In Depth: China’s Dwindling AI Chip Stockpile Leaves Little Room for Sora Copycats

The launch of Sora, a program developed by OpenAI to convert text prompts into computer-generated videos, three months ago has pressured Chinese firms in the same field. However, the high costs, significant computational resources required, and limited availability of advanced hardware in China have dampened enthusiasm for pursuing a generative video arms race [para. 1][para. 2]. Industry insiders doubt that Chinese startups will replicate Sora due to easier technical paths but challenging computing power requirements [para. 6].

Despite this, two Beijing-based generative video startups, Shengshu Technology and AIsphere, announced substantial funding rounds in March, raising hundreds of millions of yuan in venture capital and around 100 million yuan ($14 million), respectively [para. 3]. Nonetheless, the overall interest in generative video investment in China remains much cooler compared to the excitement over large language models (LLMs) triggered by OpenAI’s ChatGPT [para. 1][para. 4].

The effectiveness of American tech sanctions, which restrict China’s access to advanced semiconductors essential for next-generation computing, is evident. These sanctions force Chinese firms to focus on acquiring advanced hardware through alternative means [para. 5]. As a result, major companies are hesitant to channel resources into replicating Sora unless a clear business case emerges [para. 2]. High-profile companies like Baidu have also exhibited caution regarding entering the generative video domain despite significant AI investments [para. 7].

Sora's development underscores the substantial computational resources required to support the technology. Its capabilities combine the language processing of Transformers, used in LLMs like GPT, with the high-resolution visual generation of the Diffusion process, as seen in image generation models like OpenAI’s DALL-E and Stable Diffusion [para. 10]. Sora can create detailed, extended video scenes, a significant leap beyond previous attempts [para. 11]. However, issues like the lack of sound, bizarre anatomical features, and legal challenges related to content sourcing remain prevalent [para. 12]. The computational demands are immense, with estimates suggesting that replicating Sora's current level requires power equivalent to thousands of Nvidia’s top GPUs, with costs potentially hitting tens of millions of dollars [para. 13][para. 14].

Challenges posed by the high financial and computational costs have led to a reality check among Chinese tech giants. ByteDance, with substantial data resources and financial strength, only consolidated its efforts in text-to-video AI in the latter half of the previous year, recently launching a tool named Doubao [para. 16]. However, U.S. restrictions on AI chip sales to China have deepened the challenges, causing scarcity and driving up prices of high-performance computing resources [para. 20]. Regulatory changes in China have also imposed new requirements on generative AI firms, influencing their strategic decisions [para. 22].

Concerns around the computing power requirements make companies like ByteDance cautious, focusing on feasibility aspects such as model compression for mobile devices and meeting regulatory demands [para. 23]. Firms like Alibaba and Tencent have launched some generative video tools, but none have developed capabilities matching Sora's level [para. 25][para. 26][para. 27].

The release of Sora, unlike ChatGPT, did not stir significant investor enthusiasm due to the financial and resource-intensive nature of large models [para. 31]. However, startups like AIsphere and Shengshu Technology received considerable funding and are actively pursuing advancements in AI video models [para. 32][para. 35]. These startups face challenges distinct from those encountered with LLMs, primarily due to substantial computational power requirements and reliance on limited open-source models [para. 37].

Despite the optimistic prospects fueled by recent funding, the investment climate in China remains cautious, with many investors seeking quick returns, contrasting with the longer-term approach prevalent in markets like the U.S. [para. 38].

In conclusion, China's generative video sector grapples with significant hurdles, including high computational costs, regulatory constraints, and limited hardware access, which temper the enthusiasm for scaling efforts similar to OpenAI’s Sora [para. 5][para. 18][para. 22].