Skip to content
Encyclopaedia of PerceptionEncyclopaediaCosts & leverage
DE
Fixed costs, variable costs, profit

Cost Structure Simulator

Revenue consists of three components: variable costs, fixed costs and profit. This interactive tool shows how a revenue decline hits profit, why fixed costs are the decisive risk lever, how price elasticity works, and what all of this looks like in real life at the Volkswagen Group.

Set up a company's cost structure, then drag the revenue change slider. Fixed costs stay put, variable costs breathe with volume, and profit absorbs everything that is left over.

Sales revenue in a normal year
Rent, salaries, depreciation, as % of base revenue
Materials, suppliers, freight, per unit sold
The scenario: demand collapses or picks up

Where does the revenue go?

Base vs. scenario: the | marker shows the respective revenue. If the costs reach beyond it, a loss arises.

Break-even chart

Where the sales revenue line crosses total costs, the profit zone begins.

Two companies with exactly the same profit in a normal year (€25m on €100m of revenue), but completely different cost structures. Drag the slider and see who survives the downturn.

Company A: fixed-cost heavy

e.g. a factory with deep in-house production, lots of staff & machinery

Fixed costs €45m variable ratio 30% Leverage (DOL) 2.8×

Company B: flexible

e.g. a retailer / contract manufacturer, buys in heavily from suppliers

Fixed costs €10m variable ratio 65% Leverage (DOL) 1.4×
The same market shock hits both companies

Profit as a function of revenue

The steeper line is the bigger lever, upwards as well as downwards. ● = break-even.

Raising prices sounds tempting, but demand reacts. The price elasticity ε tells you how strongly: volume change ≈ ε × price change. Example: ε = −1.5 and +10% price → −15% volume. Whether it pays off is decided by the cost structure.

Starting price: €50 per unit

Base case: price €50 · volume 1,000 units · variable unit costs €20 · fixed costs €15,000 → profit €15,000

Volume, revenue and profit under price changes

Index: base = 100. ◆ = revenue-maximising price, ● = profit-maximising price.

A revenue decline never stops at the manufacturer: it cuts orders, its suppliers lose revenue and in turn cut their own orders and investment, and so the shock propagates through the entire chain. Two mechanisms work together here: the bullwhip effect (each tier overreacts) and the multiplier (each round triggers the next). Exactly this dynamic can ignite a downward and price spiral in an industry.

the original demand shock
Inventory rundown, cancellations, caution: orders fall more sharply than sales
Share of the revenue loss passed on as spending cuts

The bullwhip effect: the further back, the harder the hit

Percentage decline per tier of the supply chain: the end-customer shock is amplified upstream.

The multiplier: one round becomes many

Over the rounds, €100 of direct demand loss turns into a multiple: each player's cut is the next player's revenue loss.

The downward spiral

When everyone cuts at the same time, each step feeds the next: the classic deflationary spiral pattern (Fisher 1933).

Real-world example, Austria: Magna Steyr production H1/2025 −18% Magna Graz ≈ −800 jobs (2024) AVL Graz −350 Schaeffler Berndorf −450 Hella Großpetersdorf: production ends by 2027 Counter-trend: Xpeng & GAC have been manufacturing in Graz since 2025

Theory meets reality: the Volkswagen Group, half year 2026 (statement of 24 July 2026), financial year 2025, and the restructuring decision of 4 September 2026. A group with an enormous fixed-cost base (plants, around 670,000 employees, development) on an operating margin of 3.8 per cent: the fixed-cost leverage from tab 2, at full scale.

US tariff burden ≈ €5bn per year (direct + indirect) China deliveries −31.6% Write-off for ending US ID.4 output €0.5bn Restructuring of 4 Sept: 50,000 further posts Savings target €11bn Investment €125bn

Where did the revenue go? VW Group income statement

Shares of revenue of €321.9bn (2025) and €324.7bn (2024). What remains of revenue at the end is the slim operating margin.

What if: a revenue decline at the VW Group

Model calculation based on the 2025 and 2024 income statements. VW does not report fixed and variable costs separately, so the variable share is an assumption (slider).

mainly materials & suppliers; the rest is treated as fixed

The decision of 4 September, put through the model

On 4 September 2026 the supervisory board approved the largest restructuring in the company's history: around 100,000 posts including the reductions agreed since late 2024, up to a quarter of management positions, €125bn of investment and €11bn of cost savings. Four sites received an end date for vehicle production rather than a closure: Emden and Zwickau in 2031, Hannover in 2032, Neckarsulm in 2034. This card asks the only question this chapter can answer: what does €11bn less fixed cost do to the break-even?

Same model as above, same variable share. The saving is treated as a pure reduction of the fixed block, which is the friendliest possible assumption: in reality part of any such programme falls on variable cost, and the effect arrives over years rather than at once. The dates are the point. A buffer that only exists from 2031 does not protect the 2027 balance sheet.

The AI boom is financed in a circle: investors put money into AI companies that are contractually obliged to buy from those same investors. Nvidia commits up to $100bn to OpenAI, OpenAI commits around $820bn of compute to Microsoft, Oracle, AWS, Nvidia, CoreWeave and Broadcom, Microsoft holds 27% of OpenAI and books OpenAI's Azure spending as AI revenue. The Dependencies chapter shows that circle. This tab asks the question that matters for a cost structure: what does a data centre do to a profit and loss account, and how much of the answer is an accounting choice?

Set the investment

one-off, written off over the useful life
balance sheets assume 5 to 6 years, critics argue 2 to 3
the same market shock in both variants

Profit as a function of revenue

The same company, the same market, two depreciation assumptions. The steeper the fixed-cost block, the earlier the loss. ● = break-even.

The order of magnitude

What the four largest platforms plan to invest in 2026, in billions of dollars. The bar on the right is the estimated annual depreciation difference between a six-year and a three-year useful life on that sum.

How to read this

Three honest caveats. First, the model is a stylised company with revenue of 100, fixed costs of 30 and a variable ratio of 50%, not a real balance sheet. Second, a real data centre earns revenue too; here only the cost side is shown, because that is what the lever from the previous tab acts on. Third, nobody knows the true useful life of an AI accelerator. Six years is the figure in the accounts, two to three years is what the critics assume from thermal stress and the pace of new hardware generations. The truth is somewhere in between, and it moves hundreds of billions.

Agentic flooding is the name a research group gave in August 2026 to a surge in the volume or complexity of requests that a service receives, caused by AI agents, that substantially strains the service. The paper is about public offices and courts. The mechanism is not. It is a cost story and it belongs next to the fixed-cost lever: producing a document has become almost free, reading, checking and deciding has not. The cost did not disappear. It moved from the sender to the receiver. This tab shows what that does to a working week, what it already looks like in practice, and what may follow in the near, the middle and the far future.

Set the inflow

What lands on one person's desk in a week, and what checking it properly costs. All four values are your estimate; the defaults describe a mid-sized agency or department in 2026.

concepts, offers, briefs, proposals, long emails
a memo written by hand had four pages; agent concepts have fifty
reading, verifying the claims, testing whether the timeline holds
the pages that would survive if the rest were deleted

Who pays for a page

Minutes per document as a function of its length. Stylised assumptions: writing by hand costs 20 minutes a page; with an agent it costs 10 minutes plus a quarter of a minute a page for a glance; reading and checking costs what you set above, plus five minutes to decide what to do with it. ● = the page from which the reader pays more than the writer.

The flood line

Hours of checking per week as a function of the number of documents received, for three document lengths. The dashed line is one full working week of 40 hours: above it, one person's entire week goes into checking before any real work begins. ● = your setting.

What can already be seen

Nine places where the flood is measurable today. The figures are those reported by the studies named in the footer; the setting differs each time, the arithmetic does not.

The inbox Symptom 1

Everyone is fast now. Writing an email, an idea or an assignment takes a minute; reading it takes as long as before. Microsoft measured 117 emails and 153 chat messages per person and day in its own usage data, an interruption every two minutes and 57 per cent of meetings without an invitation (June 2025, before most agents were in use). The sender's time per message collapsed. The receiver's did not.

The 50-page concept Symptom 2

Partner agencies deliver long, polished documents in which little can be executed. Stanford and BetterUp surveyed 1,150 US desk workers in September 2025: 40 per cent had received such "workslop" in the previous month, each case took just under two hours to sort out, about 186 dollars per employee and month, and roughly half of the recipients afterwards rated the sender as less capable and less reliable. The polish is real. The substance is what the reader has to find.

The timeline Symptom 3

"According to the AI everything is doable in a few days." The model's estimate is text, not experience, and it meets an old bias, the planning fallacy. METR had 16 experienced developers work on 246 real tasks in 2025 with and without AI. With AI they were 19 per cent slower, had expected to be 24 per cent faster and still believed afterwards that they had been 20 per cent faster. The gap between felt and measured speed is the raw material of every unrealistic timeline.

Courts and offices Symptom 4

The paper that coined the term collected 84 cases in 11 jurisdictions. German social courts reported a 55 per cent rise in caseload in 2025 attributed to AI-written claims, single submissions ran past 4,000 pages, and Australia reintroduced fees for freedom-of-information requests. In 87 per cent of the cases the mechanism was the same: a language model writing text cheaply for a service that has to process it expensively.

The consultation Symptom 5

27 per cent of the paper's cases were not services at all but channels of participation: freedom-of-information requests, public consultations, comments on plans. When a thousand identical objections cost nothing to send, the count stops meaning anything, and the office that has to answer each one is the one that pays.

Open source Symptom 6

The curl project, whose code sits in nearly every device, ended its bug bounty on 31 January 2026 because of a "torrent" of AI-written reports; seven arrived within one sixteen-hour stretch and none described a vulnerability. GitHub introduced caps on open pull requests per outside contributor in June 2026 and wrote the sentence that fits this whole tab: the cost to create has dropped, the cost to review has not. Merged pull requests rose from 25 million a month in January 2023 to 90 million in March 2026.

The reviewer Symptom 7

A pull request is a proposed code change that somebody must read before it goes in. An analysis of 470 open-source pull requests by CodeRabbit found about 1.7 times as many problems in those co-written with AI as in human ones; validating one takes half an hour or more; one developer with an agent produces five or six a day. The arithmetic gives every producer a reviewer who does nothing else for three hours a day (The New Stack, April 2026; the article is vendor-sponsored, so treat the figures as a practitioner's estimate).

Hiring Symptom 8

LinkedIn counted about 11,000 applications a minute in mid-2025, 45 per cent more than a year earlier; a British graduate opening drew 140 on average. Recruiters answer with friction, tests and video steps, which is exactly the response the government paper describes: raise the cost of sending, at the price of fairness for those who still apply by hand.

The pilot graveyard Symptom 9

MIT's NANDA group found in August 2025 that about 95 per cent of corporate generative AI pilots showed no measurable effect on the profit and loss account. Upwork found in July 2024 that 77 per cent of employees said the tools had added to their workload and 39 per cent spent more time reviewing AI output. Both fit the model above: generation was accelerated, checking was not, and checking is where the hours are.

The observation from the shop floor, checked

The claim: the flood makes the work doable only as a punishing one-person job, and only for those who can execute the thing themselves and do. What holds, and where a caveat belongs.

What holds. When the cost of a page falls to zero, length no longer signals effort, and the only thing that still separates a good concept from a bad one is whether somebody can and will execute it. That somebody is the person who knows how long a thing really takes, which is precisely the knowledge the model does not have. The data agree so far: the METR perception gap, the two hours per workslop case and the 39 per cent who review more all describe the same shift of hours from producers to the few who can judge. Scarcity moved from writing to judgment, and judgment cannot be copied.

Where the caveat belongs. First, the judge is flooded too; being the only one who can tell substance from filler makes a person the bottleneck, not the winner, and Upwork's 71 per cent burnout figure sits in the same survey. Second, checking can be partly automated as well, and the precedent is email spam: it reached about 90 per cent of all email traffic in 2009 and email survived through filters, sender authentication and reputation, not through heroics. Third, the gains are real where the task is narrow and the output verifiable; MIT's 5 per cent and the back-office cases exist. The observation is right about the present. Whether it stays right depends on whether organisations build the filters or keep asking one person to be the filter.

What may follow

Interpretation, not forecast. Three horizons, each with the observation that would show it to be wrong.

Near · 2026 to 2027

Friction comes back

Fees, caps, forms and page limits return, in the order the paper predicts: offices first, platforms second, firms third. "Who executes this and when" becomes the acceptance question. A polished document stops being evidence of work; the workslop survey shows that the reputational bill already lands on the sender. Inside teams the hours move to the few who can judge, response times lengthen and silent triage spreads, the unanswered email as a filter. This is the punishing-job phase.

Would contradict it: the number of documents per person falls again, or the minutes per document fall because filters work.

Middle · 2027 to 2030

Verification becomes the product

Value migrates from producing to accepting and rejecting: reviewers, testers, editors, whoever signs off. Agencies are paid for the prototype that runs, not the concept that reads well, and a tender asks for one page and a named person. Identity and reputation become the gate: signed senders, allow lists, structured submissions instead of prose, the paper's "agent-native interfaces". Agents also move to the receiving side, which turns human hours into compute costs and, by the logic of the Jevons paradox, raises volume further because sending keeps getting cheaper.

Would contradict it: the email path, standards and filters, makes the problem boring within a few years; or the productivity statistics show the gains that 95 per cent of pilots did not.

Far · 2030 and beyond

Two paths, both open

On the first path flooding becomes background noise handled by protocol, the way spam did. Text costs nothing and proves nothing, so proof moves into things that cannot be generated: a signature, a deposit, a working prototype, a delivered result. The cost of proof becomes an ordinary line in every cost structure, like insurance. On the second path institutions ration by price and priority: whoever can pay for review is served, whoever cannot waits, and public services face the trade-off the paper names, access against protection. In cost terms the flood is an externality, and the four classic answers are price, quota, standard and liability. Which of the four wins decides who bears the cost.

Would contradict it: models whose output other models can verify reliably and cheaply, which would remove the asymmetry this whole tab rests on.

What a firm can do now

Six acceptance rules that follow from the arithmetic above. None of them needs a new tool.

Page one decides Rule 1

Ask for the one page that would survive if the other 49 were deleted: decision, cost, owner, date. Everything else is an appendix, and appendices are not read before the decision.

A name and a date from the person who executes Rule 2

Timelines count only when they come from the person who will do the work, not from the tool. If the concept says three days, ask who has the three days.

Pay for outcomes, not pages Rule 3

Agency contracts on prototypes, milestones and things that run. Length is no longer a proxy for effort, so stop paying for it.

Count what comes in Rule 4

Measure documents per week and minutes per document; the flood line above is a management figure, not a feeling. What is measured can be capped.

Protect the checker Rule 5

The person who can judge is the scarce resource. Give them blocked time, a queue and the right to reject without reading to the end.

Agents at the door, humans at the decision Rule 6

Use machines to triage, summarise and check claims against sources; keep the acceptance human and accountable. The sender's agent against your agent is a cost race you should not run with people.

The building blocks, cleanly separated. Rule of thumb: revenue flows from sales, income is the umbrella term, costs split into fixed and variable, and profit is what remains.

Revenue / sales revenue income statement · flow

Value of the products and services sold in a period: price × quantity. In everyday use, "revenue" and "sales revenue" are used interchangeably.

Revenue = price × quantity sold

Income umbrella term

Total value added in a period. Comprises sales revenue plus, for example, interest and investment income, rental income or disposals of assets. All revenue is income, but not all income is revenue.

Fixed costs volume-independent

Incurred no matter how much or how little is produced: rent, salaries, depreciation on plants and machinery, interest, insurance, base IT costs. Hard to cut in the short term ("sticky costs"), which is exactly why they are dangerous in a downturn.

Variable costs volume-dependent

Grow and shrink with production volume: materials and supplier parts, production energy, freight, packaging, sales commissions. When sales collapse, they automatically fall too: they "breathe".

Contribution margin management metric

Revenue minus variable costs. It must first "cover" the fixed costs; only after that does profit begin. The contribution margin ratio determines how quickly revenue changes translate into profit changes.

Contribution margin = revenue − variable costs

Profit result

What remains after deducting all costs. Operating result (EBIT): before interest and taxes, it measures the core business. Net income: after interest and taxes. When negative it is called a loss, and the fixed costs keep running regardless.

Profit = revenue − variable costs − fixed costs

Liabilities balance sheet · stock

Debts owed to suppliers, banks or the tax office at the reporting date. They are not costs, but they must be serviced regardless of revenue. High fixed costs + high liabilities = double risk: a revenue collapse then hits the result and liquidity (the ability to pay).

Break-even profit threshold

The revenue at which profit is exactly zero: the contribution margin exactly equals the fixed costs. The distance between current revenue and break-even is a business model's "margin of safety".

Break-even revenue = fixed costs ÷ contribution margin ratio

Contribution margin ratio ratio

The contribution margin as a share of revenue, so the part of every euro of revenue that is left over to pay the fixed costs. It is the number that decides how quickly a business reaches its break-even, and it is the second input to the break-even formula above.

Contribution margin ratio = 1 − variable cost ratio

Margin of safety buffer

How far revenue can fall before the loss zone begins, measured from today's revenue down to the break-even. A firm at 30 per cent has room; a firm at 5 per cent is one bad quarter away from a loss. This is the KPI in the first tab, and the fixed-cost lever is what makes it shrink.

Margin of safety = (revenue − break-even revenue) ÷ revenue

Operating margin profitability

The operating result as a share of revenue, before interest and taxes. It says what is left of every euro of revenue from the core business, and it is the figure most often quoted about a company. Volkswagen ran at 3.8 per cent in the first half of 2026; a machine builder in a good year runs at eight to twelve.

Operating margin = operating result ÷ revenue

Depreciation and useful life fixed cost by decision

A machine bought once is charged to the result over the years it is expected to serve. That expectation is the useful life, and it is an assumption, not a measurement. A shorter life means a higher annual charge and therefore a bigger fixed block. This is why the AI tab moves one slider and the whole break-even shifts: the same hardware, written off over three years instead of six, doubles the annual charge without a single thing changing in the real world.

Annual depreciation = investment ÷ useful life in years

Operating Leverage lever

The higher the share of fixed costs, the more strongly profit reacts to revenue changes, in both directions. A DOL of 2.8 means: 1% less revenue → 2.8% less profit.

DOL = contribution margin ÷ profit

Price elasticity of demand market

Measures how strongly volume reacts to price changes. |ε| > 1: elastic, price increases cost a disproportionate amount of sales. |ε| < 1: inelastic, price increases largely stick (e.g. necessities).

ε = % change in volume ÷ % change in price

Liquidity survival

The ability to pay bills as they fall due. Companies rarely fail because of a single year's loss; they fail when the money runs out. Fixed costs and liabilities determine how long a revenue collapse can be endured.

Costs vs. expenses distinction

"Expenses" is the income statement term (everything that reduces equity), "costs" the term used in internal management accounting (valued consumption for producing output). For this tool the two are treated as identical for simplicity.

Agentic flooding 2026 term

A surge in the volume or complexity of requests that a service receives, caused by AI agents, that substantially strains the service (Schmitz, Hammond and Chan, August 2026). Coined for public offices, but the mechanism is general: whoever can generate cheaply floods whoever must check expensively.

Verification asymmetry mechanism

Producing a text got cheap, checking it did not. The older folk version is Brandolini's law: refuting nonsense takes an order of magnitude more effort than producing it. In cost terms the sender's variable cost per page fell toward zero while the receiver's stayed where it was.

Checking hours = documents × pages × minutes per page ÷ 60

Workslop symptom

AI-generated work that looks polished but lacks the substance to advance a task, so that the receiver has to redo or decode it (Stanford and BetterUp, September 2025). The polish is the problem: it hides the missing work until somebody reads to the end.

Externalised cost economics

A cost that one party causes and another bears, the classic example being pollution. The flood is one: the sender's saved hour reappears as the receiver's checking hours. The four classic remedies are a price (fees), a quota (caps), a standard (formats) and liability (you answer for what you send).