Elite's Game in ToG Large Model Market
The landscape of large-scale models in China is witnessing a robust transformation, primarily driven by local state-owned enterprises. Following OpenAI's groundbreaking launch of ChatGPT in late 2022, which served as a beacon for AI advancements globally, Chinese authorities still grapple with a lack of comprehensive policy support specifically tailored for large language models (LLMs) as of 2023. Instead, current directives seem to focus on broader AI growth or computational support.
This scenario, however, is changing rapidly as regional governments unveil a wave of targeted policies aimed at accelerating the development of AI. The enthusiasm among local governments is palpable, spurring state-owned corporations (SOEs) to partner with key players like Baidu, iFlytek, and Huawei to swiftly implement large model projects and foster a competitive public bidding atmosphere.
The result? A competitive market that features established names and rising stars alike. Firms such as Baidu, iFlytek, Huawei, and notable contenders like Zhipu AI, which is heralded as one of the "six dragons" of Chinese large models, are carving out significant roles in this evolving sector.
Despite the impressive technological advancements and capabilities offered by AI large models, the transition from conceptual frameworks to tangible commercial applications remains fraught with challenges. Companies like Baidu and iFlytek, which have aligned themselves with government tenders (ToG), face uncertainty about whether this strategy will yield direct benefits or if they will have to navigate a more circuitous route to profitability.
Advertisement
Dynamics of Competition
The ToG sector offers a substantial market opportunity, yet the dynamics within this arena are complex. As of 2023, there has been a marked increase in procurement demands, with public biddings witnessing a surge. According to available data, approximately 190 large model procurement initiatives were launched throughout the year, amounting to a staggering 595 million yuan; this already skyrocketed to 498 projects by mid-year 2023, totaling 1.34 billion yuan—more than double the previous year’s figures. The month of August alone saw a record 127 projects being awarded, with 84 disclosing transaction values that reached 390 million yuan.
In July, the State Council held a press conference discussing the strategic impetus for high-quality development, announcing that central enterprises are expected to allocate over 3 trillion yuan towards upgrading equipment and enhancing technological integration over the next five years. This initiative aims to accelerate digital transformation and the adoption of innovative technologies such as artificial intelligence across all production facets.
This cascade of project demands hits at the heart of the commercialization challenges faced within the Chinese large model market. The 3 trillion yuan allocated is not solely aimed at the large model domain; however, Xu Wenqiang, director of the Forward Industries Research Institute, has expressed optimism that AI applications will evolve from current business-centric scenarios into more comprehensive decision-making frameworks. By 2028, the AI large model sector in China is projected to surpass 100 billion yuan, with an anticipated compound growth rate exceeding 50% over five years.
Despite the market-centric approach being utilized by central enterprises, the backing of state-owned capital amplifies their motivation to cultivate industry growth and stimulate demand. The bidding landscape reveals an array of traditional industries—including energy, telecommunications, finance, and education—embracing digital upgrades. Well-known companies like China Mobile, China Tower, Southern Power Grid, and the State Energy Group frequently appear on procurement lists.
Yet, the attitudes of large model enterprises in bidding processes exhibit divergence. According to self-media platform "Silicon Star Pro," Zhipu AI led in bidding success in the first half of the year, with considerable competition from Baidu and iFlytek. The other notable players like Huawei, SenseTime, Alibaba, and Tencent have significantly trailed in securing procurement, with the ‘six dragons’ label not fully represented in the bidding landscape.
Baidu, iFlytek, and Zhipu AI are often deemed the triumvirate of ToG market competition. Recent statistics illustrate that iFlytek, during July and August, secured 112 and 127 projects respectively, with the latter amounting to a staggering 150 million yuan in a single month.
Industry experts suggest that major players possess inherent advantages in navigating the ToG domain. Governance bodies and large SOEs tend to gravitate towards bidders with established credentials, often relegating smaller firms to the role of subcontractors under the wing of these larger entities.
Both Baidu and iFlytek's established footholds in governmental business bolster their large model portfolios. In contrast, Zhipu AI initially pursued a ToB trajectory and now sees a variety of project types, with financial scales that range from tens of thousands to millions.
Within this hyper-competitive arena of the “Three Kingdoms,” the ultimate victors remain uncertain.
Winners and Losers
The AI large model ToG sector is shifting the nature of deliverables and services, yet the clientele and market characteristics persist. Famous venture capitalist Yi Feifan pointed out that approximately 70% of the top 60 enterprise service companies in China emphasize ToG, with half reporting gross margins below 50%, in stark contrast to American counterparts averaging above 70%.
This dichotomy stems from the fact that most demand from ToG clients, typically government agencies or massive SOEs, tends to come with hefty order sizes. Nevertheless, intricate implementation scenarios and stringent risk management standards inflate the execution costs, resulting in thin profit margins for successful bidders. The transparency of procurement pricing increasingly stabilizes profit margins in many orders.
Furthermore, G-end clients often favor integrators capable of delivering comprehensive solutions, who then manage project needs or source out other subcontractors. This reality positions smaller suppliers in a precarious position. While they must rely on integrators for orders, their operational flexibility can lead to higher quality income streams. Integrators face the dual challenge of managing extended payment cycles from G-end clients while ensuring the delivery standards of suppliable partners, which can compel them to accept an increasing number of contracts.
Consequently, companies like Baidu and Huawei, which have emerged as frontrunners in the ToG market, stand out because they can function as both integrators and suppliers, thus preserving reasonable profit margins. Furthermore, their diverse operations across sectors enable them to maintain sustainable revenue streams from their ToC activities even when ToG segments underperform.
The strategy employed by iFlytek is notably distinctive; they have maintained a dual approach of integrating ToG and ToC initiatives for years. Hu Yu, one of the co-founders, articulated their mission to leverage core technologies to address national planning challenges while also catering to mass consumer markets with products like translation devices, a duality often challenging for outsiders to grasp.
Despite the significant investments made by Baidu and iFlytek in the large model sector, their financial reports reveal divergent impacts. For instance, Baidu reported a second-quarter revenue of 33.9 billion yuan, with a core operating profit of 5.6 billion yuan—reflecting a 23% year-on-year growth, largely attributed to accelerated growth in its cloud services, consistent with the push of large model initiatives. Conversely, iFlytek’s half-year report showed revenues of 9.325 billion yuan, an 18.91% increase, yet revealed a net loss of 401 million yuan—its first-ever half-year deficit since going public, a reflection of its heavy investment in large model infrastructure.
This performance discrepancy can be traced back to insights shared by Baidu's founder Li Yanhong during recent discussions at a quarterly meeting. He noted that ToB business models must be standardized. This entails a project-based approach that requires significant resources for personnel deployment and extensive backend development.
Products like Comate may currently struggle to compete price-wise, but Li remains optimistic about their potential. He believes that with sustained investment and gradual enhancement of product standards, they could carve out a more competitive market niche.
Focusing on mid-tier clients has become imperative. As Li pointed out, securing profits from larger clients can often be elusive, owing to their substantial investment needs, while revenues from smaller clients also present challenges due to limited budgets.
Within this competitive triad of large model ToG ventures, Zhipu AI—a startup—continues to seek funding. Just last month, they completed their 11th round of financing. Their aggressive pursuit of ToG opportunities appears to supplement their ToB strategy more than reliance on large contracts from SOEs.
This raises questions about the intentions behind the participation of the trio, with competitors remaining dormant in their approaches.
Paths and Pitfalls
Entrepreneurs focused on China's ToB landscape find themselves grappling with the ToG market, juggling both excitement and frustration. Some companies soar with government reforms, while others remain ensnared in a web of unpredictable business relationships.
Former chief AI scientist at Alibaba, Jia Yangqing, noted two central dilemmas surrounding the commercialization of large models: the direction of revenue flow diverges from previous norms, and the brief time frame allowed for large models to generate revenue compared to traditional software.
Zheng Yu, vice president of JD.com, likened the government to a "super B-end customer." The essence of the large model ToG market still adheres to ToB principles, such as the separation between product users and decision-makers.
One product manager actively engaged in the ToG sphere emphasized the necessity of interpreting policy documents accurately to discover metrics that bolster leadership performance, all while considering practical needs that must be met upon product delivery.
Moreover, the procurement processes for AI models might influence traditional workflows, necessitating internal negotiations among various executive leadership to align organizational restructuring.
This reflects the challenges apparent in the SaaS sector's profitability model. As disclosed by one SaaS entrepreneur, a digital project initially concluded that it could significantly reduce offline labor costs, successfully traversing stages of requirements gathering and prototype development, only to falter during internal approval processes.
Unlike overseas SaaS companies that predominantly appeal to small and medium clients with high product standardization, Chinese SaaS firms chiefly derive income from major clients, who exhibit a greater demand for custom solutions—unwittingly increasing overall cost structures and reducing gross margin.
A case in point is ByteDance’s collaboration tool, Feishu, which recently faced layoffs due to its pursuit of high-end clients amidst intensifying competition for the SMEs market from Alibaba’s DingTalk and Tencent’s WeChat Work. While aiming for exemplary case studies through large clients—entailing higher customization demands—Feishu ultimately expanded its team excessively, hindering efficiency and leading to a one-third reduction in staff.
When overlooking the unique subjective factors of ToG projects, the age-old cost conundrum remains an inherent challenge for large model businesses. Outside of substantial single projects for government and SOEs, most B-end customer revenue relies on continued API usage, necessitating continual client engagement to ensure consistent cash flow.
As major firms aggressively engage in price competition in the small to medium business segment, Zhipu AI made headlines in June by slashing its entry-level GLM-3 Turbo model's usage cost from five to one yuan per million tokens—a staggering 80% decrease.
Despite the definite drawbacks observed in the ToG sector relative to its ToC and ToB counterparts, many large model enterprises still view it as the quickest route to recoup investments. Yet, indulging too excessively in the immediate rewards may obfuscate the broader strategic vision, which could lead the necessary profits back to funding further developments.
Leave a comment