Counsel: request a conflict check →

Cloud · Economics · Infrastructure · Bots

Systems Up, Margins Down: The Cost of Processing Every Event

Most infrastructure failures do not take a system down. They keep it working on events no one needed, and the meters keep counting. An account of event processing, metered computing, and the cost of not asking what an event is worth.

SDShailin DharFutureproof TMT
Share
Published
November 18, 2026
Reading time
16 minutes
Format
Article
Exhibits
4 figures

When infrastructure fails, we ask one question: is it up? Dashboards, status pages and service agreements are built around it. But a system can answer yes while it repeats work, processes traffic that has no buyer behind it, and bills for every step. A system can remain available while its economics deteriorate. Availability and economic efficiency need separate tests.

This briefing is about the second test. Every business that runs software now processes events: page views, form fills, API calls, transactions, sensor readings, ad requests. Most of that processing is rented, and most of it is metered by activity rather than value. We describe how one event becomes many units of work, how layered services turn that work into charges, and what changes when a system asks what an event is worth before it processes it. We call that question value discernment.

Fig. 01 // Illustrative model
One event, two pipelinesMaximum attempts per event across four nested call boundaries. Illustrative.Without value discernment1 event1arrivesNo check—admitted as isCollector2max attemptsRouter4max attemptsAPI8max attemptsDatabase16max attemptsWith value discernment at the input (one shared retry budget)1 event1arrivesQualify1stop or passCollector1one attemptRouter1one attemptAPI1one attemptDatabase1–2retry once, hereAttempts are not successful writes, people or billable units. Each metered layer may still charge for them.
Fig. 01Without a value check, every layer retries on its own and the work multiplies. With one, unwanted events stop at the input and the rest share a single retry budget. The counts are maximum attempts, not successful operations or billable units.Source: Visual Systems Library, Futureproof TMT · Illustrative model, not measured data

01Two ways infrastructure fails

Infrastructure can fail a business along two paths. They overlap, and they are not a complete taxonomy: failures also affect integrity, security and safety. But they call for different measurements, and most organizations measure only the first.

Table 01 // Two exposure paths
PathWhat the business experiencesWhat to measure
Service denialRequests fail, slow down or cannot completeAvailability, latency, completion and recovery time
Denial of MarginUnproductive work consumes capacity, triggers charges or adds operating cost, while service continuesCost per qualified outcome, accepted billable units and contribution
An outage is not automatically an attack, and Denial of Margin is our proposed economic framing, not a standard incident classification.

02What last week showed, and what it did not

Three unrelated disruptions in mid-November 2026 made dependence visible. From the evening of 12 November, a Microsoft 365 outage left police forces in England and Wales unable to receive crime reports online and sent NHS England staff to paper records; Microsoft called it a “technical issue not related to cyber or AI agents,” and a UK minister said it “mainly affected back-office functions” (PA, via The Irish News). On 13 November the Federal Aviation Administration halted flights nationwide for nearly five hours after what it called “a temporary telecommunications outage.” In remarks quoted by FedScoop, officials said a router being replaced in a switch from copper to fiber “started flooding other routers in our air traffic control system,” making “over 1,000 connection attempts every four to five seconds”; the Department of Homeland Security saw “no indication of a nefarious attack” (FedScoop). On 16 November JPMorgan Chase attributed transaction delays to “a software bug in an on-premise hardware system,” adding that other banks were affected “only to the extent that they were a party to a transaction with us” (AP, via OPB).

These events establish interdependence: services with different names shared dependencies, and trouble in one place surfaced in another. The FAA case also shows amplification at work: one component’s repeated connection attempts became everyone’s problem. But the events do not establish a common cause, and no public source attributes billing losses to them. The FAA said it would share a root-cause analysis once complete. The margin argument below rests on separately documented mechanisms, not on these incidents.

03Six things we count as one

Much of the confusion about computing costs comes from using one word, “event,” for several different quantities. They diverge exactly when systems are under stress.

Table 02 // From arrival to outcome
TermDefinitionWhy it diverges
ArrivalSomething knocks: a request, a page load, a callAutomation can produce arrivals at any volume
AttemptA try at doing work for that arrival at one layerRetries multiply attempts across layers
OperationWork that actually completed: a write, a query, a sendFailed attempts still consume resources
Measured eventWhat a tool records and reportsDuplicate tags and instruments count the same arrival twice
Billable unitWhat a contract charges for: events, users, credits, gigabytesDefined by each vendor, not by the buyer
Qualified outcomeA result someone values: a sale, a lead, a verified viewThe only quantity tied to revenue
Value discernment is any step that estimates, before processing, whether an arrival is likely to become a qualified outcome.

04One event, more work

Distributed systems retry, and that is usually good engineering. The problem is where the retries happen. Amazon’s engineering guidance describes a call that passes through a five-deep stack of services: “If each layer retries independently, the load on the database will increase 243x, making it unlikely to ever recover,” because “the retries at each layer multiply” (Amazon Builders’ Library). Its recommended practice is to retry at a single point in the stack, with backoff and jitter.

Apply the same arithmetic to a hypothetical measurement workflow. One served ad impression passes from a pixel to a collector, a router, a measurement API and a database. Allow two attempts per call at each of four boundaries, exhaust every allowance, and the impression can generate up to 16 database-write attempts. That is not 16 successful writes, 16 people or 16 billable events; idempotency can prevent duplicate side effects. But the repeated requests still consume resources, and in a metered stack, resources are where charges come from.

Volume is also a poor proxy for burden. The Wikimedia Foundation found that bots produced about 35% of its pageviews but at least 65% of its most resource-intensive traffic (Wikimedia). And chargeability is specific, not universal: in 2024 AWS stopped charging for certain unauthorized requests that a customer’s storage rejected (AWS). A rejected request should not be assumed billable everywhere, and an accepted one should not be assumed free. Billing follows the agreement, not a universal rule.

05Meters on meters

Most of the work above runs on someone else’s computer. A software contract is also a relationship with a production system: processors, storage, networks and operators that may sit several vendors down. Four things that feel like one are separate: who holds the data, who controls the software, who owns the hardware, and who is liable for the bill.

Fig. 02 // Concept diagram
Clouds on cloudsEach layer runs on the one below and adds its own meterYour marketing functionAnalytics, CDP, tags, email, verificationYou see one invoiceSaaS vendorPrices the functionEvents · MTUs · profiles+Data and platform servicesWarehouse, queues, CDN, APIsCredits · requests · GB+Cloud infrastructureRented servers and storageInstance-hours · egress+Data center and hardwareProcessors, memory, power, coolingWhere the work happens+Every seam addsa meter, a margin and ahandoff no one fully ownsOpacityone invoice, five metersInefficiencyidle capacity, billedBloatduplicate tools and dataSecurity gapsshared, split responsibility
Fig. 02The buyer sees one invoice. Underneath it, each layer resells the one below with its own meter and margin.Source: Visual Systems Library, Futureproof TMT

The meters stack. HubSpot says it serves the majority of its platform from Amazon Web Services and reports an 84% gross margin (2025 10-K). Snowflake tells investors its margins are “significantly influenced by the prices we are able to negotiate with these public cloud providers” (10-K). AWS itself earned a 37% operating margin in 2024 and 35% in 2025, computed from Amazon’s segment results (10-K). The UK’s competition regulator found that AWS and Microsoft have earned returns “substantially above their cost of capital” and that egress fees are “a key commercial barrier” to switching (CMA, July 2025). None of this is wrongdoing. It is how a layered market prices.

The units are printed on the pricing pages, and almost all of them count activity. Marketing and data tools make the pattern easy to see:

Table 03 // How common software meters activity: examples
ToolWhat it countsPublished terms
Twilio SegmentMonthly tracked users, including anonymous visitors$120 a month for 10,000, then $10 per 1,000
MixpanelEvents$840 a month at 10 million events, billed annually
SnowflakeCompute credits, billed per second60-second minimum; credits double with warehouse size
HubSpot Marketing HubMarketing contactsNo contact-tier downgrades mid-term
AWS (underneath many of them)Data transferred out$0.09 per GB for the first 10 TB a month
Vendor pricing pages, accessed 13 November 2026. Prices vary by tier, region and contract.

An anonymous visitor is a tracked user. A bot that fires an event is an event. A scripted form fill is a contact, and on some contracts the tier it pushes you into does not come back down until renewal. None of these meters asks whether the activity was worth anything. Layering adds four side effects: opacity, because no vendor reconciles its slice to the hardware; inefficiency, because each layer provisions its own headroom (Kubernetes clusters averaged 8% CPU utilization in 2025, per Cast AI); bloat, because duplicate tools record the same activity again; and security gaps at every seam, as the 2024 theft of credentials for about 165 organizations’ Snowflake accounts showed (Mandiant).

06Exposure by company size

The bill scales with the company; the exposure does not shrink with it. Large enterprises waste the most in absolute terms. Smaller ones have the least room to absorb a surprise and, according to the CMA, pay the most to leave.

Table 04 // Metered software and cloud spend by company size
Company sizeUnused SaaS licenses a yearPublic cloud spendWhat else the evidence shows
Small$3.8M at 1–500 employees63% of firms under 1,000 employees spend $1.2M a year or lessManaging cloud spend is the top challenge for 86% of SMBs. Egress fees weigh most on smaller customers (CMA).
Mid-size$9.5M at 501–2,500; $29.8M at 2,501–10,000No separate public breakdownSaaS apps at firms of 1,500–4,999 employees rose 41% in 2026, from 116 to 164 (BetterCloud).
Large$80.6M at 10,001+40% of firms over 1,000 employees spend more than $12M a yearZylo customers, a large-company sample, average $55.7M a year on SaaS (median $20.6M).
Everyone—29% of cloud spend self-reported as wasted; spend ran 17% over budgetSaaS prices rose 16.4% in June 2026, more than four times US consumer inflation (Vertice).
Sources: Zylo waste by size, 2026 and 2026 SaaS Management Index (Zylo customers, who skew large; the 1–500 band is not typical of small businesses); Flexera State of the Cloud 2025 and 2026 (the 63% is our sum of Flexera’s bands); BetterCloud; Vertice SaaS Inflation Index; CMA. Each source draws size bands differently, so compare within a column, not across.

07From availability risk to billing risk

On servers you own, cost is mostly fixed. A flood of traffic cannot raise the bill; it can only exhaust capacity. That is the classic threat, distributed denial of service (DDoS): the system goes down. On metered infrastructure the same flood meets no ceiling. Services scale, stay up, and send the invoice.

Fig. 03 // Illustrative model
Same flood, two billsMonthly cost as traffic rises to five times normal. Illustrative units.Owned capacityBare metal: fixed cost, hard ceiling2461×3×5×Traffic, multiples of normalFloodOver capacity:slows or failsCapacity ceilingFlood = availability risk (DDoS)Metered servicesCloud and SaaS: cost follows traffic2461×3×5×Traffic, multiples of normalFloodCheaper at normal traffic5× the bill, still upFlood = billing risk (EDoS)
Fig. 03Owned capacity turns a flood into an availability problem. Metered capacity turns it into a billing problem. Metered is cheaper at normal traffic, which is why the trade looks free until the flood arrives.Source: Visual Systems Library, Futureproof TMT · Illustrative model, not measured data

Security strategist Christofer Hoff named this in 2008: traffic “designed to make seemingly legitimate requests for service to generate an economic denial of sustainability (EDoS)” by scaling service “beyond the economic means of the vendor to pay their cloud-based service bills” (Rational Survivability). Researchers have since formalized it, from attributing EDoS in public clouds (Karami, Wang and Chen, 2018) to “denial of wallet” in serverless computing (Kelly, Glavin and Barrett, 2021). The documented cases share one feature: the bill arrived before anyone decided the traffic was a problem.

Table 05 // When unwanted traffic became an invoice
CaseWhat happenedOutcome
Empty private storage bucket, 2024About 100 million denied requests from misconfigured software ran up more than $1,300 in a day (The Register)AWS stopped charging for certain unauthorized requests (AWS)
Read the Docs, 2024One AI crawler downloaded 73 TB in a month: over $5,000 in bandwidth charges (Read the Docs)Blocking AI crawlers cut download bandwidth by 75%
Stolen API keys, 2026Unauthorized use of Google AI APIs left customers with bills above $10,000 (The Register)Customers fought for refunds case by case
Only the last case involved an attacker. The others were software and crawlers doing what they were built to do.

08The flood changed motive

Automated traffic has not become friendlier. On the vendors’ own definitions it has become more hostile: in Thales’s decade series, bots labeled “bad” rose from about 20% of all web traffic in 2016 to 40% in 2025, while humans fell to 47% (2026 Bad Bot Report). Volumetric attacks have not gone away; Cloudflare mitigated 47.1 million DDoS attacks in 2025, more than double the year before (Cloudflare).

Fig. 04 // Data
Bad bots doubled their share of web trafficShare of all web traffic by source, 2016–2025Bad botsGood botsHumans0%50%100%2016: bad bots 20%, good bots 19%, humans 61%20162017: bad bots 22%, good bots 20%, humans 58%20172018: bad bots 20%, good bots 18%, humans 62%20182019: bad bots 24%, good bots 13%, humans 63%20192020: bad bots 26%, good bots 15%, humans 59%20202021: bad bots 28%, good bots 15%, humans 58%20212022: bad bots 30%, good bots 17%, humans 53%20222023: bad bots 33%, good bots 18%, humans 50%20232024: bad bots 37%, good bots 14%, humans 49%20242025: bad bots 40%, good bots 13%, humans 47%202520%Bad 40%Good 13%Human 47%“Bad” is the vendor’s label. It includes scrapers, scalpers and ad-fraud bots, not only attacks.
Fig. 04The share of web traffic labeled “bad” doubled in a decade. The label covers scrapers, scalpers and ad-fraud bots as well as attacks. The 2023 value is 33% in this chart and 32% in earlier editions.Source: Visual Systems Library, Futureproof TMTData: 2026 Bad Bot Report, global internet traffic profile 2016–2025, Thales (Imperva)

What changed is what “bad” means. In 2016 Imperva wrote that “the most active bad bots are all impersonators used for DDoS attacks” (Imperva). By 2025 the largest categories were scraping and general automation, API abuse and business-logic abuse, and 41% of bad bot traffic declared itself as Chrome. Alongside it runs traffic that is not an attack at all: about 80% of AI crawling is for model training, and one AI company’s crawlers requested pages as many as 500,000 times for every visitor they referred back (Cloudflare). The modern flood is less often trying to take a system down and more often trying to take something from it, at the operator’s expense.

09Denial of Margin, with and without value discernment

We use Denial of Margin for avoidable economic degradation from unproductive work: activity whose processing cost exceeds the value it creates. It can arise from mistakes, routine automation or malicious activity; intent must be established separately. A Denial-of-Margin attack is EDoS with evidence of intent. Passive Denial of Margin needs no attacker: retries, duplicate tags, crawlers and synchronization jobs, each one billable, none of them a customer. Not every infrastructure failure produces it, and not every instance requires a failure.

The economics reduce to one comparison. Let D be the downstream cost an arrival incurs once admitted, retries included, and q the share of arrivals that are worth processing. Without discernment, every arrival costs D. With a qualifying check that costs c per arrival, the expected cost per arrival is c + (q × D). Discernment pays whenever the check costs less than the waste it prevents: c < (1 − q) × D. Two forces make that inequality easier to satisfy each year: layered meters raise D, and automated traffic lowers q.

No check is perfect, and the errors are not symmetric. Rejecting a valuable arrival loses revenue that never shows up in a report; admitting a worthless one shows up in every meter it touches. We examine that trade-off in Scam Likely, and why a bounded check is cheaper than an open-ended hunt in Detection vs. Verification. It is the same infrastructure-drag loop we described in The Traffic Monetization Paradox, now with a price on each step.

10Ownership changes the curve, not the problem

Owning compute changes the shape of cost, not its existence. Owned capacity exposes a business to DDoS; metered capacity exposes it to EDoS. The first fails visibly, the second fails on the invoice. Large, steady workloads often run cheaper on owned hardware; spiky, early or specialist workloads are often cheaper rented, and many organizations can recover waste without moving anything. We take up when each wins in a forthcoming briefing, Cloud Economics Are Counterintuitive. Owned, rented and hybrid systems can all waste work and all fail. Value discernment is useful under every model, because it reduces the work before the cost structure applies.

11What to measure

Uptime remains necessary. It is not sufficient. The measure that connects infrastructure to the business is cost per qualified outcome: the full processing cost of the arrivals behind each sale, lead or verified view, across every layer that touched them. Five questions bring it into view:

  • Which inputs can create cost, and who has authority to initiate them?
  • Which tools share an underlying provider, identity service or data dependency?
  • What happens to retries, queues and spend when a dependency slows or fails?
  • Which records substantiate the units on each invoice, and which cannot be reconciled?
  • What changes when redundant, invalid or otherwise unproductive activity is removed?

Keep efficiency and accounting apart. Fewer events do not automatically mean lower invoices: minimum commitments, fixed licenses, reserved capacity and tier rules can hold the bill steady. Report freed capacity, avoided future spend and realized savings separately, and measure any legitimate demand a filter suppresses. Match each control to its cause. Engineering controls include bounded retries, backoff with jitter, deduplication, idempotency and queue limits. Commercial controls include eligible-unit definitions, reconciliation rights, disputed-charge procedures and spend limits that actually stop spending. Alerts alone are not caps.

12Limits of this account

  • This is a mechanism-based account, not an estimate of market-wide losses or of how often margin failures occur. Public sources do not establish that margin-only failures are more common than outages; the defensible claim is that economic degradation can exist without service loss, and an uptime dashboard alone will not detect it.
  • The retry illustration is hypothetical. Recent incident details rely on attributed reporting and may change when providers publish root-cause reports.
  • High traffic is not proof of fraud, and clean traffic is not proof of incrementality.
  • Denial of Margin is proposed terminology, not an established security classification.

13Measure availability. Audit the outcome.

The cloud did not make computing expensive. It made computing billable by the unit, by someone else, several times over. That was a good trade when most units were customers. It is a worse one when much of the traffic is automated, retries multiply quietly, and the meters cannot tell the difference. The systems that hold their margins will be the ones that ask what an event is worth before they pay to process it.

“A system can remain available while its economics deteriorate.”

— Shailin Dhar, “The System Is Up. Your Margin May Be Down.” (2026)

14Sources

Filed underCloudEconomicsInfrastructureBots

Futureproof TMT

Seeing something like this in your own data?

Our team advises enterprises, counsel, and families on the automated internet.

Start a conversation

Essential reading

All briefings