Why Haven't AGI Concept Stocks Surpassed ChatGPT Despite Sora's Popularity?

[baoshi.rao]

On February 16, OpenAI released its first text-to-video model, Sora, once again igniting the AI industry. Sora can produce videos up to one minute long, not only accurately presenting details but also generating characters with rich emotions, marking a significant milestone for the AIGC industry. Elon Musk responded on social media with "gg humans," implying that AI is so powerful that humans must concede defeat.

The popularity of Sora has also shifted attention back to the domestic scene, prompting discussions about who could create a "Chinese version of Sora." This has led to a surge in Sora-related concept stocks, with many seeing a 20% limit-up in trading. The question has evolved from "Why wasn't ChatGPT born in China?" to "Why wasn't Sora born in China?"

Comparing Sora with similar products like Runway and Pika, the most noticeable differences are twofold: First is duration – expanding from typically 3-15 seconds to a full minute. This one-minute duration already meets the basic requirements for short videos and advertising, marking the beginning of AI automation in the film industry and heralding the commercialization and mass adoption of large AI models.

Second is visual quality. Sora generates videos with far higher fidelity than similar products, achieving realism in depicting human skin, facial expressions, animal fur, and demonstrating exquisite handling of camera work and scenes. Most importantly, it evolves text-to-video generation from single shots to producing multiple stylistically consistent shots.

According to OpenAI's technical report, Sora adopts a "Diffusion+Transformer" architecture for its video generation model. As analyzed by ICBC Credit Suisse, compared to the U-Net diffusion model previously used, the Transformer architecture offers stronger parameter scalability – performance accelerates with increased parameters while supporting training data of arbitrary resolution, aspect ratio, and duration without quality degradation from compression. This represents a milestone advancement. Sora can sometimes simulate actions that affect the state of the world in simple ways. For example, a painter can leave new brushstrokes on a canvas that persist over time, or a person can eat a hamburger and leave bite marks. In other words, Sora can "simulate the world," becoming a high-capacity simulator for developing both physical and digital worlds, including the objects, animals, and people within them.

However, as a simulator, Sora currently has many limitations. For instance, it cannot accurately simulate many basic physical interactions, such as glass shattering. This "shortcoming" also provides some relief to those anxious about "AI replacing humans," allowing them to face technological progress more calmly.

According to Morgan Business Research, in China, neither the suddenly popular Sora concept stocks nor the "experts" making quick money by selling Sora courses—as the first beneficiaries of the Sora concept—are likely to achieve significant success through these means. This isn't meant to pour cold water on technological progress. In the torrent of tech development, whenever a new trend emerges, everyone talks about the future or profits from information asymmetry. However, from the perspective of developmental patterns, looking at the past provides much clearer insights than predicting the future. The explosive popularity of ChatGPT a year ago serves as a prime example. From the sudden emergence of concept stocks to the "hundred-model war," after a year of "separating the wheat from the chaff," the trend in the secondary market is far from optimistic.

Take Danghong Technology, which rode the waves of both ChatGPT and Sora, as an example. When the Sora frenzy hit, on February 8, Danghong Technology's stock price rebounded from its bottom, surging by 13.49% that day. Subsequently, the company experienced two consecutive 20% limit-up gains. By the close on February 21, Danghong Technology's stock had risen over 70% in just four trading days. This price movement was almost identical to what happened when the ChatGPT trend emerged last year.

On January 31 last year, Danghong Technology stated on an interactive platform: "The company is actively promoting AI technology to empower various industries, including AIGC, NLP, and other technologies and products related to ChatGPT, which have been gradually refined and implemented in some businesses." Riding the wave of ChatGPT, its stock price soared from 27.87 yuan to a peak of 69.99 yuan by March 28. As the hype around ChatGPT began to fade and leading players successively launched their large model products, Danghong Technology's stock price entered a downward spiral, even hitting a low of 16 yuan on February 6 this year—a 77.14% drop from last year's peak.

The rollercoaster-like trajectory of its stock price is evident in Danghong Technology's financial reports. In 2021, the company's net profit was 61.3475 million yuan, a sharp year-on-year decline of 40.34%. By 2022, its net profit had swung to a loss of 98.5883 million yuan, plummeting 260.7% year-on-year. As of the third quarter of 2023, Danghong Technology reported a net loss of 96.1575 million yuan, down 132.01% year-on-year, with little hope of reversing its unprofitable trend.

Danghong Technology is just one example of many 'concept stocks' stuck in the R&D phase with immature technologies: they are related to cutting-edge fields but not leading in them, their product commercialization remains lackluster, and their finances are either unbalanced or persistently in the red. This suggests that 'concept stocks' are far from synonymous with 'high-quality stocks,' and suspicions of 'riding the hype' or 'pump-and-dump schemes' outweigh their actual capabilities. As one of the leading AI companies, iFLYTEK represents a typical case of a 'concept stock'. In terms of products, the Spark cognitive model was among China's earliest ChatGPT-like releases, undergoing continuous iterations with commercial integration into educational tablets and business laptops. According to iFLYTEK's 2023 financial forecast, its net profit has shown steady growth.

However, the secondary market hasn't witnessed the expected sustainable growth for iFLYTEK. Instead, each technological upgrade has triggered a 'darkest hour' in stock performance:

• On August 15, 2023, when iFLYTEK launched the Spark V2.0 upgrade, its stock price dropped 8.62% the next day.
• On October 24, 2023, with the release of Spark V3.0 claiming overall superiority over ChatGPT and medical domain advantages over GPT-4, the stock closed down 9.97%.
• On January 30, 2024, during the Spark V3.5 upgrade announcement by Chairman Liu Qingfeng, the stock fell 6.66% the following trading day. This isn't hard to understand. For tech companies, the easiest story to tell is before a product's actual launch. iFlytek, for instance, has consistently promoted the idea of 'surpassing' in its marketing. However, with ChatGPT setting a high benchmark, every release and iteration of domestic large models provides users with a better opportunity to compare.

Repeatedly falling short in user experience makes it increasingly difficult to convince users of the 'surpassing' vision. Technical iterations and innovations that bring no surprises only lead to disappointment and disillusionment. How, then, can they maintain a moat in the secondary market?

What people anticipate are sudden 'surprises,' not long-planned 'responses.' Are you disappointed with domestic enterprises in related fields? To say no would be insincere. However, the disappointment does not stem from "backward" technology, but rather from the actions of many domestic companies—making empty promises, riding hype waves, manipulating capital, and exploiting retail investors.

Technologically speaking, whether it's disruptive large-scale model products like ChatGPT and Sora, or the design and production technologies of various high-end chips, the delays due to late starts and the "no ingredients for a skilled cook" scenario caused by product and technology blockades are understandable. As many joke, many domestic companies developing large models can't even afford their electricity bills.

Taking computing power chips as an example, according to estimates by third-party data agency SemiAnalysis, OpenAI used approximately 3,617 HGX A100 servers, containing nearly 30,000 NVIDIA GPUs. Moreover, following Moore's Law, with technological and product iterations, OpenAI's demand for computing power will multiply in the future. This will be a terrifying financial requirement, one that is difficult for small and medium-sized enterprises to sustain. For most domestic enterprises, a better path would be to build applications based on existing open-source large models or their ecosystems. However, possibly influenced by capital pressures, many have instead embarked on the path of surpassing OpenAI.

It's important to note that beyond GPUs, investor Microsoft helped OpenAI establish customized computing clusters for large models, which further enhanced the efficiency of these GPUs. Even with such capabilities and investment, OpenAI still took over eight years to develop the groundbreaking product GPT-4, which still has imperfections.

This demonstrates that the path of large model development will be extremely long and difficult. Even for well-funded companies like Tencent, Baidu, Huawei, and Alibaba, they are affected by the US 1017 new regulations, which restrict the export of GPU chips including A100, H100, A800, and H800 to China, creating a bottleneck. In fact, whether it's OpenAI, Microsoft, Google, or the numerous companies in China's "Hundred Models War," due to difficulties in product implementation, commercialization challenges, and the lack and lag of ethical, legal, and regulatory frameworks, it's difficult for them to monetize their technologies and products in the short term.

The "paradigm revolution" brought by ChatGPT and Sora has not only fueled speculative capital but also made Nvidia, the seller of computing chips, the real beneficiary.

Financial reports show that Nvidia's revenue for the fourth fiscal quarter was $22.1 billion, a 265% year-over-year increase; net profit was $12.3 billion, a staggering 769% year-over-year surge; diluted earnings per share were $4.93, up 765% year-over-year. In the secondary market, NVIDIA stood out with a staggering 239.02% annual gain in 2023, followed by another 46.63% surge in 2024 (as of February 16). NVIDIA's stock price has hit a new all-time high, pushing its market capitalization to $1.8 trillion, surpassing Google and Amazon to become the world's fourth most valuable company. This explains why major tech giants, beyond investing in large models, are now turning their attention to chips.

In early February, a Meta spokesperson revealed that its second-generation in-house AI chip, Artemis, will officially enter production this year, initially for data center use. However, the production capacity and computing power of self-developed chips cannot meet demand in the short term. Meta plans to deploy 350,000 H100 GPUs by year-end, bringing its total GPU count to over 600,000.

OpenAI CEO Sam Altman aims to raise $7 trillion to enter the chip industry, seeking self-sufficiency by controlling the upstream supply chain. Meanwhile, SoftBank Group founder Masayoshi Son, after a long hiatus, is seeking $100 billion (approximately ¥7 trillion) to establish an AI chip company to challenge NVIDIA's dominance. In other words, this AI battle isn't simply about the metaverse, virtual beings, or large models. It's about using disruptive technologies to concentrate attention, capital, and resources, thereby reshuffling and redefining global technological dominance. For domestic enterprises in China, this remains a long and arduous journey.

Of course, we all believe that Chinese tech companies will eventually break through the 'bottleneck' constraints, and China's AI technologies and products will stand at the forefront of the world. However, before any milestone 'surprises' materialize, all claims of 'intention to surpass' or 'aspiration to become' can only be defined as cowardly acts of 'exploiting one's own people'.

In the AI era, talent, data, and chips form the foundation for training and development. As humanity transitions from 'carbon-based' to 'silicon-based', we need to pursue technological limits with down-to-earth determination. Disclaimer: This article is based on the company's legally disclosed content and publicly available information for commentary, but the author does not guarantee the completeness or timeliness of such information.

Additionally: The stock market carries risks, and caution is advised when entering. This article does not constitute investment advice, and investment decisions should be made at your own discretion.