Cisco's $9.3 Billion AI Order Year Included Nearly 4,000 Job Cuts
At 1:30 p.m. Pacific on August 12, Cisco executives began explaining the strongest annual results the networking company had reported in decades. Fiscal 2026 revenue had reached $63.3 billion. Fourth-quarter revenue was $17.3 billion. Hyperscaler customers had placed $9.3 billion in AI infrastructure orders during the year, more than four times the prior-year amount.
Chief Financial Officer Mark Patterson added a workforce measure. Cisco had achieved its highest revenue, operating margin, and earnings per employee in 30 years, according to the company’s fourth-quarter release.
Three months earlier, that denominator had carried a different announcement. On May 13, Chief Executive Chuck Robbins told employees that Cisco would eliminate fewer than 4,000 jobs, or less than 5% of its workforce, before the fiscal year closed. The company filed an estimated restructuring charge of up to $1 billion. By late June, California notices identified 471 affected employees in software engineering, product management, design, business operations, and other functions.
Cisco did not say AI performed the work of those employees. Its filing described a transfer of resources toward silicon, optics, security, and artificial intelligence. The company also said it expected to reinvest substantially all the savings, rather than keep a material reduction in total costs.
A purchase order, an AI prompt, and a severance notice produce different evidence. AI demand can expand revenue while investment moves between product lines and people leave. Internal tools can raise output without causing a specific layoff. A smaller employee denominator can lift an earnings-per-employee ratio before anyone proves that a model produced the gain.
Cisco’s year narrows the debate to a reporting problem. A company needs one account for customer demand, product investment, restructuring, automation, training, and employee exits when they happen at the same time.
A record close with a smaller roster
Robbins’s email reached employees on the same afternoon Cisco announced record third-quarter revenue.
On May 13, Cisco reported third-quarter revenue of $15.8 billion, up 12% from a year earlier. In the employee memo published the same day, Robbins said the company needed the focus and discipline to shift investment toward areas with stronger demand and long-term value. Notifications for affected employees would begin on May 14 and continue across countries under local rules.
The planned reduction was fewer than 4,000 jobs. Cisco’s last annual employee count before the announcement was about 86,200 as of July 26, 2025. That baseline came from its fiscal 2025 annual report, which assigned 29.3% of employees to research and development, 29.7% to sales and marketing, 30.1% to cost of sales, and 10.9% to general and administrative work.
Those percentages do not identify who left. The baseline preceded a year of hiring, acquisitions, attrition, and the new reduction. Cisco has not published a final global exit count, the affected functional mix, the number of contractors, or the number of new roles created in the investment areas. Dividing 4,000 by 86,200 would turn an announced ceiling and an old denominator into a false realized rate.
Cisco had an internal process that could have tracked some of the movement. The same annual report described “one company, many careers,” a skills-intelligence system for matching people to opportunities, and about 2 million manager check-ins submitted by roughly 65,700 employees during fiscal 2025. The current public record does not say how many people in the fiscal 2026 reduction entered that matching system or received an internal offer.
Local notices provide one narrow view. The Los Angeles Times reported that 471 workers across San Jose, Milpitas, and San Francisco were included in California filings. The listed roles crossed software, product, design, and operations. That evidence rules out a story in which only one obsolete back-office function disappeared. It still covers less than one-eighth of the announced global ceiling and cannot establish the worldwide pattern.
Notification, finance, and product time did not match. Employees experienced notice and termination dates. Finance recorded severance liabilities and restructuring charges. Product teams worked against customer orders that could turn into revenue months later. An earnings release compresses all of them into one quarter, but an affected engineer and a hyperscaler purchase order do not move through the company at the same speed.
An Associated Press review of Cisco and other technology companies found that executives increasingly mentioned AI beside job reductions while still citing restructuring, changing demand, and wider cost decisions. Public language rarely made AI the sole reason.
Cisco’s own documents fit that pattern. The company named AI as one investment destination. It did not report how many removed tasks were automated, which jobs had shrinking customer demand, which work moved to another team or supplier, or how many affected employees qualified for internal openings. Any claim that AI replaced nearly 4,000 Cisco workers would run past the evidence.
Calling the plan a reallocation cannot prove that AI played no role either. Cisco was investing in employees’ use of AI across the company while reporting record per-employee economics. Task and role evidence would have to connect those facts.
Orders and revenue run on different clocks
Cisco recorded $9.3 billion in hyperscaler AI infrastructure orders. It recognized about $4 billion in revenue from that business during the year.
The company took $4 billion of those orders in the fourth quarter alone. Its prepared investor remarks said the fiscal-year total was about 4.5 times the fiscal 2025 result. Roughly 60% involved Silicon One systems and 40% involved optics. Cisco also reported more than $1 billion in annual AI infrastructure orders from neocloud, sovereign, and enterprise customers outside the hyperscaler measure.
The products were concrete. In February, Cisco introduced its Silicon One G300, a 102.4-terabit-per-second chip for large AI clusters, along with new systems and optics. Cisco reported a 28% improvement in job-completion time for the G300 design and a 50% reduction in optical-module power use for its new linear pluggable optics compared with retimed modules. Those are company test claims rather than customer results, but they explain why silicon and optics appeared in both the order mix and the workforce memo.
Revenue followed a different schedule. Cisco recognized about $4 billion in hyperscaler AI infrastructure revenue during fiscal 2026 and forecast $7.5 billion for fiscal 2027. A large customer can place an order before hardware ships. Delivery can span quarters. Supply availability, customer construction, acceptance, and contract terms can change when revenue appears.
Put the commercial figures side by side:
| Measure | Cisco’s disclosed figure | What it establishes | What it does not establish |
|---|---|---|---|
| FY26 hyperscaler AI infrastructure orders | $9.3 billion | Contracted demand recorded during the year | Revenue, cash, gross profit, or customer use |
| FY26 hyperscaler AI infrastructure revenue | About $4 billion | Product revenue recognized under Cisco’s accounting | Total order conversion or workforce return |
| FY27 expected AI infrastructure revenue | $7.5 billion | Management’s forward revenue expectation | Completed sales or realized margin |
| FY26 total company revenue | $63.3 billion | Cisco’s company-wide recognized revenue | Revenue generated by internal employee AI use |
| Q4 networking product order growth | 40% year over year | Broad demand beyond one product | Permanent demand or jobs attached to it |
Large infrastructure orders are often placed before delivery, so the gap between orders and revenue is ordinary. A strong order book can justify hiring and investment across supply chain planning, silicon engineering, optics, security, sales, customer deployment, and support.
Yet the mix matters for labor. Silicon and optical systems do not require the same organization as older networking products, acquired software, collaboration tools, support operations, or regional sales structures. A company can need more engineers in one architecture and fewer people maintaining another. It can also buy a capability through acquisition, shift work to suppliers, or use software to absorb coordination work.
Cisco closed two acquisitions in the fourth quarter: Galileo Technologies in observability and Astrix Securities in non-human identity security. Those deals can add employees and revenue while a restructuring removes roles elsewhere. A year-end employee count without an acquisition bridge would hide both movements.
Acquired talent also had a retention price. In its filing for the first half of fiscal 2026, Cisco estimated that acquisition agreements could require up to $400 million in future cash compensation tied to continued employment of certain acquired-company workers. The amount covers acquisition agreements rather than the later Galileo and Astrix deals alone. It still shows why a plain net employee count is inadequate: Cisco could pay to retain one group of specialists while funding exits elsewhere.
Customer concentration adds another limit. Hyperscaler orders can be large and irregular. Four major hyperscalers each grew AI infrastructure orders at triple-digit rates in the fourth quarter, Cisco said. That is strong demand, but it also means a few customers can move the annual total. The staffing plan needs to distinguish durable product capability from labor added for a temporary order peak.
Cisco clearly won substantial AI infrastructure demand. But an order tells the company what customers want to buy, not an employee why a particular role ended.
One billion dollars buys a reorganization
The SEC filing gave Cisco’s reorganization a cash price before the company could report a final workforce result.
In its May 13 Form 8-K, Cisco estimated up to $1 billion in pre-tax charges. The amount consisted mainly of cash-based severance, one-time termination benefits, and other costs. The company initially expected about $450 million in the fourth quarter and the remainder in fiscal 2027.
The August results showed $511 million in restructuring and other charges for the quarter and $693 million for the full year. Those figures are broader accounting lines, so they should not be treated as the exact severance cost of the new plan. They do show that reorganization was material enough to sit beside the record revenue numbers in the financial statements.
Cisco’s quarterly filing for the period ended April 25 said the company expected to reinvest substantially all the cost savings from the plan in its growth opportunities. Overall cost savings were not expected to be material.
The job cuts were therefore a transfer of capacity rather than a simple expense reduction. Cisco was paying cash to remove positions, then moving recurring resources toward other products and work. The announced destinations were silicon, optics, security, AI, and internal employee use of AI.
The filing did not allocate dollars among them. It did not say how much would fund new hires, compensation, chips, product development, cloud use, acquisitions, training, or deployment. It also did not disclose how much annual payroll the removed roles represented. A reader cannot reconcile the $1 billion charge to a future AI investment line.
Cisco had recently completed another plan. Its fiscal 2025 restructuring accumulated $926 million in charges and was substantially complete in the second quarter of fiscal 2026. The new plan arrived before the year ended. Repeated restructuring is important context because it weakens any clean story that a single new AI capability suddenly made one group unnecessary.
An organization change also preceded the new plan. On April 27, after Common Hardware Group leader Martin Lund decided to return to an early-stage AI company, Robbins moved the hardware group into the wider product organization under Jeetu Patel. Cisco said the move would align silicon, systems, optics, and applications. Sixteen days later, the company announced the workforce reduction and named many of the same product areas as investment destinations.
Large technology companies have always reorganized around product cycles, acquisitions, margin targets, geographies, and customer demand. Cisco has lived through transitions from routing to broader networking, software subscriptions, security, observability, and cloud-managed products. AI entered a company already practiced at moving costs between portfolios.
Recurring portfolio changes provide the strongest counterargument to an automation-first reading. Customer demand shifted sharply toward high-speed networking, silicon, and optics. Cisco had a business reason to change its workforce even if no internal AI tool eliminated a single job. Holding every role constant while the product mix changed could delay delivery and weaken competitiveness.
That counterargument shifts the reporting burden. Cisco should identify roles removed because demand fell, roles redesigned around AI, work transferred to vendors, positions left open after attrition, new roles funded, and affected employees moved internally. Without that bridge, “strategic reallocation” remains a destination label rather than evidence of a completed transfer.
Customers have a stake in the bridge. A hyperscaler buying a new switching system needs engineering support, supply coordination, deployment, and fault resolution after the purchase order. Cisco’s own risk disclosure says restructuring may fail to make the business more efficient or effective and may harm operating results. A transfer that lifts earnings per employee but delays a customer deployment would fail a commercial test even if the accounting ratio improved.
Shareholders received another transfer. Cisco returned $12.7 billion through repurchases and dividends during fiscal 2026, equal to 99% of free cash flow according to its prepared remarks. That capital return is neither restructuring spend nor AI investment. It belongs in the same allocation picture because management chose all three: cash for shareholders, cash for restructuring, and reinvestment in growth areas.
Cisco could afford each choice. Employees were asked to understand why their roles funded the next portfolio while investors received almost all free cash flow. A complete account would show the scale, timing, and expected result of each destination.
Productivity per employee leaves out the transfer
Patterson defined Cisco’s 30-year productivity high through revenue, operating margin, and earnings per employee. Each measure is real. None isolates AI.
Revenue per employee has a numerator and a denominator. Revenue can rise because Cisco sells more units, raises prices, improves product mix, completes an acquisition, recognizes prior orders, or benefits from a customer spending cycle. The employee count can fall through layoffs, attrition, divestitures, or slower hiring. Work can also move to contractors and suppliers who disappear from the employee denominator while their invoices remain elsewhere in the accounts.
Earnings per employee adds more variables. Gross margin, operating expenses, stock compensation, restructuring charges, amortization, taxes, and capital structure can all move earnings. A reduction in employees can lift the ratio even if output per remaining worker stays unchanged. Internal AI could contribute, but the public ratio cannot calculate its share.
Cisco had not yet published a fiscal 2026 employee denominator in the August earnings release. The last annual baseline was 86,200. The company did not disclose total hours worked, overtime, contractor headcount, accepted product output, software defects, customer response times, or work moved across borders. It also acquired companies during the year. A record ratio without that bridge is a corporate result, not a labor diagnosis.
Consider two simplified cases. In the first, revenue rises 12% and employee count stays flat. Revenue per employee rises 12%. In the second, revenue rises 12% while employee count falls 4%. The ratio rises about 16.7%. The extra increase comes from the smaller denominator, regardless of what caused the headcount change.
Nothing in that step requires AI.
The arithmetic does not make the metric dishonest. It exposes an allocation choice inside the label “productivity.” A company reports more revenue or earnings for each person it still counts, while the people who left no longer appear in the denominator.
| Workforce question | Needed denominator | Current public status |
|---|---|---|
| Did internal AI reduce task time? | Same workflow, baseline minutes, review and correction time | Not disclosed |
| Did AI raise accepted output? | Comparable output volume and quality before and after use | Not disclosed |
| Did fewer employees produce the revenue? | Current average employees, contractors, acquisitions, hours | Incomplete |
| Did portfolio demand shift? | Revenue, orders, open roles, exits, and hires by product line | Partly disclosed |
| Did employees benefit? | Pay, hours, promotion, training, placement, and retention | Not disclosed as a joined record |
| Did customers benefit? | Deployment time, reliability, security, support, and renewal outcomes | Not tied to workforce change |
Microsoft ended fiscal 2026 with 30 million paid Copilot seats and a smaller workforce. Paid seats, capital spending, and headcount could all be verified, but the bridge from product adoption to labor outcome remained missing. The site’s earlier analysis, Microsoft Closed FY26 With 30 Million Paid Copilot Seats and a Smaller Workforce, treated those measures as related management facts rather than a causal chain.
Productivity Rose. Labor’s Share Reached a Record Low. examined a national productivity increase and a 52.9% labor share without assigning either movement to AI. Company accounts face the same attribution problem. A ratio can improve while the distribution of gains remains unknown.
Cisco managers need the next planning cycle to explain the work behind the record metric. If a team shipped more because demand grew, record the demand. If engineers used Circuit or coding tools, record accepted work and review time. Count work moved to a supplier. Keep positions that disappeared before the year began in the bridge.
Inside Cisco’s people-first promise
Seven weeks after the job-cut announcement, Cisco published an internal-workforce account under the heading “Building a people-first AI strategy”. It described role-specific learning, live labs, time for experimentation, human oversight, and an internal assistant called Circuit.
Mary de Wysocki, Cisco’s senior vice president for learning and future readiness, argued that companies should invest in people with the same urgency as technology. Fran Katsoudas, the executive responsible for people, policy, and purpose, described AI as team work.
Chief Learning Officer Marci Paino focused on trust, transparency, data use, and expectations for employees learning the tools.
Jeetu Patel, Cisco’s president and chief product officer, described agents as a new set of coworkers that could expand the work an organization attempted. Austin Roth-Eagle, who oversaw an AI acceleration office in communications, emphasized domain expertise and human oversight. Gianpaolo Barozzi, the technology leader in the people organization, used a mixed team of artists, engineers, and agents for an installation at the Venice Biennale.
Circuit routes employee requests among Cisco models and external systems such as Claude, ChatGPT, and Gemini. Cisco executives described it as useful for brainstorming, work quality, security, and process improvement. The company did not publish active users, role coverage, messages, accepted outputs, time saved, error rates, or business results.
No published denominator means the program cannot yet be reconciled with the finance narrative. The July article said AI should support people. The August release said earnings per employee had reached a 30-year high. The May memo said nearly 4,000 jobs would leave while investment shifted toward internal AI use.
A joined record would show who experienced each part.
Robbins offered affected employees two forms of support. Cisco’s placement service had seen 75% of participants find a next role, and departing employees would receive one year of access to Cisco U courses and certifications in AI, security, networking, and other subjects.
Both offers need denominators. Cisco did not disclose how many people used the placement service, the measurement window, whether the roles were internal or external, how comparable the pay and level were, or how many participants were from the current reduction. “Found a next role” could cover a quick internal transfer or an external job months later. Those outcomes matter differently to an employee and to the company.
The California role list makes that distinction practical. Software engineers, product managers, designers, and business-operations workers can have skills relevant to the portfolio Cisco said it was funding. Their titles alone do not prove they qualified for the new work, and growth teams may need different hardware, systems, or customer knowledge. An internal-mobility record would show how often the company tested that match before an exit.
One year of course access stops well short of a worker outcome. Access does not show enrollment. Enrollment does not show completion. A certificate does not show placement, pay recovery, or use of the skill at work. Training can help a network engineer move into AI infrastructure or security, but a severance clock and a learning clock may not match.
Remaining teams absorb customer demand, handoffs, vacant work, training, and new tool expectations. If Cisco reinvested substantially all savings, some teams should gain hiring capacity or new colleagues. Others may carry more work while the new investment takes time to arrive. A per-employee record should include workload, hours, defects, and retention, not only earnings.
Their managers face a sequence that never appears in an earnings-per-employee ratio. They receive the notification plan, decide which work stops, arrange access and customer handoffs, interview internal candidates, keep a delivery calendar moving, and introduce new AI tools. Cisco already had a weekly check-in system. Adding work disposition and transfer status to that existing rhythm would cost less than reconstructing the story after people leave.
Cisco’s people-first language sets a test. Customer orders may require different skills, and a responsible reallocation can support employees through internal transfers, portable credentials, severance, and external placement. Growth-area hiring does not require pretending every former job has an exact successor.
Completed exits, internal moves, external placements, training completion, comparable pay, time to placement, new hires, and the work attached to them would strengthen that case. Internal AI results need a workflow view: shorter queues, better quality, added product capacity, and any work transferred to remaining employees.
Trust cannot rest on a collection of executive statements. It accumulates when an employee can trace the decision from a removed role to a funded destination and then see whether the promised outcome arrived.
A workforce reallocation account
At a quarterly review, Cisco’s numbers would arrive from different owners. Sales brings orders; finance brings revenue, expense, and restructuring. HR holds employee counts and placement, while learning teams record course activity. Product groups own output and quality. Procurement brings supplier cost. Investor relations selects the measures that reach shareholders.
The separation makes attribution easy to suggest and hard to test. A workforce reallocation account joins the records without forcing them into one causal score.
| Account field | Minimum record | Decision it supports | Common reporting error |
|---|---|---|---|
| Workforce baseline | Employees, contractors, open roles, hours, function, geography | Establish the real denominator before change | Using last year’s employee count as the current workforce |
| Exit event | Announced ceiling, notified people, completed exits, voluntary departures, date | Separate plan from realized reduction | Reporting “up to” as completed layoffs |
| Work disposition | Stopped, automated, transferred, outsourced, combined, or still open | Explain what happened to the work | Assuming a removed job means removed work |
| Investment destination | Dollars, positions, product, owner, start date | Test whether savings moved where management said | Naming AI without an allocated amount |
| Hiring and mobility | New hires, internal moves, level, location, source team | Show whether new capabilities came from people already employed | Counting a job posting as a hire |
| AI use | Eligible users, active users, workflow, model, review requirement | Measure actual adoption | Treating tool access as use |
| Accepted output | Units, quality, errors, rework, customer acceptance | Test whether AI changed productive work | Counting messages or generated tokens as value |
| Learning | Offered, enrolled, completed, practiced, certified | Separate access from skill development | Counting course seats as reskilling |
| Placement | Internal or external destination, time, level, pay recovery | Evaluate support for affected employees | Publishing a percentage without cohort or window |
| Commercial result | Orders, revenue, gross margin, renewal, delivery date | Connect demand to delivered business | Calling orders revenue or profit |
| Capital allocation | Restructuring cash, growth investment, acquisition, buyback, dividend | Show competing uses of cash | Presenting all spending as AI investment |
| Revisit | 6- and 12-month result, variance, owner | Test the original plan after execution | Closing the record on announcement day |
Cisco already disclosed pieces of almost every row. The missing work is not a new data-collection system. It is a reconciliation.
A Cisco review would begin with average employees during the period rather than one year-end point. Contractors and acquired employees belong beside them. Open requisitions matter because a company can reduce planned hiring without laying off an incumbent. Broad function and product-group splits can show where capability moved without exposing an individual.
The exit row would retain the wording “fewer than 4,000” as a ceiling. A complete close would show notifications, completed exits, rescinded notices, voluntary departures, and internal transfers. California’s 471 notices belong in the geography column, not as a proxy for the world.
Work disposition is the missing center. A role can end while its tasks continue. Product work may move to a different group. Support can move to a partner. A manager can absorb coordination. Software can automate a step. A product can be retired. Each route creates a different cost and employee effect.
Cisco named silicon, optics, security, AI, and employee use of AI as destinations. The account should show funded positions and spending for each. A new silicon engineer, an internal transfer, a model contract, a security acquisition, and a learning program do not produce the same labor result.
AI use and accepted output should stay separate. An active Circuit user is not a completed customer outcome. A generated code change needs review, testing, deployment, and maintenance. A marketing draft needs approval and performance measurement. A support answer needs resolution and customer feedback. Review time belongs in the cost.
Placement requires a worker outcome. A useful rate could report the share of participating affected employees who accepted an internal role within 90 days, the share who reported an external role within 180 days, and the share whose new compensation stayed within a defined range. Participation, privacy, and missing responses should remain visible.
The account should not promise a neat causal result. Cisco’s revenue can rise for reasons unrelated to internal AI. Orders can expand before new hires arrive. Restructuring can remove work that was already shrinking. The aim is narrower: prevent incompatible measures from being presented as one proof story.
Boards can use the account to ask concrete questions. Did the planned investment receive the savings? Which teams gained capacity? Did contractor spending replace employee cost? Were customer delivery times better? Did remaining employees work longer? How many affected workers moved into the growth areas? Did the $7.5 billion revenue expectation arrive with the workforce management expected?
Employees can use the same record differently. It tells them whether “reallocation” meant a real internal path, a training offer after exit, or a transfer of work without people. Managers gain a staffing plan. Finance gains a cleaner return model. Investors can separate demand from cost removal.
An earlier digidai analysis, AI Spend Turns Hiring Into an Adoption Test, argued that spending does not automatically reduce payroll. Cisco presents the reverse accounting challenge. Removing roles does not automatically prove AI adoption. Both claims need workflow evidence between the budget and the workforce result.
Next year’s missing denominator
Cisco expects about $7.5 billion in hyperscaler AI infrastructure revenue in fiscal 2027. Robbins and Patterson also told investors to expect the per-employee measures to improve again.
Those expectations create a clean verification window. The next annual report can provide a new employee count. Later filings can close the restructuring charge, identify whether the plan finished by the end of fiscal 2027, and report whether the expected revenue arrived. Cisco can also publish a workforce bridge before investors have to infer it.
Five disclosures would change the quality of the story: average employees and contractors; final exits and internal transfers by broad function; new hires in named investment areas; placement and training outcomes for affected workers; and workflow-level results from internal AI use. None requires Cisco to reveal an individual’s performance or a customer’s confidential order.
Companies change portfolios, and some work ends. A traceable transfer would name the work Cisco stopped, the capability it funded, the employees it moved or supported, and the customer result that followed.
For now, two Cisco records sit nine weeks apart. In May, an employee could read that fewer than 4,000 jobs would leave and that one year of Cisco U access would remain available. In August, an investor could read that AI orders reached $9.3 billion and earnings per employee reached a 30-year high.
Fiscal 2027 will show whether those two pages ever meet.
This article examines Cisco’s fiscal 2026 AI infrastructure orders, workforce restructuring, per-employee metrics, and employee transition commitments. Published August 21, 2026.