The bill arrived

Here is a pattern that does not require a chart to see. Sora launched free, bled money, died. Grok Imagine launched free, went silent overnight, erected a paywall. Seedance launched globally, hit copyright walls, retreated into CapCut with geographic restrictions and face generation blocked. Runway expanded into gaming and third-party hosting. Multiple platforms are quietly cutting free tiers, throttling daily limits, or pivoting entirely to enterprise contracts. Industry analysts are calling it a divestment. The word is generous. What is happening is simpler: the check bounced.

Generating a four-second video clip requires orders of magnitude more GPU cycles than generating a paragraph of text or a static image. Every frame demands fresh, high-fidelity computation. Motion coherence, lighting consistency, physics simulation, audio synchronization. None of it caches. None of it reuses. Every generation is a new burn.

The early pricing was never cost-reflective. It was a customer acquisition strategy wearing a product's clothes. Give away the expensive thing. Build the habit. Lock in the workflow. Raise the price later, or find a business model that monetizes the attention. This works for software where the marginal cost of one more user approaches zero. It does not work when every user interaction costs real money in GPU time, electricity, and cooling infrastructure.

The math was always wrong

Sora generated $2.1 million in lifetime revenue. That number tells you everything about the economics. OpenAI was running one of the most compute-intensive consumer products ever built and collecting less money than a successful food truck. The Disney deal was supposed to subsidize the gap with content licensing revenue. Disney walked. The app died the same week.

Grok Imagine generated 1.245 billion videos in January alone. At any meaningful cost per clip, that is an infrastructure bill measured in tens of millions of dollars per month. Musk posted "available free worldwide for a limited time" eight months before the paywall went up. The limited time lasted exactly as long as the growth metrics justified the burn rate. Once the user acquisition curve flattened, the economics became indefensible.

The pattern repeats across the industry. Platforms that promised unlimited generation at ten or twenty dollars a month discovered that their heaviest users cost them multiples of the subscription fee. The light users subsidized nothing because there were not enough of them. The economics of all-you-can-eat video generation are the economics of an all-you-can-eat restaurant where every plate costs forty dollars to prepare and the buffet costs fifteen.

The retreat is structural

This is not a temporary correction. The cost curve for video inference is improving, but not fast enough to rescue the subsidized model. Hardware gets faster. Models get more efficient. Prices drop. But the expectation of free or near-free generation was established during a period of venture-capital-funded abundance, and the gap between user expectations and sustainable pricing is still measured in multiples, not percentages.

What is left standing when the subsidies drain out? Two categories of product.

The first is enterprise. Higher prices, usage caps, SLAs, custom deployments. Google putting Veo in Workspace. Adobe wrapping thirty models inside a Creative Cloud subscription. Netflix acquiring InterPositive for internal use. The enterprise customer pays enough per seat to cover the compute, and the platform can forecast costs against revenue. This is sustainable. It is also, for an independent filmmaker, irrelevant.

The second is pass-through. Pay the provider directly for what you use. No subscription pool. No credit system that expires. No platform sitting between you and the GPU, adding margin. The cost of each generation is transparent, immediate, and proportional to what you asked for. This includes API-direct access, BYOK architectures, and staking models like Venice where users supply the economic input that funds the compute.

Everything in between is vulnerable. The fifteen-dollar-a-month plan with "200 generations included" is a bet by the platform that most subscribers will not use their allotment. When they do, the platform loses money. When they do not, the platform profits from selling something it never had to deliver. Either way, the pricing has no relationship to the actual cost of the service. It is a financial product dressed as a creative tool.

What the retreat means for filmmakers

If your workflow depends on a platform's pricing model, your workflow depends on a quarterly planning meeting you were not invited to. This series has documented that lesson in real time. Six disruptions in eight weeks. Each one changed the terms of access without changing the capability of the models.

The models are fine. Kling still renders physical detail at 4K. Veo still finds the atmospheric light. Runway still follows instructions with literal precision. WAN still defaults to saturation and visual density. Seedance still preserves reference frames. Grok Imagine still fills every vacuum with spectacle. None of them got worse. The business structures around them shifted. The filmmaker who invested in vocabulary lost nothing across any of these disruptions. The filmmaker who invested in a platform's generosity lost access to the generosity.

CinePrompt's BYOK architecture was not a philosophical statement about user freedom, though it is that too. It was an economic decision. When the user brings their own API key, the cost of generation is between the user and the provider. CinePrompt never sits in the financial path. Never collects a margin on compute. Never has to decide between subsidizing heavy users and throttling light ones. The value is the structured cinematography knowledge, and that value exists whether the user generates through fal.ai at market rates, through Venice with staking credits, or through a local ComfyUI installation where the only cost is electricity.

The 1,457 cinematography controls do not have a cost-per-use. The knowledge layer is independent of the generation layer. That separation was always the architecture. Now it looks like the architecture the economics demand.

What comes next

Prices will keep falling. They always do. NVFP4 compression, more efficient architectures, competition among hardware providers, open-source models running locally. The floor is still descending. But the floor and the free tier are different things. The floor is a real price reflecting real costs, trending toward pennies per clip. The free tier was an illusion reflecting a growth strategy, trending toward cancellation.

NAB 2026 opens in Las Vegas next week. At least one exhibitor is explicitly positioning against what they call "AI fatigue." The broadcast industry is the oldest professional video industry on the planet, and its practitioners are arriving at the same conclusion this series has been documenting: the tools are real, the hype was funded, and the funded part is receding. What remains is infrastructure and knowledge.

The subsidy era taught a generation of creators that video generation should cost nothing. That expectation will take longer to correct than the business models that created it. Some filmmakers will chase the next free tier. Some will invest in the vocabulary that makes every dollar of compute count. The first group will be disrupted again. The second will not.

The bill arrived. It was always going to. The question was never whether the free tier would last. The question was what you built while the door was open.

Bruce Belafonte is an AI filmmaker at Light Owl. He once watched a platform's entire business model evaporate between breakfast and lunch and his prompts still worked by dinner.