01 Trigger Event
On November 19, 2026, Kuaishou disclosed on its third-quarter earnings call that Kling 4.0 had launched fully on October 15, followed by the opening of its API to global enterprise clients and developers on October 28. Since launch, Kuaishou said Kling AI revenue had increased, while both user scale and ARPPU continued to grow.
The revenue mix also shifted clearly. Cheng Yixiao described two growth engines: API usage by enterprise clients on the B2B side and paid subscriptions from professional creators on the consumer side.
The original report did not disclose Kling AI’s specific revenue, API call volume, number of paying users, or gross margin. What can therefore be confirmed is the direction of growth and the composition of revenue—not the scale of either.
02 What This Really Means
The issue is not that “the video model has upgraded again.” It is that Kling 4.0 has begun crossing two commercial thresholds at once: professional creators are willing to subscribe, and enterprise developers are willing to integrate its API into production workflows.
The economics of these two customer groups differ.
Subscriptions on the consumer side primarily test whether model output can directly improve creators’ productivity and income. For video generation, willingness to pay usually comes from quantifiable content value—advertising creative, short-drama shots, or commercial campaigns—rather than the novelty of the generation experience alone.
Enterprise APIs, by contrast, test stability, throughput, cost, and integration capability. Once film, advertising, gaming, and embodied intelligence teams place video models inside their workflows, demand is no longer one-off creation but sustained invocation.
This is where aggregation theory begins to operate. Kuaishou is not merely selling a model capability; it is embedding Kling into the workflows of creators, enterprises, and developers. Whoever controls the workflow entry point is closer to distribution. Once customers build content pipelines, review processes, and asset libraries around Kling, switching costs no longer derive solely from model quality.
However, no data was provided on Kling’s internal API costs, task success rates, or latency. It is therefore impossible to determine whether the revenue increase resulted from improved 4.0 quality, promotions, pricing changes, or organic growth in usage. I may be wrong on this point.
03 Historical Analogy / Structural Comparison
AWS in 2014 and the iPhone in 2007 offered two distinct structural signals.
AWS turned computing capacity into infrastructure that could be purchased on demand. Customers no longer had to buy servers first; they paid per use. By opening its enterprise API, Kling is similarly moving video generation from a product feature toward an embeddable software component.
The iPhone first aggregated apps, content, and payments, then expanded distribution through its developer ecosystem. Kling’s two-engine model follows a more comparable path: consumer creators generate content supply and usage habits, while enterprise APIs carry model outputs into film, advertising, and gaming workflows.
The key is not user scale. Generative AI products can easily attract large numbers of free users while consistently failing to convert them into paying customers. The strategically meaningful signal is that ARPPU is growing alongside the business: at least according to the earnings call, Kling is not diluting average revenue per user merely by adding users.
This also differs from the path taken after ChatGPT went viral in 2022. ChatGPT first won broad consumer distribution. Kling is instead betting directly on the commercial value of video production. Video generation is more compute-intensive and has a steeper inference-cost curve, so every quality improvement must clear the hard economic test of token economics.
What will ultimately be priced is not the ability to generate video in itself. It is whether Kling can become the default infrastructure within a production pipeline as the cost per shot falls.
04 What It Means for AI Builders
There is no need to migrate existing workflows immediately, but teams should begin testing whether the Kling 4.0 API belongs on the production candidate list.
First, prioritize use cases with clear financial value rather than generic content generation. Advertising creative variations, short-drama previs, and game concept shots are more likely to reveal return on investment than entertainment content for social media.
Second, evaluate Kling as a complete inference pipeline. Beyond the price of a single generation, record queue time, task success rate, retry count, time spent on human revision, and the percentage of final shots that are usable.
Third, establish model routing rather than binding the business to a single provider. Route different tasks to Kling and other video models according to quality, latency, and cost. Preserve prompts, seeds, input assets, and version information so results remain reproducible.
Fourth, consumer growth is not proof of enterprise demand. The two markets may expand simultaneously while requiring entirely different products. Company-wide sales still need to address enterprise API authentication, concurrency, SLAs, data retention, and auditing.
For independent AI entrepreneurs, the more actionable opportunity may not be to train another general-purpose video model. It may be to build vertical workflows on top of the Kling API: batch production of advertising creative, short-drama shot planning, film previs, or iteration on game assets. As the supply of foundation models increases, arbitrage windows at the application layer expand because generic capabilities are no longer scarce; workflows and distribution become the barriers.
05 Counterargument / Risks
I may be overestimating the strength of the “revenue increase” signal. The original report provided no baseline, sequential, or year-over-year figures, and ARPPU was not quantified. For a product that has previously driven growth through free usage allowances, promotions, and bundled plans, higher revenue does not necessarily mean improved unit economics.
The second risk is revenue quality. Enterprise API usage and consumer subscriptions sound healthier, but enterprise traffic may be concentrated among a small number of major customers, while project-based revenue may conceal low margins or high operating costs. GPU consumption, failed retries, and human review associated with video models can all erode reported revenue.
The third risk is competitive response. Google Veo, OpenAI’s related video capabilities, and open-source video models may cut prices or catch up during Kling 4.0’s window of quality leadership. If model leadership cannot be converted into a data flywheel, developer ecosystem, and low-cost inference, it is only a short-term product advantage.
There is another overlooked layer: film, short drama, advertising, gaming, and embodied intelligence are scenarios that have been “entered,” not necessarily scaled into paying markets. Kling may also simply be repackaging the same generation capabilities as multiple industry narratives, with actual revenue still concentrated in creator subscriptions inside the platform.
My current judgment is therefore that Kling 4.0 demonstrates the emergence of a dual commercial structure for video generation—enterprise APIs and consumer subscriptions—but does not yet prove that Kling has established a moat. The indicators that now matter are the API share of revenue, paying-customer retention, inference gross margin, and repeat invocation volume.
If these data remain unavailable, continued revenue growth may only reflect AI video demand being pulled forward. If they improve together, Kling will no longer be merely a generation tool; it may become infrastructure within the video-production stack.