Churn and Burn: Why 93% of Ads Run on Under $10K of Media Spend

What 633,000 ads say about churn and burn

Churn and burn is the 2026 advertising model: brands produce far more assets than their media budgets can support. CreativeX data covering 633,000 ads shows content production up 29% while media spend per ad fell 15%, leaving 93% of ads running on under $10,000. For commercial video teams, that gap is now the brief.

The figures come from a CreativeX study of 633,000 ads across two years, published in a WARC interview in July 2026. Content production rose 29%. Media spend per ad fell 15%. And 93% of ads ran on less than $10,000 of media - a long tail that still absorbed 30% of the total media budget. WARC's label for the pattern is churn and burn: manufacturers of content outrunning the money that makes content work.

The last number is the one production teams should sit with. The under-backed majority is not a fringe at the edge of the market; it is where three in every ten media dollars go. An asset that ships without media behind it rarely fails loudly. It quietly burns a slice of budget, delivers a fraction of the frequency a backed asset would, and returns almost nothing the brand can reuse.

None of this argues for making less video. It argues that the production plan and the media plan have come apart. When the two were welded together - one hero film, one flight, one quarter - volume and backing grew at the same rate. Generative tooling broke that weld by making the asset the cheapest part of the system and leaving the spend untouched.

Why CMOs keep ordering more content anyway

If the economics are this lopsided, why does the volume mandate keep arriving? Dentsu Creative's 2026 CMO report, Making It Real, surveyed 1,950 senior marketing decision-makers across 14 markets in April 2026 and found an industry that can see the problem and cannot stop. Some 76% of CMOs worry their teams are producing more content for less impact, and 74% agree that more assets do not necessarily mean more impact.

Demand pulls the other way. 81% of CMOs expect to need far more content in future, and 71% say they do not know how to balance the cost of extra assets against what those assets return. AI has improved speed to market for 71% of respondents, yet 70% still have not seen major cost savings - the hidden labour of workflow integration and brand training eats the gain before it reaches the P&L.

Expectations have quietly reset as well. The share of CMOs who expect generative AI to greatly help them produce assets quickly and cheaply fell from 89% in 2025 to 82% in 2026. The drop is small in percentage points and large in meaning: the industry has stopped confusing cheap production with cheap marketing, and the next brief you receive will reflect that.

There is a quieter cost as well. Every unbacked asset still consumes review cycles, version control and approval gates - the human labour Dentsu's respondents keep naming as the reason savings have not arrived. Volume that no one can fund is volume that someone still has to manage, and that management shows up as the overhead that eats the efficiency case.

The production-side trap: volume is cheap, backing is not

For a commercial video team this looks like an opportunity. Every brief now carries a volume number - more cuts, more aspect ratios, more variants - because production finally feels free. The impression ceiling has not moved with the render queue, though, and as our breakdown of {{link}} explains, unlimited variants redistribute the same finite attention rather than creating more of it.

Auction mechanics reward volume from the other side as well, and our piece on {{link}} explains how generative output became the cheapest way to manufacture auction surface. Both effects are real. Both also carry a precondition the churn-and-burn data exposes: an ad can only win an auction it actually enters, and 93% of ads never get the media backing to enter at scale.

The arithmetic is unforgiving. A brand that ships fifty variants a month on a budget that can fund five has not built a testing system; it has built a warehouse. Each unbacked asset carries a production cost that is now near zero and a media cost that is not, which means the marginal asset is the one most likely to starve. The fleet grows. The fleet's performance does not.

The trap springs shut because the incentives point the same way. Production is rewarded for throughput, the agency is rewarded for retaining the volume mandate, and no one in the chain is rewarded for asking what the asset will meet when it goes live. Breaking that alignment does not require a new tool; it requires a question in the brief that someone is accountable for answering.

The impression ceiling has not moved with the render queue, though, and as our breakdown of AI video frequency capping explains, unlimited variants redistribute the same finite attention rather than creating more of it.

Auction mechanics reward volume from the other side as well, and our piece on creative auction weight explains how generative output became the cheapest way to manufacture auction surface.

Where the savings actually go

The same Dentsu data shows what disciplined teams do with the efficiency AI did deliver. Some 82% of CMOs say AI-driven production efficiencies let them put more money into upfront creative craft, and 80% say the savings help fund real-world experiences. The money is moving to the two places volume cannot substitute: the idea, and the live touchpoint.

That reinvestment only works if an organisation can move money across silos, and here the survey is blunt: 77% of CMOs believe agency structures that separate creative, media, data and production stop content from working as hard as it should. A production team that never sees the media plan cannot know which assets deserve backing. A media team that never sees the pipeline cannot plan around what is coming.

Measurement is where the two sides meet. CreativeX's client work makes the point: Nestle mapped nine creative best practices against a single business metric, lifted its overall creative quality score by 15%, and reported 66% ROI on Meta for ads with a high score. Quality scoring does not replace media judgement, but it gives production and media a shared language for which assets earn support - and which are headed for the 93%.

For production companies the signal is direct. The budgets that remain are moving toward craft and experience, which means the work that wins is the work that raises the quality of a small number of assets rather than the work that multiplies a large number of them. That is a different sales conversation than the one most teams are still having.

What commercial video teams should do about it

The practical response starts with how the work is priced. Usage-based models charge per second of generation rather than per finished asset, which ties the invoice to volume - the exact metric the 2026 data says stopped paying, as our breakdown of {{link}} spells out. Price the assets that earn media, not the ones that merely exist.

Second, split the pipeline by destination. Route most variants into a structured test programme - cheap, fast, disposable - and reserve a small hero set for assets that will actually carry media. Learning speed, not output volume, is what separates teams now, and our breakdown of the {{link}} covers how fast test-measure-double-down cycles beat fleet size. A variant that never faces an audience is not an asset; it is inventory.

Third, write the media plan into the brief. Before a shoot or a generation batch, ask what spend the finished asset will carry and what frequency ceiling it will hit. If the answer is 'we will decide later', the asset is already headed for the 93%. Teams that can answer the question get briefed first, because they are the only ones who can tell a client whether the volume being ordered will ever be seen.

Churn and burn is not an argument for making less video. It is an argument for making video that gets bought. The 2026 data hands production teams something rare: evidence that the asset is only half the job, and a commercial reason to charge for the half that actually works.

Usage-based models charge per second of generation rather than per finished asset, which ties the invoice to volume - the exact metric the 2026 data says stopped paying, as our breakdown of usage-based billing spells out.

Learning speed, not output volume, is what separates teams now, and our breakdown of the learning-speed moat covers how fast test-measure-double-down cycles beat fleet size.