Cisco’s $9.3B AI Orders Put Networking Back in Charge
Cisco booked $9.3 billion in AI-infrastructure orders while everyone watched the chips. The AI bottleneck is data movement — and networking is getting paid again.
Cisco just booked $9.3 billion in AI-infrastructure orders from hyperscalers in fiscal 2026. If you still think AI is mainly a software story, you are missing the bill that makes the whole thing work: moving data between the chips.
The fashionable view is that Nvidia makes the picks and shovels, OpenAI makes the magic, and everyone else gets crumbs. That is lazy thinking. AI systems do not work because you bought a pile of clever chips. They work when vast numbers of those chips can exchange data quickly, reliably and without melting your economics.
Cisco’s numbers are the bluntest proof yet that networking has become a toll road again.
The $9.3 billion signal everyone should notice
On August 12, Cisco reported a record fourth quarter: $17.3 billion in revenue, up 18% year on year, and $3.9 billion in GAAP net income. Fine results. But the number that matters is elsewhere.
Cisco said hyperscalers placed $4 billion of AI-infrastructure orders in the quarter, taking the full-year tally to $9.3 billion. That is roughly 4.5 times its fiscal-2025 total. Cisco also said it delivered about $4 billion of AI-infrastructure revenue in fiscal 2026 and expects $7.5 billion in fiscal 2027.
Read that again: orders are $9.3 billion, recognised revenue is about $4 billion, and the company is guiding to $7.5 billion next year. There is a lot of work already sold that has not yet turned into revenue.
That is not a vague CEO promise about “AI opportunity”. It is purchase orders from customers with enormous capex budgets.
Cisco’s fourth-quarter product orders rose 35%, or 25% even excluding hyperscalers. Networking product orders rose 40%, its eighth straight quarter of double-digit growth. The company is forecasting fiscal-2027 revenue of $72.2 billion to $73.4 billion, after reporting $63.3 billion for fiscal 2026.
The networking bloke your startup friends stopped talking about has wandered back into the room carrying a very large invoice.
AI has a data-movement problem, not just a chip problem
Here is the bit most AI commentary gets wrong: the model is only one layer of the machine.
A serious AI deployment needs compute, memory, power, cooling, storage, security and networking. The moment you spread a workload across thousands of accelerators, moving information among them becomes central to performance. A bottleneck in the network can leave staggeringly expensive compute sitting around waiting.
That is why this matters. The bill for AI does not stop at the GPU rack. The more chips the hyperscalers add, the more they need switches, silicon, optical systems and architecture that can move data at absurd speed.
Cisco is explicitly pitching into that reality through networking, Silicon One chips and optics. In the fourth quarter, it reported three new hyperscale AI design wins across a Silicon One P200 scale-across system, a G200 scale-out system and an optical-line system. Those product names are less important than the pattern: Cisco is trying to sell the connective tissue, not merely a commodity box at the edge of the data centre.
And that is a far better place to be than selling generic IT gear while customers cut budgets.
There is an old business lesson here. When a market moves from experimentation to industrial scale, the sexy front end attracts attention. The boring constraints collect the revenue.
Cisco is not “back” because it found a chatbot
Let’s not get carried away and pretend Cisco has suddenly become a hot startup in a hoodie. It has not. It is still a massive incumbent with all the baggage that implies: big customer procurement cycles, legacy products, fierce competition and the usual danger of management calling every decent quarter a supercycle.
But that is exactly why the story is more interesting.
Cisco does not need to become the next Nvidia to win. It needs AI demand to make its existing strengths more valuable: enterprise relationships, network operating expertise, systems integration, security and the ability to sell and support enormous infrastructure deployments.
Its installed base matters because moving a company’s network is painful. It is not like swapping a project-management app because your team is bored on a Tuesday. Once AI workloads, security controls, data-centre links and campus networks are tangled together, customers place a premium on reliability and accountability.
That gives incumbents an advantage startups hate admitting exists: when the buyer is spending hundreds of millions, “nobody got fired for buying Cisco” can become a commercial weapon again.
Cisco also said data-centre-networking orders grew 25% in fiscal 2026, while orders from neocloud, sovereign-cloud and enterprise-AI customers added $1.3 billion beyond its hyperscaler AI-order figure. That does not prove every customer will keep spending at this pace. It does show the demand is wider than one flashy deal or one customer having a spending fit.
The overlooked angle: the AI boom may favour operators over inventors
The contrarian take is this: the next tranche of AI wealth may not go mainly to the people building the most impressive model. It may go to the operators who make expensive systems run cheaply, securely and at scale.
Everyone loves a demo. Nobody loves network architecture until it fails.
That is good news for disciplined operators. AI rewards people who can map a bottleneck, measure the economics and execute without romance. It punishes businesses that buy technology because a competitor posted an enthusiastic LinkedIn video.
Cisco’s figures also contain a warning for investors. Orders are not revenue, and revenue is not free cash flow. A $9.3 billion order book is valuable only if products ship, margins hold and customers continue building. Cisco itself flagged risks around customer demand, component costs, supply constraints, competition and the timing of orders.
The hard question is whether this is a durable re-rating of networking economics or a capex surge that gets ahead of real AI demand. You do not answer that by chanting “AI”. You watch conversion: backlog into revenue, revenue into margin, margin into cash.
Cisco expects about $7.5 billion in fiscal-2027 AI-infrastructure revenue. That is now the scoreboard. If it delivers, the company has earned a stronger claim that this is structural. If it does not, then a lot of people have confused a burst of orders with a permanent transformation.
Why founders should care even if they will never buy a Cisco switch
Most founders will read this and think, “Great, big tech bought some networking gear. What has that got to do with me?”
Plenty.
First, it tells you the AI arms race is still capital intensive. The winners at infrastructure level are spending real money on physical capacity, not just talking about agents and prompts. If your product depends on cheap, instant, infinitely available AI inference, build your model with the possibility that underlying costs remain stubbornly real.
Second, it is a reminder to search for bottlenecks, not buzzwords. In every market, the obvious innovation grabs headlines while the scarce complement quietly gains pricing power. For AI, that could be networking, power, cooling, data rights, security, deployment expertise or distribution.
Third, do not confuse an impressive technology with a complete business. Cisco’s edge is not that it invented intelligence. Its edge is that it can package infrastructure into something huge customers can purchase, install and rely on. That is boring right up until it becomes a $9.3 billion order book.
What this means for you
If you run a business, do three things this week.
1. Find the expensive waiting time in your operation. In AI data centres, chips waiting on data are a disaster. In your company, it may be sales waiting on legal, customer support waiting on product fixes, or a founder approving every minor decision. Measure it. Waiting time is where margin goes to die.
2. Build around the constraint, not the demo. Before buying another AI tool, ask what stops your team producing more valuable work today. If the answer is bad data, slow approvals or no distribution, another chatbot will not save you.
3. Follow revenue conversion, not announcement volume. Whether you are investing in public markets or backing a startup, separate orders, contracts, revenue, gross profit and cash. They are not interchangeable. Anybody can announce ambition. Cisco’s next test is converting its AI order book into the $7.5 billion of fiscal-2027 revenue it has forecast.
The big lesson is simple: when everyone is chasing the gold, do not only watch the miners. Watch who owns the road between the mine and the refinery.
Right now, Cisco is making a serious case that it does.