Large Models Still Can't Save Struggling AI Companies
-
Last year, ChatGPT sparked a wave of global internet technology innovation, making large models the undisputed trend. This injected new vitality into the somewhat stagnant AI industry in recent years, bringing AI-related companies back to the center stage of the internet.
Entrepreneurial trends around large models have also surged, with many startups launching large model-related products, and a continuous stream of small AI products focusing on niche market applications.
But a year has passed—what has this hot generative AI technology boom actually brought to China's AI industry? What breakthroughs have AI companies achieved with the assistance of large models? Has it provided new developments in technological R&D and commercialization exploration? If we ponder these questions, we might arrive at a pessimistic answer. At a closed-door meeting on generative AI earlier this year, a former R&D lead from one of China's 'AI Four Dragons' remarked, '2023 hasn't brought any real change—the problems AI companies faced before still remain unsolved today.'
2023 has truly earned its reputation as a 'capital winter.' According to PitchBook data, U.S. venture capital firms raised $67 billion in 2023, marking a 60% year-over-year decline and hitting a six-year low. China's market mirrored this trend: in the first half of 2023, the venture capital sector reported 4,367 financing deals (down 38.73% YoY) totaling 301 billion RMB (down 24.92% YoY).
Amid these conditions, the ChatGPT frenzy and the technological wave around large models have concentrated a significant portion of funding within the AI sector. According to the latest data from The Information's Creator Economy Database, among the more than 350 global startups covered in the database, the total funding raised in 2023 continued its downward spiral, reaching approximately $1.7 billion. AI startups accounted for the largest share of funding, with over $324 million.
The resurgence of funding in the AI sector can be seen as a timely relief for China's AI industry and related companies. Since 2019, the number of investment and financing events in China's AI sector has shown a clear downward trend. While there was a brief surge in 2021, 2022 saw a sharp decline, almost halving the previous year's figures. This has left all AI companies, still in the stage of heavy capital expenditure, facing survival challenges—even the leading ones are no exception.
For example, Mobvoi completed seven rounds of financing between February 2013 and September 2019 but has not secured any new funding in the past four years. 4Paradigm underwent eleven rounds of financing from August 2015 to June 2021 but has seen no new funding activity in the past two years. Zhipu AI, one of the top-tier companies in the large model startup sector, completed a significant funding round of 2.5 billion yuan last year. However, prior to that, the company also experienced a prolonged period of funding drought. In the past, investors paid little attention to the field of large models. However, after the emergence of ChatGPT, the financing situation for startups related to large models has noticeably improved. Even AI companies that merely ride the wave of the large model trend have seen their valuations soar, reviving new narratives. Yet, this fervor seems somewhat fleeting.
According to data from the China Commercial Industry Research Institute, as of November 24, 2023, there were 531 investment events in China's artificial intelligence sector, totaling 66.048 billion yuan. Among these, May, June, and July saw the highest number of investment events, with 62, 60, and 60 cases, respectively. However, by November, the number of investment events had decreased significantly, with only 26 cases recorded by the 24th. These included 5 strategic investments, 4 Series B rounds, and 4 Pre-A rounds.
Looking at the investment and financing events in the third quarter, the hot sectors were concentrated in five areas: new energy, semiconductors, healthcare, enterprise services, and artificial intelligence, with 97, 154, 278, 155, and 117 cases, respectively. Except for a slight increase in the semiconductor sector, the number of investment and financing events in other sectors declined, aligning with the overall downward trend in the primary market. An angel investor who evaluated multiple projects but ultimately declined to invest stated, "Basically none had investment value. About 90% of projects in the market are open-source models, but they lack mature ecosystems."
With a watch-and-wait approach or even pure observation without investment, the capital "winter" remains bitterly cold.
In recent years, the common dilemma for AI companies has been that technology R&D resembles a bottomless pit while commercial returns are meager. Worse still, beyond profitability challenges, many firms can't even identify viable commercialization pathways. While the explosion of large model concepts has brought financing to the AI sector, these generative AI startups face the same fundamental question as earlier AI enterprises: Where are the business models? In the second half of 2023, OpenAI's commercialization process has noticeably accelerated, first by introducing a paid version for individual users, followed by customized solutions for enterprise clients. Throughout this process, the company has continuously reduced operational costs for its free-tier services while enhancing the capabilities of paid versions to improve conversion rates.
However, seeking commercial monetization from individual consumer demand is not an easy path in China.
For users, while ChatGPT's explosive popularity has refreshed their understanding of AI technology and capabilities, large language models and generative AI remain relatively vague or even distant concepts. This is primarily because "content generation" is not a universal or core need for most people. Therefore, attempting to generate revenue through user subscriptions inherently faces the challenge of whether users are willing to pay. Of course, embedding large models or generative AI technologies in high-demand scenarios to create a breakthrough application-layer product could be a better choice. Unfortunately, by the end of the year, China still hadn't produced a consistently popular AI application. Even the once-trending Miaoya Camera gradually faded from public view just two months after its launch, due to issues like repeated payments and insufficient user retention.
Many industry insiders believe the broader application market for large models lies in the B2B sector. Most AI companies before the large model wave also found their business models in the B2B space. Now, does the integration of large models with their current operations bring new possibilities?
It must be said that most AI companies' revenue situations remain dismal. For example, SenseTime Group reported total revenue of 1.433 billion yuan in the first half of 2023, a year-on-year increase of 1.3%, with gross profit of 649 million yuan, down 30.6% year-on-year, and a net loss of 3.143 billion yuan. Cambricon's revenue for the first three quarters of 2023 was 146 million yuan, down 44.84% year-on-year, with a net loss attributable to shareholders of 808 million yuan. CloudWalk's revenue for the first three quarters was 346 million yuan, down 24.13% year-on-year, with a net loss of 401 million yuan. Investment and research and development in the fields of large models and AIGC represent a long-term commitment. This means that products and services related to AIGC and artificial intelligence have not yet generated actual revenue and are unlikely to make a significant contribution to a company's short-term financial performance. However, the fundamental issue lies in the fact that large models still cannot provide a clear path for AI companies to overcome their commercialization challenges in the B2B sector.
On one hand, this is because large models and AIGC technologies inherently face difficulties in commercial implementation.
As an AI company that developed large models, AIGC technologies, and applications relatively early, Mobvoi serves as a typical example. According to a report by CIC, Mobvoi ranked first among Chinese AI technology companies in terms of revenue from AIGC products and services in 2022 and launched the first commercial AIGC application in China. However, in terms of revenue structure, the proportion of income from AIGC solutions at Mobvoi was only 0.2%, 1.7%, and 8% from 2020 to 2022, respectively. On the other hand, in addressing and solving the diverse needs of enterprises, most AI companies' current technology-based services have not brought qualitative upgrades. This is why their B2B commercial prospects are increasingly being questioned. The emergence of large models might be a new boost, but their actual utility is doubtful.
"If large models cannot be used flexibly or fully adapted to one's business scenarios, it may be difficult to completely achieve the goal of reducing costs and increasing efficiency to some extent," professionals said.
Before large models can drive transformation, huge bubbles have already emerged. In the first half of 2023, the AI sector saw numerous "bull stocks" emerge. Among them, Wanxing Technology was crowned the "most popular," with its stock price more than tripling for the year, reaching a peak increase of 380%. Kunlun Tech was not far behind, with its stock price doubling and peaking at over 340%. Companies like Yinsai Group, Kaipuyun, and Xinguodu also saw annual increases exceeding 100%. Many of these AI concept stocks, however, have not made significant technological breakthroughs, and some are even operating at a loss, yet their stock prices have surged ahead.
As Gary Marcus quipped, "A few years ago, if your startup had '.ai' in its domain name, you could add a zero to your valuation. Now, you might add two zeros, especially if you claim to be working on generative AI."
But after elevating these AI companies or large model startups to such heights, once the hype subsides, will those AI firms engaged in costly large model investments fall even harder? On one hand, there are continuously rising operational costs. It's reported that last year SenseTime invested 10,000 GPUs for large model development and increased its SenseCore AI infrastructure to 30,000 GPUs. The increasing hardware costs and depreciation have severely impacted SenseTime's profits. Leading startups like Baichuan Intelligence, Minimax, and Zhipu AI are accelerating fundraising precisely to cope with these massive future investments.
Zhipu AI CEO Zhang Peng stated in an interview that 2.5 billion yuan in funding is far from enough, frankly admitting, "No matter how much we raise or earn now, it's just travel money on our road to AGI."
On the other hand, investors are becoming increasingly impatient with their portfolio projects, urgently wanting to see returns or clear commercialization paths. As Hua Capital founding partner Xiong Weiming said, "Building startups just for VC funding doesn't work anymore - you must generate profits. Today's AI startups can't expect multiple funding rounds like A, B, C, D series anymore. Now it's basically just two rounds - angel and Series A. So we've become very pragmatic in our investments, focusing strictly on whether the project can generate profits, and substantial ones at that." Certainly, the longer a company operates in the field of large models, the more capital it requires. With each additional round of financing, investors impose stricter demands on the company's technological capabilities and profitability.
Under such combined internal and external pressures, AI companies are likely to find themselves in a dilemma: focusing on enhancing technological capabilities while neglecting commercial exploration may lead to abandonment by investors. Conversely, prioritizing commercialization before achieving technological maturity could hinder breakthroughs in technology.
At some point in the future, if all the invested funds are exhausted without yielding technological results, capital withdrawal will inevitably expose the company to even greater risks. This test probably won't wait too long, after all, in just half a year, the wind direction has subtly changed.