On October 3, OpenAI officially announced that it had successfully raised $6.6 billion (approximately RMB 46.69 billion) in funding. After the completion of this round of funding, OpenAI's valuation exceeded $150 billion. The new round of funding was substantial, yet it was only enough to sustain OpenAI for a year. The company expects to generate revenues of $3.7 billion this year, but with losses of $5 billion.

Previously, US media outlet The Information reported that, based on financial data analysis, OpenAI is projected to become profitable in 2029 with revenues reaching $100 billion (approximately RMB 707.48 billion). However, significant losses are expected in the interim, with a projected loss of $14 billion (approximately RMB 99.05 billion) in 2026. Therefore, OpenAI's potential to generate annual revenues of $10 billion is predicated on its ability to survive until 2029.

Even a formidable player like OpenAI struggles with financial sustainability, underscoring the intense competition in the AI large model industry. Following the 'Hundred Models War' in 2023, the industry finally recognized the importance of 'application implementation' in 2024. However, the journey from laboratory to real-world applications is fraught with challenges, and only a few companies are likely to endure. How many AI companies will survive? Robin Li's answer is: 1%.

Recently, Robin Li, the founder of Baidu, a leading Chinese AI large model company, once again sounded the alarm. In a conversation with Adi Ignatius, Editor-in-Chief of Harvard Business Review, he bluntly stated that the technology bubble in generative AI is inevitable, but fortunately, the industry is on the right track. Li predicted that during the AI bubble bursting phase, "pseudo-innovations that fail to meet market demands will be weeded out. After that, 1% of companies will emerge as winners, continuing to grow and create significant value for society. We are merely going through this phase, and the industry is calmer and healthier than last year."

(Image source: Harvard Business Review's HBR Live: The Future of Business)

From the Hundred Models War to a Frenzy of Applications: The Inevitable AI Bubble

During the deep learning craze around 2017, there were numerous online discussions about whether AI was a bubble. The debate intensified during the era of large models, with pessimistic netizens outlining a complete bubble bursting process: "Invest - Develop - Harvest - Burst - Settle," even predicting that the generative AI bubble would be much larger than any previous technological bubble.

(Image source: Zhihu)

After the emergence of ChatGPT, the AI industry witnessed a large model boom in 2023. In China alone, hundreds of new players launched foundational large models, competing with established players like Baidu, igniting the "Hundred Models War." However, few of these players possessed genuine capabilities. Many foundational large models were hastily assembled, relying on open-source frameworks for their underlying technology, and their market presence was often inflated through artificial means.

The result? While many foundational large models claimed to surpass GPT4.0, they quickly proved to be inadequate to technical experts. Even one of the "Big Five" unicorns was accused of using an open-source framework. More concerning was that most foundational large models lacked real-world usage; the combined usage of over 200 domestic large models was less than that of a single Wenxin large model.

Duplicating large model development is futile for entrepreneurs and a colossal waste of R&D, talent, and computing resources for society. In 2024, the market finally cooled down, and Robin Li's call for a shift from model competition to application competition was heeded by many AI companies. As a result, the number of foundational large model projects dwindled, while the number of upper-level AI applications increased, even surpassing that of foundational models. However, many AI teams still pursued the wrong direction in AI application development, plagued by what Robin Li termed "pseudo-demands." Not only AI startups but even giants and tech titans made mistakes and detours.

In the first half of 2024, AI hardware emerged as a form of AI large model implementation. At CES (Consumer Electronics Show) and MWC (Mobile World Congress), the most notable new species of AI hardware were AI devices like Rabbit R1 and AI Pin. Many appliance manufacturers incorporated AI large models, even in products like range hoods and gas stoves. Meizu even announced it would abandon its mobile phone business to focus solely on AI hardware. However, AI hardware's fortunes took a sharp turn in the second half of the year, with AI Pin facing massive returns due to poor user experience. By September's IFA (Internationale Funkausstellung Berlin), AI hardware had all but disappeared.

(Representative of new AI hardware species: AI Pin)

It's not just AI hardware that missed the mark. Many AI software products were also criticized for exaggerated marketing claims. Users lured by false advertisements often found the actual experience vastly different from what was promised. Consequently, the US Federal Trade Commission even sued five AI companies for significant discrepancies between their claims and actual AI performance. One of the sued companies was DoNotPay, a robot lawyer service that claimed to "provide services comparable to human lawyers," which turned out to be exaggerated.

By the second half of 2024, it became evident that warnings about the AI large model bubble were increasing, and the industry was cooling down.

Before Robin Li's warning, the industry was witnessing an increasing number of sober voices and actions. For instance, Apple researchers recently published a paper titled "Understanding the Limitations of Mathematical Reasoning in Large Language Models," questioning their mathematical reasoning abilities and even suggesting that they lack genuine reasoning capabilities. Apple, which abandoned car manufacturing to strategically bet on AI, even passed on investing in OpenAI's latest funding round.

Meanwhile, AI large model investments ceased to be frenzied. Some AI startups struggled, and AI talent and technology consolidated towards giants, signaling accelerated industry consolidation. For example, InsurStaq.ai, hailed as a dark horse in the insurance industry, abruptly shut down after just a year of operation. Unicorn companies Adept AI and Inflection were hollowed out by acquisitions from Amazon and Microsoft, respectively. In the pragmatic and application-focused Chinese market, the AI venture capital market underwent intense shuffling. According to TiMedia statistics, from November 30, 2021, to July 29, 2024, nearly 80,000 AI-related companies in China were dissolved, suspended, or revoked.

Bursting the bubble will ultimately accelerate technology democratization

Historically, bubbles have been inevitable in major technological waves.

In his conversation with Harvard Business Review, Robin Li recalled the "bubble phenomenon" in past technological waves:

'Just like many previous technological waves, after the initial excitement phase, bubbles are inevitable. Then, when the technology fails to meet the high expectations set during that excitement, people feel disappointed. We've experienced similar situations many times, such as the dot-com bubble that burst in March 2000, following the rapid growth of the internet in the 1990s. A similar scenario played out during the mobile internet era. In the era of generative AI, we'll go through this process too.'

Regarding the development patterns of technological waves, the renowned consulting firm Gartner outlined the 'Hype Cycle' model as early as 1995. This model suggests that the development of a technology passes through five stages: the Technology Trigger, Peak of Inflated Expectations, Trough of Disillusionment, Slope of Enlightenment, and Plateau of Productivity.

(Image source: Gartner)

When large models first emerged, humanity was as excited as our ancestors discovering fire, vastly expanding the imagination of what AI could achieve. However, excessive media coverage, often sensationalized with terms like "explosive," "nuclear," and "disruptive," propelled AI to the forefront. Capital rushed in, further crowding the AI landscape with a mix of genuine players like OpenAI, Baidu, Alibaba, and iFLYTEK, alongside opportunistic speculators. Amid the chaotic competition, from the "Hundred Models War" to the "frenzy of implementation," AI's shortcomings, problems, and limitations gradually surfaced. Industry failures outnumbered successes, and the bubble gradually materialized.

Considering Robin Li's warning and Gartner's 'Hype Cycle' model, today's AI technology may be on the eve of the 'Trough of Disillusionment.'

In this new phase, more and more AI players will be weeded out. Those without core technological capabilities will quickly be exposed and eliminated. Even players with technological prowess cannot rest easy, as productization, commercialization, and positive business cycles are crucial for long-term success. This underscores the importance of Product-Market Fit (PMF), emphasized by Robin Li in his conversation with the Editor-in-Chief of Harvard Business Review.

(Image source: Baidu AI Assistant)

Fortunately, Chinese entrepreneurs, known for their proficiency and emphasis on technological application, pay closer attention to PMF. "China focuses more on application-driven development. We're more concerned with which applications will benefit from large models, and many startups are exploring how to leverage large model capabilities," said one observer.

As the bubble bursts, only genuine players with both technical prowess and the business acumen to bridge the technology-market gap will remain. In Robin Li's prediction, only 1% of AI players will navigate the bubble phase and emerge into the 'Slope of Enlightenment' and 'Plateau of Productivity.'

AI's journey mirrors that of PCs, the internet, mobile internet, short videos, and other technological waves. After the dot-com bubble burst in the early 2000s, internet giants like Baidu, Alibaba, and Tencent emerged as the '1%' that survived and thrived. Post-bubble, these companies prioritized commercialization ('how to make money'), launching products tailored to user needs, attracting more internet users, fueling the consumer internet boom, and reinvesting profits into R&D and service upgrades, forming the tripod of internet commercialization: advertising, value-added services, and e-commerce.

Inevitably, technological waves create bubbles, and their bursting is a crucial step towards maturity. This is the deeper meaning behind Robin Li's statement that "bubbles aren't necessarily bad; they purge pseudo-innovations."

Becoming the 1% to survive and claim victory

Since AI bubbles are inevitable and will eventually burst, causing painful disruptions, how should AI players confront this destiny? The answer is straightforward: enhance 'Product-Market Fit' (PMF).

For instance, duplicating foundational large models is unlikely to find a market. Tech giants are investing heavily in foundational large models, engaging in an arms race that most companies cannot afford, especially in terms of computational power. With Microsoft's support, OpenAI bears the costs of inference, training, and labor, with labor being the lowest at around $1.5 billion annually; inference costs for ChatGPT and underlying LLMs, powered by rented Microsoft servers, are around $4 billion; training costs, including data expenses, are around $3 billion annually and expected to exceed $10 billion in the future. Similarly, Baidu, a leading Chinese AI player, has invested over RMB 100 billion in AI over a decade to achieve its current achievements.

Therefore, startups competing with giants in foundational large models is akin to 'throwing pebbles at a boulder.' If a startup claims its foundational large model surpasses GPT4.0, even the laws of physics might object. The only opportunity for startups in foundational large models lies in becoming 'agents of giants,' akin to OpenAI's reliance on Microsoft.

Regarding AI applications, whether AI software, hardware, or B2B services, AI entrepreneurs must focus on genuine user needs rather than "bringing a hammer to find a nail" or blindly following trends. They must also prioritize profitability, ensuring users/clients are willing to pay or exploring new business models like reliable advertising to achieve positive business cycles and survive.

Not just startups but even giants must prioritize PMF in AI deployments.

In 2024, many smartphone and appliance manufacturers faced anxiety due to shrinking market demand and slumping sales. AI seemed like a lifeline, turning smartphones into AI phones, smart TVs into AI TVs, and PCs into AI PCs. However, when changing titles, did these AI-infused devices genuinely enhance user experience or offer value worth paying for? Many manufacturers might struggle to give a confident answer.

In 2024, embracing AI large models became fashionable across industries worldwide. AI large models' technological potential can indeed benefit various industries, enterprises, products, scenarios, and individuals. However, how many enterprises genuinely leverage AI to enhance PMF or business value? The answer is likely bleak. A fascinating viewpoint shared by the Chief Analyst at Forrest, a market consulting firm, might capture the mindset of many 'AI embracers': "No one wants to attend a costume party in street clothes."

In the future, the true AI era will only officially arrive when AI is no longer a "fad" but is "ubiquitous yet imperceptible" like the internet and 5G networks.

A few days ago, Hinton, a former Google leader and AI expert, won the Nobel Prize in Physics, while two AI researchers from Google DeepMind, Demis Hassabis and John Jumper, shared the Nobel Prize in Chemistry for their work on "protein structure prediction." This underscores the accelerating penetration of AI into the physical world and metaphorically suggests that AI will eventually take over this realm. While the current AI bubble is inevitable, the future of AI remains optimistic.

Source: Lei Technology