Cisco is using a silicon-first strategy to lock in AI hardware supply, reshaping sourcing, inventory and capacity decisions across its global network.
In Brief
- Cisco is moving upstream into silicon and optics to gain direct control over critical AI hardware supply and reduce dependence on merchant components.
- The company is using its balance sheet to secure long-term semiconductor and memory capacity, accepting higher working capital to avoid supply failures.
- Margin pressure from hardware-heavy AI demand is being managed through product redesign, pricing discipline and tighter commercial terms.
The Strategic Break: Treating Silicon as a Primary Supply Node
Cisco describes roughly half of its current AI hyperscale business as systems based on its in-house Silicon One portfolio. That is more than a product choice. It is a structural shift in how upstream supply is managed.
AI infrastructure orders from hyperscalers reached 1.9 billion dollars in the third quarter, up from 600 million dollars a year earlier, with year-to-date orders of 5.3 billion dollars already above the prior full-year expectation. Cisco now expects around 9 billion dollars of AI infrastructure orders from hyperscalers in fiscal 2026 and at least 6 billion dollars of AI hyperscale revenue in fiscal 2027. Those numbers define a new, multi-year volume base that must be backed by predictable silicon and optics supply.
Instead of relying mainly on merchant chips, Cisco is reshaping its operating model around control points it owns or tightly steers: Silicon One for routing and switching, Acacia for coherent optics, and a dedicated memory strategy. A restructuring programme is explicitly reallocating resources toward silicon, optics, security and AI. This is less a cost-cut and more a decision to anchor growth on components that sit at the centre of current supply risk.
In parallel, Cisco’s Acacia optics business booked over 1 billion dollars of orders in the quarter and is on track to grow more than 200 percent year over year, with more than 750,000 400G and over 40,000 800G coherent pluggable optics shipped to date. Five new hyperscaler design wins in the quarter, including systems based on Silicon One P200 and G200, lock in further future demand.
In supply chain terms, this is a pivot from managing a broad commodity electronics stack to operating a few strategic component franchises that set the tempo for the rest of the network.
How The Silicon-first Model Is Being Executed
At operating level, the shift shows up first in sourcing decisions and balance sheet use.
Cisco reports that silicon supply is secured through the end of calendar 2026, with negotiations under way for 2027. Across silicon, substrates, memory, photonics, printed circuit boards and power, the company is signing long-term agreements where possible, engaging directly with sub-tier suppliers and building what it calls strategic inventory.
Inventory and advanced purchase commitments increased by 6.7 billion dollars in the last 90 days alone, a 48 percent jump, and by 11.6 billion dollars year over year. Operating cash flow of 3.8 billion dollars for the quarter was down 7 percent, explicitly linked to investment to meet growing AI infrastructure demand. Cisco is leaning on its financial strength to carry more stock and commit to future capacity in exchange for supply certainty and the absence of decommits.
Memory is treated as its own supply problem. Non-GAAP product gross margin fell 330 basis points year over year to 64.3 percent, with management tying the decline directly to mix and higher memory costs. The response combines three execution levers:
- design changes to cut memory usage, with over 20 active programmes to reduce memory per product and new wireless products in the fourth quarter that require 50 percent less memory;
- a technical migration from DDR4 to DDR5 across multiple programmes; and
- a three-year strategic investment and supply agreement with memory supplier Nanya.
On top of that, Cisco is tightening commercial terms to contain cost-price timing risk. Notice periods for price increases have been cut from roughly 30 days plus 30–45 days of quote protection to 15 days plus 15 days. The shorter window reduces the period in which orders can be booked at old prices while component costs are already rising.
Price increases are already visible in order patterns. For non-webscale customers, product orders grew 19 percent year over year in the quarter, up from 10 percent in the prior period, with 4–5 percentage points of the acceleration attributed purely to price rather than volume.
In practical terms, Cisco is resetting how upstream and downstream commitments are sequenced: components are locked in earlier and for longer; commercial exposure to cost swings is shortened; and design teams are tasked with structurally reducing sensitive materials like memory.
Aligning Capacity With Volatile AI Demand
AI-related demand is both large and uneven. Orders from service provider and cloud customers grew 105 percent year over year in the quarter, with five top hyperscalers each posting triple-digit growth. AI infrastructure orders from non-hyperscaler customers, including neocloud, sovereign and enterprise buyers, added around 300 million dollars in the quarter and 900 million dollars year to date, with a 3 billion dollar pipeline reported.
Against that volatility, Cisco is using design wins and product standardisation to make capacity more fungible. The Silicon One P200 and G200-based systems mentioned earlier are positioned as scalable anchors for different AI use cases. Management notes that the first large-scale P200 deployments, labelled ‘scale across’, will not contribute at scale until fiscal 2027, providing lead time to align fabrication slots, test capacity and logistics.
Networking product orders grew more than 50 percent in the quarter, led by triple-digit growth in service provider routing and compute and double-digit growth in data center switching, campus switching, wireless, enterprise routing and industrial products. AI-tagged Nexus data center switch orders were up almost 50 percent sequentially. Campus networking orders grew more than 25 percent year over year, with record wireless orders up more than 40 percent and WiFi 7 representing about half of the wireless mix.
To keep this mix flowing, Cisco reports that lead times for traditional networking have returned to non-extreme levels. That suggests a planning cadence that separates relatively stable campus and enterprise refresh cycles from more volatile hyperscaler AI projects, while using common silicon and optics platforms to support both where possible.
Peers are converging on similar patterns from different starting points. Equinix is upgrading its Fabric interconnection platform from 10G to 400G and designing new data centers for flexible power densities, while Cloudflare is running all products across a single global network and optimising its hardware stack for AI inference. Cisco’s emphasis on in-house silicon and optics sits alongside these approaches as another way to stabilise capacity decisions under AI-driven traffic growth, but with a heavier focus on component control than on colocation or edge software.
Working-capital And Margin Tension as a Design Constraint
The silicon-first approach comes with clear trade-offs. Carrying more inventory and committing billions of dollars in advance purchase commitments tightens working-capital headroom. The 7 percent decline in operating cash flow, despite record revenue of 15.8 billion dollars and product revenue up 17 percent year over year to 12.1 billion dollars, reflects that tension.
At the same time, the revenue mix is shifting toward hardware. Hardware grew around 30 percent in the quarter, while software grew 1 percent. This mix depresses gross margin because hardware generally carries lower percentage margins than software, even when hardware margins are healthy. Total non-GAAP gross margin fell 260 basis points to 66 percent.
Cisco is signalling that it will defend operating margin more than gross margin. Operating expense as a percentage of revenue declined from 34.1 percent to 31.9 percent year over year, and non-GAAP operating margin held at 34.2 percent, in line with a stated target of around 34 percent. Management explicitly acknowledges that some of the new high-volume hardware businesses will have different margin structures and that discipline on operating costs is part of the response.
Tariff assumptions add another boundary. Guidance for the fourth quarter and full fiscal year assumes current tariffs and exemptions remain in place. Any change in trade policy would feed straight into the cost base for imported components and finished goods, testing both the sourcing model and pricing agility.
Taken together, these factors make clear that Cisco’s silicon-first posture is constrained not only by manufacturing physics but by cash, margin and policy friction. The operating model now has to balance high fixed commitments in silicon and memory against the ability to reprice and redesign fast enough to keep margins within target bands.
What The Model Now Enables and Constrains
Cisco has turned its AI infrastructure opportunity into a supply chain problem centred on a small set of critical components. Ownership of silicon and optics designs, backed by long-term agreements and strategic inventory, increases control over availability and delivery reliability at a time when AI demand is straining global semiconductor capacity.
The cost of that control is a heavier working-capital profile and sustained pressure on reported gross margins. The company is addressing that through product engineering to reduce sensitive inputs, faster pricing adjustments, and tighter operating cost management rather than through a narrow focus on unit cost cuts.
The result is an operating model that trades some balance-sheet flexibility and short-term margin purity for delivery assurance and long-term volume commitments in AI hardware supply. That trade now shapes how sourcing, inventory and capacity decisions are made across Cisco’s global network.
