growth
  • AI infrastructure orders hit $4 billion in Q4 alone, totaling $9.3 billion for the full fiscal year; FY27 guidance points to $7.5 billion in revenue

02 What This Really Means

This isn't just "Cisco had a good quarter."

For the past two years, I've been watching Nvidia's data center revenue to gauge GPU shipment cadence—but GPUs can't be sold in isolation. A single H100 consumes 1-2 network ports; an NVL72 rack requires 72 GPUs plus an entire stack of ToR/spine switches and optical modules. Cisco's order book is effectively a dependent variable of GPU shipments and a leading indicator of hyperscaler cluster build-out speed.

Two inferences:

First, hyperscaler AI capex is no longer just GPUs. With $9.3 billion in annual AI infrastructure orders alongside Nvidia's $30-40 billion in annual GPU shipments, the magnitudes are now comparable. This means the optimistic curve I'd been using—"token prices fall as compute scales"—needs a discount, since networking depreciation also flows into token costs.

Second, the AI infrastructure SKU is being independently repriced by capital markets, carved out from traditional networking. Cisco separating AI infrastructure guidance on the earnings call is a deliberate move—they're trying to split "general Cisco" from "AI Cisco" in valuation terms. If the market buys it (and I expect it will), Cisco's P/E gets revised upward, which in turn reanchors pricing for all networking names (Arista, Juniper).

03 Historical Analogy

The closest parallel is the 2013-2015 AWS region-build cycle. Arista Networks was the sole beneficiary of 100G switches back then, with its stock rising six-fold over three years—essentially because hyperscalers concentrated core orders on Arista during their 10G-to-100G upgrade. Arista was the leading indicator for AWS/Azure capex; Cisco's current role is identical, except the customer base has shifted from generic cloud to AI cluster.

An older negative mirror is the 2000 fiber bubble. Ciena, Lucent, and Nortel peaked on those orders and then lost 90%+ within three years. Today's narrative structure—"AI infrastructure is a decade-long supercycle, and networking vendors eat everyone's lunch"—bears a striking resemblance to "the internet is a decade-long supercycle, and fiber vendors eat everyone's lunch." I haven't run an internal analysis on backlog conversion rates, so I won't claim this time is different, but this analogy stays in my mind as a warning sign.

04 What This Means for AI Builders

Short term (6 months): Tokens won't get cheap. The hyperscaler hardware capex cycle is still accelerating, and depreciation costs will enter token pricing models over the next four quarters. Anthropic and OpenAI have zero incentive to price below the loss line voluntarily. My earlier call that "API prices will halve within 12 months"—I'm walking back half of that. Networking-side costs will eat a meaningful share of the price-down room.

Medium term (12-18 months): Networking is the hidden bottleneck for MoE long-context workloads. This generation of switches—Cisco Silicon One and Tomahawk Ultra—determines whether 100K+ context windows hold up under multi-user concurrency without latency degradation. If you're building agent products (like Claude Code or Cursor) that demand long context and high concurrency, your inference TCO will carry a higher networking cost share than the headline number suggests.

Long term: The moat concentrates further toward players who own private fabric. Verticalized GPU clouds like Crusoe and CoreWeave, aiming to push latency below hyperscaler levels, must purchase Cisco or Arista gear directly—which raises the barrier for new entrants. The supply-chain negotiating leverage of token gateway players like opcx.ai is also fundamentally tied to upstream fabric capacity.

05 The Bear Case / Risks

I may be over-reading three things:

Number classification issues. I don't have a detailed SKU breakdown of Cisco's "AI infrastructure" definition—it could include campus switching, Wi-Fi 7, enterprise campus upgrades, not just GPU fabric. The Q4 report listed traditional networking product orders at +40% separately, which suggests the non-AI infrastructure book is also growing. That's actually a mixing signal.

InfiniBand substitution risk. Nvidia Spectrum-X, this generation of Ethernet, is increasingly competitive. If Spectrum-X takes further share from Cisco inside hyperscalers in FY27, Cisco's high growth may not be sustainable. I haven't run an internal TCO comparison between Spectrum-X and Silicon One on GB200/GB300, so I won't call that one.

Capex cliff risk—this is the biggest one. Cisco's orders are evidence that hyperscalers are still spending, not evidence they will continue to spend. Once any single hyperscaler's token business gross margin falls below 30%, capex stops overnight, and Cisco's order model can be revised down 30%+ in a single quarter. Jensen Huang's early-August claim that "AI capex still has several years of strength"—if that flips to a downward revision, Cisco is the first-wave vendor to get cut.

My personal odds on that last point are roughly 50/50. If it happens, gateway players like opcx actually see margin structure improve (upstream GPU prices loosen). But if it doesn't happen, I need to rewrite my downside expectations for token economics. The real value of this earnings report isn't Cisco's valuation—it's adding another independent data point to the AI capex cycle.