From 5% to 18%: Cisco's AI-Driven Growth Leap and the Durability Question

From sleepy giant to record quarter

Three years ago, one number defined Cisco for outside investors: roughly 5% annual growth. At the time, sell-side analysts framed even that modest company target with glass-half-empty caution, writing notes about how such growth could conceivably be achieved. According to Cisco Investor's fourth-quarter results, the picture has flipped: quarterly revenue rose 18% year over year to 17.3 billion dollars, full fiscal 2026 revenue grew 12% to 63.3 billion, and management guides fiscal 2027 revenue to a range of 72.2 to 73.4 billion dollars, close to 15% growth. The comparison guest Sami Badri offers is striking: on a normalized basis, nothing like this has been seen since the dot-com boom.

The prior twelve months had not been slow either; the company was already growing at a low double-digit pace, and the new guide promises acceleration on top of that. The core question host Steve Eisman keeps returning to is simple: how does a company of this size triple its growth rate, and how much of it comes from AI? The answer hides in two engines, one directly tied to AI and the other only partly so.

What Cisco actually sells

The part of the business cameras never see is its essence. In Badri's plain telling, Cisco builds the invisible fabric that carries a message from one phone to another and once connected one university database to another: switches, routers and wireless access points. The overwhelming majority of customers are large enterprises and institutions, a supplier profile whose fate hangs on corporate and cloud spending cycles rather than consumer enthusiasm. That is why the company's destiny is tied to the capital budgets of big organizations and hyperscalers. A silicon program has been layered onto that core over the past decade. Per the Cisco Newsroom announcement of February 2026, the Silicon One G300 delivers 102.4 terabits of switching capacity aimed at gigawatt-scale AI clusters, powering the new N9000 and 8000 series systems. The sample chip Badri holds up on the show is visibly smaller than a graphics processor, because it is a network processing unit , the processing layer of light that moves information from point A to point B. Two claims in the announcement stand out: a 28% improvement in job completion time and nearly 70% energy-efficiency gains with liquid-cooled designs.

Engine one: the doubling hyperscale business

The first engine of the growth story is sales to household-name cloud giants. Badri says this business will nearly double in fiscal 2027; its share is still below 20% of total revenue, yet the momentum comes from there. Cisco Investor's order data sharpens the picture: fourth-quarter AI infrastructure orders from hyperscalers reached 4 billion dollars, bringing the fiscal 2026 total to 9.3 billion. Roughly 60% came from Silicon One based systems and 40% from optics, with revenue from this line expected to climb from about 4 billion in fiscal 2026 to 7.5 billion in fiscal 2027.

Two main products draw from the hyperscaler's connectivity budget: switches carrying the company's own chips to move enormous traffic across data centers, and coherent pluggable optical modules tying fiber and ethernet links into those switches. Per Yahoo Finance's account of the earnings call, networking product orders rose 40%, the eighth straight double-digit quarter; total product orders climbed 35% and enterprise orders 21%. So while hyperscale takes the headline, the order book is swelling broadly.

Engine two: retiring 15-year-old boxes

The second engine looks unrelated to AI at first glance, though Badri argues it partly rides the same wind: the campus refresh cycle. Campus here does not mean a university; it means anywhere people walk, talk, transact and work, from a Manhattan office tower to an airport or a factory. Some switches installed in such places have been running for over 15 years, a situation Badri likens to still living on a 15-year-old phone. The company has now declared end of support, patching and updates for that vintage gear, and customers are answering with rip-and-replace orders.

The scale shows up as record campus orders in the company's own figures, with call disclosures pointing to roughly 20% growth on the campus side. The AI link is indirect but real: modern campus gear is designed for the capacity that agent-based workloads and dense wireless traffic demand. Fiscal 2027's guided 15% growth therefore stands on two legs, AI clusters in the data center and a replacement wave across offices and sites.

The inflection and the Arista gap

When did this story turn real for Cisco specifically? Badri's answer is candid: the sector trend was already flowing, with graphics processor sales and cloud spending long rising; Cisco's own inflection arrived just six to nine months ago, when orders for Silicon One based systems landed far above historical scale. Until then investors had a ready objection: nice silicon story, but too small a revenue slice. With both scale and growth rate changed, the market takes the story more seriously, and the company stresses multiple design wins across different zones of hyperscaler networks. The gap with arch-rival Arista starts in hardware philosophy. Cisco designs its own silicon and puts it in its own switches, while Arista buys merchant silicon and leans on software engineering. In Fierce Network's market analysis, Arista passed Cisco in 2024 to become number one in data center ethernet switching; in the first quarter of 2026 it posted 2.71 billion dollars in revenue, up 35%, and nearly doubled its AI networking target to 3.5 billion. Badri's explanation rests on accounting reality: Arista's revenue mix is indexed almost entirely to the fast-growing hyperscale market, while Cisco's legacy in slower fields like security, collaboration and services drags the average.

Concentration risk: five customers, 70% of receivables

Eisman's toughest question probes the fragile spot in this fine picture: while the planet's largest company grows triple digits, Nvidia's 10-Q filing with the SEC shows five direct customers accounting for 22, 14, 13, 11 and 10% of receivables, fully 70% concentrated in five accounts. On the revenue side, two customers make up about 39% of total sales. With Anthropic and OpenAI's health sitting at the bottom of the food chain, everyone above looks tied to those two names. Eisman's Mercedes analogy sticks: you do not need the priciest, strongest model for every job; Chinese and open-weight models may cover perhaps 95% of tasks far more cheaply.

Badri answers on two levels. First, the two companies' dominance is not a supply conspiracy but the result of end-user token appetite: a mass, accelerating demand for the tokens frontier models produce has pooled traffic at two doors, a gravitational center like a privately owned Google search. Second, how capacity gets financed and built is a separate mechanism; the demand quality metric he tracks is token consumption itself, and it is exploding. OpenRouter figures published via Yahoo Finance show the platform processing 25 trillion tokens a week, five times the level of six months earlier and equivalent to 100 trillion a month.

The skills file: 90% still beginners

The episode's most practical stretch concerns how people actually use AI tools. On his travels Badri asks investors three questions: are you a simple user, do you have a skills file, have you put an agent into production? The answers stall shockingly at step one: roughly 90% of users never move past simple prompting. The skills file idea means uploading an analyst's old reports into an AI harness and distilling a 40-page playbook of what to look for in a stock, which Badri likens to training a new hire. Eisman's hedge-fund son-in-law doubling his research capacity is the living example of that layer.

This picture also explains the division of labor between open-weight and frontier models. Routine workflows go to cheap open models, while intense, competitive teams like silicon engineering pay up for the best intelligence; in Badri's bank-desk analogy, not everyone must be a first-string player, but the sharpest minds must sit in the critical seats. The investment thesis that follows is crisp: power users today are programmers, the next wave will be financial services, and token consumption compounds as each new profession learns the tools. Deloitte's research finds 67% of enterprises already consuming over a billion tokens a month.

Capital intensity and the grid wall

Eisman's second big theme is how onetime cash-printing, capital-light giants turned into capital devourers. According to TechCrunch, Alphabet unveiled an equity sale plan of about 80 billion dollars to fund AI infrastructure, including a 10 billion private placement with Berkshire Hathaway. In CNBC's tally, the company lifted its 2026 capital expenditure ceiling toward 205 billion dollars, turned free cash flow negative for the first time in the second quarter, and more than doubled debt in six months to 98 billion. Amazon and Microsoft carry spending plans around the 200 billion mark and Meta a 139 billion program, all part of the same picture. That spending appetite is hitting a physical wall. Under the audit directive published by the Texas administration, the grid operator's interconnection queue holds 474 gigawatts of requests, more than five times the state's record peak demand, with roughly 90% coming from data centers. New projects cannot connect until they document how much of their own power and water they provide, which tax incentives they take, and who truly owns them. Badri, speaking as a Texan, presents this retreat as evidence that friction has arrived even where permits were fast and power plentiful; his answer on the water debate breaks the script: the water footprint of 38,000 prompts equals the water needed to grow a single almond.

Megawatt racks and the IR lesson

The scale shift reads best in electricity numbers. Where a 100 to 200 megawatt facility made headlines in 2021, gigawatt campuses are now discussed; per-rack power is marching from 35-40 kilowatts into the 150 to 650 kilowatt band, with roadmaps uttering megawatt racks. As IEEE Spectrum reports, the industry is therefore moving to 800-volt direct-current distribution: 85% more power through the same conductor, 45% less copper needed, 5% efficiency gained. With a one-megawatt rack demanding some 200 kilograms of copper busbar, wiring architecture alone becomes an engineering problem at gigawatt scale.

In the closing stretch Badri explains what a sell-side past brings into the company: representing Cisco outside while acting as a shareholder advocate inside, asking both sides of the Street what they want to hear and recalibrating disclosure every quarter. His closing message pairs the company's claim of record revenue, record operating margin and record earnings per employee over the past twelve months with a one-line sector conviction: as long as token consumption charts keep exploding, this cycle counts as early. Just as offices once resisted learning Excel, each newly converted profession learning skills files and agents will compound demand another layer.