Why Seedance beat Sora in the race for AI video generation

On 24 March 2026, OpenAI officially announced the shutdown of its AI video generation project Sora. This meant all technical iterations and operational maintenance were terminated for its three core components: the standalone consumer app, the developer API interface and the video generation feature built into ChatGPT.

On the same day, OpenAI also announced the termination of its partnership agreement with The Walt Disney Company, which Disney confirmed in a statement. In December 2025, the two parties announced with much fanfare a three-year collaboration involving a planned US$1 billion investment and IP licensing by Disney. However, the agreement was terminated before the funds were even transferred, following Sora’s shutdown. 

Thus was Sora, the video generation model that had astonished the world in February 2024, formally consigned to history.

The dominant position in global video generation models has now been firmly taken over by the once low-profile underdog, Seedance, developed by ByteDance. In the first quarter of 2026, Seedance ranked first globally across the three most critical metrics: overall technical score, total token usage, and estimated aggregate revenue.

The demise of Sora had long been foreshadowed. Despite its once dazzling “cinema-grade generation capabilities”, the model was unable to escape the dilemma of high costs and low returns. 

Industry sources revealed that generating a single one-minute high-definition video cost as much as US$300, with daily computing costs of over US$1 million. Meanwhile, its commercial deployment remained largely stagnant. Since Sora’s launch, total revenue had amounted to less than US$10 million — far from sufficient to cover its enormous R&D and operational expenses — ultimately forcing its withdrawal amid sustained losses.

In stark contrast, the explosive growth of Seedance 2.0 has been remarkable. As a multimodal video generation model independently developed by ByteDance, it was rapidly integrated into products such as AI assistant Doubao and generative AI app Jimeng, as well as the Volcano Engine cloud service platform.

During internal testing alone, daily active users exceeded ten million. After submitting generation requests, users faced queue times of up to eight hours, with daily video output reaching between five and eight million clips. 

Its commercialisation has advanced at breakneck speed: within just three months of launch, Seedance’s revenue surpassed US$300 million. Paying users exhibit high frequency, high spending and strong retention, with professional sectors such as short drama production and computer animated films contributing over 75% of total revenue. Both user growth and revenue scale have shown exponential expansion, making it a phenomenon in the global AI video generation landscape.

The meteoric rise of Seedance and the exit of Sora are by no means coincidental. They reflect a fundamental divergence in the technological development paths of China and the US in AI video generation, a theme that I have explored throughout this series.

From its inception, China’s video generation models have been deeply embedded within the mature and profitable ecosystem of short drama production, following a closed-loop path of technological deployment, scenario feedback and iterative optimisation.

By contrast, the US has fallen into a trap of style over application, driven by financing and technological showmanship. China’s short drama industry has already formed a vast and mature ecosystem. According to industry statistics, the market size exceeded 80 billion RMB (US$11.7 billion) in 2025, of which AI-generated short dramas accounted for 32 billion RMB — 40% of the total — and continues to grow at a monthly rate of 15%.

In comparison, the US short drama market lags significantly behind, with a 2025 market size of only around 12 billion RMB — less than 15% of China’s — and largely characterised by niche productions lacking large-scale, industrialised systems.

In terms of distribution channels, Chinese mobile short drama platforms dominate 68% of the global market, supported by a massive user base and well-developed distribution systems. US platforms, by contrast, account for only 12%, highlighting a stark disparity. This difference in industrial foundations has determined the divergent fates of Seedance and Sora.

Seedance’s development has consistently revolved around the needs of industrialised short-drama production. Its industrial-grade capabilities, such as automatic storyboard planning and cross-shot character consistency control, have reduced production cycles from 21 days to just three days, while significantly lowering costs. This has created a virtuous cycle of “technology empowering industry, and industry feeding back into technology”, demonstrating strong practical value.

Meanwhile, Sora became overly absorbed in technological demonstrations and capital hype, lacking the support of mature industrial scenarios. It ultimately devolved into a “laboratory toy”, unable to translate into real commercial value. This reflects the core logic of China’s AI industry: cost efficiency and concrete applications — and not fundraising — are the primary objectives.

Seedance’s overtaking in the AI video generation field is not an isolated case, but rather an inevitable outcome of China’s “industry-driven, practical application-oriented” model of AI development. This model is rapidly expanding into other sectors and will drive China to catch up with — and even surpass — the US in more domains.

Looking at the broader landscape of China-US AI competition, China has already surpassed the US in global large-model usage over the past three years. The rise of Seedance marks the extension of this advantage from general-purpose models into vertical application domains.

China’s AI industry has consistently adhered to a pragmatic approach, closely aligned with industrial needs, and leveraging its vast manufacturing and internet service base to tightly integrate technological breakthroughs with commercial applications. This “cost-performance plus scenario adaptability” model is far more resilient than the US approach, which emphasises heavy R&D, slower iteration and reliance on capital-driven spectacle.

As this development model continues to deepen, China’s AI sector is poised for explosive growth across more vertical industries. The competition between China and the US in AI has now entered a turning point, with the balance of initiative shifting. I would boldly predict that the next domain in which China may overtake the US will be AI programming or intelligent agent systems.

At present, models such as GLM-5 and DeepSeek V3.2 already rival US counterparts in coding efficiency and accuracy. In the field of intelligent agents, Chinese companies are accelerating iteration by leveraging rich application scenarios. Compared with the US approach — which emphasises compliance over deployment — China’s agents prioritise deep integration with industrial contexts. Given time, a historic overtaking in this field appears inevitable, potentially ushering in an era of Chinese leadership in AI.

Related: When cost and practical application takes priority: China surpasses US in AI adoption | Who steers AI: China’s industrial state vs America’s frontier builders?