HomeArticle

Is AI video entering the DeepSeek moment?

AIX财经2026-11-19 19:39
The gap among models is narrowing, but the gap between major tech players is widening.

The AI video industry is approaching a moment similar to that of DeepSeek, where the gap in model capabilities is narrowing, prices have dropped to the level of cents per second, and creators have started to call related services in batches. 

In early November, Google officially released its video model Gemini Omni 1.2 Flash, which not only supports functional upgrades such as 15-second video generation and multi-turn conversational editing, but also cuts prices. The 720p resolution tier is priced at $0.07 per second (about 0.47 yuan per second), 30% lower than the previous generation. Calculated based on the official API, the cost of a 720p video of the same duration is about one-third of that of the same tier of ByteDance's Seedance 2.5. 

Prices in China are also trending downward. In early October, Shengshu Technology released Vidu Q4 Preview, with its debut preferential price reduced to 0.09 yuan per second. 

However, just over half a year ago, this business just eliminated a heavyweight player. This March, OpenAI shut down Sora. When this product was launched in September 2025, it exceeded one million downloads faster than ChatGPT, but it only operated for half a year. According to media disclosures, Sora lost about 1 million US dollars per day before its shutdown. Industry data estimates that the computing power cost of a 10-second video may be equivalent to thousands of ordinary text conversations. 

While OpenAI stepped back, major domestic manufacturers are accelerating their layout. ByteDance's Seedance 2.0 was exposed to have a monthly revenue of more than 10 billion yuan. Although Tan Dai, President of Volcano Engine, also publicly responded that the publicly circulated revenue data is "generally on the high side", there is no doubt that it has become one of the important starting points for ByteDance's AI commercialization. Kuaishou split Keling out as an independent entity, which completed nearly 3 billion US dollars in financing with a post-investment valuation of about 18 billion US dollars. 

For this same capital-intensive business, OpenAI chose to exit, while Google, ByteDance and Kuaishou are still increasing their investment. Are they calculating the same account? 

01. The cost account of AI video: Revenue is generated, but the business is not yet profitable

First of all, there are four main revenue sources for current AI video products: C-end subscriptions, APIs, enterprise projects (providing customized models, private deployment and long-term services), and ecological cooperation (technical services, project revenue sharing). 

Among them, subscriptions and APIs cover the widest range of users, and are also the most important revenue source for all players. Gemini Omni 1.2 Flash can be called through API, or used directly in Gemini; ByteDance's Seedance collects API fees through Volcano Engine, and charges through membership or points when using the Jimeng platform; Keling AI's revenue in the first half of 2026 exceeded 1.5 billion yuan, with enterprise APIs and creator subscriptions as the main sources. 

Short dramas are an important driving force behind the API increment. Chinese Online disclosed in June that among all Token consumption shares in China, AI short drama and video generation ranks first, accounting for as high as 55%. According to 36Kr reports, Seedance's penetration rate in the short drama industry has reached about 95%, and the Seedance 2.5 model alone brings Volcano Engine more than 1 billion US dollars in monthly revenue. Volcano Engine then raised its 2026 MaaS revenue target to 15 billion yuan, while the actual revenue of this segment in 2025 was about 1.5 billion yuan. 

Part of Keling's high growth also comes from the B-end. In the first half of 2026, Keling's revenue exceeded 1.5 billion yuan, surpassing the total revenue of about 1.1 billion yuan in 2025; its annualized recurring revenue (ARR) in March was close to 500 million US dollars. A data obtained by National Business Daily shows that in the first quarter of this year, Keling's API revenue proportion exceeded subscriptions for the first time, reaching 60%, and overseas revenue accounted for nearly 80%. 

While revenue is growing rapidly, costs are also rising. The cost of AI video needs to be viewed in two separate accounts. 

The first is fixed investment, including model training, team building and data center construction. Kuaishou expects its capital expenditure in 2026 to be about 260 billion yuan, of which the newly added 110 billion yuan is mainly invested in the computing power of Keling and other basic large models, including server procurement and data center construction. This amount is Kuaishou's total capital expenditure, which cannot be directly equated with Keling's investment, but it is certain that Keling is still in a period of heavy investment. 

The second is the inference cost that has to be paid for every generation. The cost side is more difficult to calculate. An investor focusing on AI film and television said that every call by users is accompanied by actual computing power consumption. The higher the resolution and the longer the duration, the higher the cost. For companies providing AI video services, this is a continuous expenditure that will grow with the increase of users. 

Creators have a more obvious perception of this point. The person in charge of an AI short drama production company publicly calculated an account: for a 720p video, the model charges 1.5-1.6 yuan per second, and generating a one-minute video costs 90 yuan. Calculated based on three times of "random generation" for the same shot, the computing power cost per minute is nearly 300 yuan. 

Xiao Qi, an AI film and television producer, said that many AI short films screened at film festivals are basically modified dozens of times, and the cost of high-quality short films can reach thousands of yuan per minute. But she also said that for many super creation teams cooperating with platforms, the computing power is provided by the platform. This money has not disappeared, it has just been moved from the creator's account to the platform's account. 

This is also the biggest difference between AI video and SaaS. The more SaaS users there are, the lower the marginal cost; the larger the call volume of AI video, the higher the computing power expenditure. What really determines whether this business is viable is the second account: if a single call itself is profitable, the loss mainly comes from the upfront investment, which can be diluted when the scale rises; if every single call is losing money, the more you use it, the more you lose. 

Sora is the extreme version of the latter case. It had high popularity in the early stage of its launch, but the popularity declined later. According to foreign media reports, Sora lost about 1 million US dollars per day, and also occupied a large amount of AI chip computing power. At the node where OpenAI is preparing for IPO and focusing on business contraction, Sora became the first product to be cut off. 

Other players also have the problems that Sora has. From public information, no AI video model company has publicly achieved profitability. 

Keling is one of the samples with the largest revenue scale at present, but according to media reports, Keling is expected to achieve break-even in a single quarter in the fourth quarter of 2028. 

Since everyone is losing money, why haven't Google, ByteDance and Kuaishou retreated, but cut prices to seize users instead? 

02. Major manufacturers increase investment: the money lost in video can be earned back in the ecosystem

In this round of price cuts, Google's moves are the clearest and its pricing is the most transparent. Its approach is to exchange price cuts for usage frequency. Gemini Omni 1.2 Flash is priced at $0.07 per second for 720p, and the synthesis cost of a 15-second clip is about 7 RMB. Xiao Qi said that this price is so low that a single generation no longer constitutes a threshold. 

After the threshold is lowered, what Google sells is "editing". The newly launched conversational editing supports users to repeatedly modify the same video without restarting from scratch every time, but each modification needs to go through an API call and be billed once. 

Some customers have already given usage feedback. Shivam Rastogi, Vice President of Engineering at invideo, admitted frankly after testing that if large-scale and fast video production is required, this is one of the best models they have tested. Matt Rouif, CEO of Photoroom, calculated the account more specifically: they chose this model to make product video advertisements, favoring the "balance of quality, speed and cost", and the native 15-second duration exactly matches the most common needs of customers. 

ByteDance maintains the price for B-end APIs, and uses C-end subsidies to seize creators. Seedance 2.5's API is billed by Token, with 720p at about 1.51 yuan per second, and the unit price of Token is about 1.5 times that of version 2.0. The cheaper part is the C-end channel. During the Double 11 period, the publicity floor price of Seedance 2.5 on the Jimeng member channel was about 0.16-0.19 yuan per second, of which 0.16 yuan per second was the extreme converted price after superimposing membership, bonus points and consumption discounts. 

In addition, open source models such as MiniMax H3 leave room for low-cost deployment on third-party platforms. 

The confidence behind this round of price cuts comes from two things. 

On the one hand, the capability gap of leading video models is narrowing. The Arena video list uses Elo blind test scoring (thousand-point scale), and the first place is only 65 points ahead of the seventh place. In the Elo system, this gap is far from the threshold of "stable advantage". The gap between domestic Wan 3.0, Seedance 2.5 and MiniMax H3 is even more concentrated within 20 points, which is almost statistically tied. 

Many creators said that both Keling and Seedance can generate the same type of shots, and the choice when switching mainly depends on price, speed and ease of access. Some people use Keling to generate shots and switch to Seedance for storyboarding; some others used Keling as their main tool last year, and transferred some links to Seedance this year. 

On the other hand, there is the ecosystem of major manufacturers. Although it is difficult to break even the cost of AI video as a single product, large manufacturers can share the cost through cloud, advertising, subscription and distribution. As long as the video capability can be connected to more businesses, there is a chance to generate revenue there. 

Focusing on video, Google has launched at least six tools including models: on the video model side, there are Gemini Omni and Veo. Veo is Google's video generation model line, and Gemini Omni is positioned as a unified model for creating and editing videos from any modality, both focusing on video creation; on the product entry side, there is Flow for film and television creators, Vids embedded in Workspace for office services, Vertex for enterprise cloud calls, and Ads for advertising delivery. 

Therefore, its video generation charging methods cover four types: API billing by second, Vertex cloud contracts, Gemini subscriptions and advertising delivery. YouTube Shorts has also accessed Veo, allowing creators to directly generate short video clips with text. 

Google can make up for the loss caused by video price cuts from the increment of cloud business. In the third quarter of 2026, Google Cloud's revenue increased by 103% year-on-year to 30.7 billion US dollars, and its operating margin increased from 23.7% a year ago to 36.2%. The management specifically mentioned that TPU system sales are one of the important contributions. Self-developed TPU also means that Google designs everything from chips, servers to software scheduling on its own, so its computing power cost is lower than that of rivals relying on external chips. 

Looking at ByteDance and Kuaishou in China, AI video is not an independent business either. 

ByteDance connects the video model to its own content production chain: Tomato Novel provides IP, Doubao and Xiaoyunque process scripts and storyboards, Jimeng and Seedance generate images, CapCut completes post-production, and then distributes through channels such as Hongguo Short Drama and Douyin, and finally realizes commercial monetization through Ocean Engine and Douyin E-commerce. 

The short drama track is the first to be fully verified. Seedance provides generation capabilities at the most upstream of this chain, and at the same time sells APIs externally through Volcano Engine. Moreover, the industry consensus is that traffic delivery in AI short drama projects usually accounts for 70% of the total revenue. AI reduces production costs, but most of the saved money is paid back to traffic delivery, and the traffic delivery platform is also owned by ByteDance itself. 

Xiao Qi admitted that although multi-person multi-scenario scheduling and complex long shots still require human intervention, problems such as jumping of character positions and mismatched costumes still exist in dozens of episodes of continuous short dramas. But from the perspective of business closed loop, ByteDance has the most complete system among the three players. 

Keling's path also has two aspects. 

Internally, Keling still supports multiple business lines of Kuaishou, including advertising materials for Magnetic Engine, storyboard preview for Xingmang Short Drama, product short videos for e-commerce merchants, and virtual scenes for film and television dramas. In the second quarter of 2026, the consumption of Kuaishou's AIGC short video marketing materials increased by more than 70% year-on-year, and more than 850,000 merchants used the platform's free AI operation tools in the first half of the year. 

In terms of external opening methods, it is significantly different from ByteDance. Seedance connects APIs to third-party platforms such as libtv and Wondershare Drama Lab through Volcano Engine. Xiao Qi describes it as more like a supplier, while Keling implements independent pricing and independent incentives, with its external opening concentrated in the professional creation tool chain. 

The spin-off and financing is a longer-term arrangement. Keling's nearly 3 billion US dollars financing has introduced investors such as Tencent, Alibaba Cloud and Baidu, giving it more sufficient capital reserves. 

AI video is a business that requires continuous investment, but all three companies can find outlets in their respective ecosystems. OpenAI, which chose to shut down Sora, is exactly lacking this point. If model capabilities continue to converge, the party that only sells models will most likely end up becoming a supplier of traffic platforms. 

*The cover image and the pictures in the text are all from Unsplash. 

This article is from the WeChat official account "AIX Finance", author: Lu Wang, published with authorization from 36Kr.