Produced by | Zidan Finance

Art Design | Qianqian

Edited by | Songwen

Where should big models go?

At a time when the competition between China and the United States for big models has come to a head, more and more model developers and enterprise users are falling into anxiety. This is because the scale and training methods of general models are approaching their limits, but the improvement in effectiveness has not reached the highest level expected by everyone, nor can it solve specific problems encountered by anyone at a specific point in time.

This is especially true for enterprise users. The widespread anxiety in the industry and the mismatch in the commercialization of big models have become core challenges restricting the development of big model technology.

On the one hand, since 2023, the rapid development of big model technology has attracted widespread attention from entrepreneurs.

They generally recognize that big model technology, especially when combined with various AI technologies, has a significant positive impact on reducing corporate costs and improving efficiency. However, how to integrate their own business with basic models? How can big models help enterprises improve efficiency in front-line production?

These needs cannot be addressed by a simple general big model. At this point, there is an urgent need for teams that understand big models, or are even big model developers, to step forward and find the intersection between enterprise operations and big models from the perspective of enterprise transformation. They need to train industry-specific models separately for this purpose and provide links and tools for their use.

Only in this way can entrepreneurs truly integrate their needs with general big models quickly, thereby enabling AI to integrate deeply with their own enterprise development and driving improvements in business efficiency.

On the other hand, more and more model development enterprises are beginning to focus on the B-end market, hoping to make breakthroughs in this field. The question is, how to successfully enter the market and establish a foothold, meeting enterprise needs while achieving growth in their own revenue and forming a sustainable competitive advantage. These issues also plague model development and operation enterprises.

Jeff Bezos, the founder of Amazon, once pointed out in a media interview that the essence of digital transformation lies in leveraging information technology and capabilities to drive core business changes. He believes that the three key elements of enterprise digital transformation include: establishing a digital enterprise strategic model and culture, mastering digital technology capabilities, and treating data as a strategic asset for the enterprise.

The realization of these capabilities requires a solid technical foundation to support them.

In the past era of digital intelligence, public clouds and cloud-native technologies provided a solid foundation for enterprises' informatization capabilities. However, entering the era of AI+ and big models, the foundation that supports enterprises in completing their own development capabilities based on big model technology should become the development direction of the entire industry.

Currently, whichever model development and operation enterprise can take the lead in achieving this goal will occupy a first-mover advantage in the future B-end market, thereby standing out in the fierce market competition.

1. Who is leading the industrialization of big models?

As everyone knows, the underlying foundation of big models is computing power, and the basis of computing power is chips. After the United States imposed increasingly stringent sanctions on Chinese chips, domestic chip manufacturers led by Huawei have become the driving force behind the development of big models in China.

In fact, this year should be considered a year of significant development for Huawei's Ascend series of computing chips. The Ascend 910B, which matches the performance of NVIDIA's A100, has already begun mass production. Meanwhile, rumors have circulated online that the Ascend 910C, which is comparable to NVIDIA's H100, has already begun shipping.

For domestic big model developers, the rapid development of iFLYTEK's Spark big model, which is fully built on the Ascend computing chip's ten-thousand-card matrix, demonstrates the capability of the Ascend series to support the development of big models in China.

With chips in hand, it indicates that domestic model developers and operators have the ability to continuously serve enterprise users, which is why everyone is beginning to focus on enterprise users.

According to the "China Big Model Bidding Project Inspection Report (July 2024)" compiled by the media, the company with the highest number of big model bids in the first half of the year was iFLYTEK. The data shows that in July, a total of 112 big model bidding projects were recorded in the first half of the year through public channels, with iFLYTEK leading in the number of bids and showing a steadily increasing trend. Among them, state-owned enterprises are typical customers where iFLYTEK's Spark business is implemented.

This actually represents a trend.

First, the current investment in big models by various types of manufacturers has shifted from a broad approach to a more refined one. Big model developers have realized that simply increasing computing power and investment indefinitely cannot achieve dynamic business balance, and the development momentum of big models is insufficient.

Therefore, on the basis of stable computing power, companies are beginning to seek breakthroughs in knowledge, data, and algorithms, using technological adjustments to improve model operation efficiency.

Among them, iFLYTEK is doing the best job.

Upgraded to Spark V4.0 on June 27, it not only fully benchmarks GPT-4 Turbo in terms of base capabilities but also launched multiple software and hardware products for education, healthcare, automotive, and enterprise agents, demonstrating its leading capability in Spark application implementation.

Crucially, this is an upgrade accomplished through technological advancements, algorithmic adjustments, and knowledge adjustments on the basis of previously established computing power, with less reliance on hardware investment than previous versions. This actually represents the development direction of the industry.

Second, in the bidding statistics for B-end enterprises, iFLYTEK consistently holds the top position.

The reason behind this phenomenon is that iFLYTEK has a deep understanding of the needs of these enterprises, especially the ability to guarantee the practical application of big models. This in-depth understanding and strong guarantee capabilities give iFLYTEK a significant advantage in competition.

Specifically, the needs of these enterprises are very clear. They hope to apply advanced models to actual operations while reducing related costs and technical investments. Whoever can effectively accomplish this transformation will win the favor of these enterprises.

Currently, iFLYTEK is the first choice.

2. Why iFLYTEK?

So why do B-end enterprises prefer to choose iFLYTEK when selecting a model provider?

The core reason is simple: iFLYTEK truly understands what these enterprises need and puts effort into model training and matching. Moreover, they have separately adjusted and trained models for some key areas, launching industry-specific big models directly derived from the general Spark big model. These models outperform third-party trained industry big models in terms of efficiency, accuracy, and concurrency.

In addition, through adjustments to model training and algorithms, iFLYTEK can create enterprise-specific big models with less computing power and higher efficiency, better aligning with enterprises' input-output requirements.

Currently, relying on big model technology in areas such as speech recognition, translation, and OCR, iFLYTEK has achieved the best in the industry with its existing computing power. The application value in areas such as education, healthcare, and automotive is also constantly demonstrated and has initially formed a commercial closed loop. At the same time, because it can provide big model solutions with the best cost-effectiveness in different sizes, Spark is leading in the industrialized business model.

Moreover, iFLYTEK is exploring the outcomes of big models at various stages in the AGI process, using various technical means such as distillation and pruning to produce outputs along the way and implement them in Spark big models of various sizes, reducing the size of the big models without compromising their effectiveness.

Coupled with their understanding of customer needs, they creatively provide many tools for interfacing with these models, significantly reducing the deployment difficulty for B-end customers.

More crucially, on the base capabilities of general big models, iFLYTEK collaborates with enterprise customers to iteratively develop industry models. Meanwhile, leveraging enterprise knowledge bases and industry knowledge bases, big models form successful cases in specific scenarios within enterprises, creating a virtuous circle of investment returns after customers invest in big models in specific scenarios.

The effectiveness of iFLYTEK stems from its commitment to research and development investment and its industrialized perspective from the beginning of model training.

The recently released 2024 interim report shows that R&D investment in the first half of the year reached 2.19 billion yuan, an increase of 32.23% year-on-year, accounting for 23.5% of revenue. Moreover, R&D investment continues to favor the Spark big model, with a total investment of 1.3 billion yuan related to the big model in the first half of the year. There was an additional 400 million yuan invested in R&D for the general big model, 160 million yuan in R&D for engineering the big model training and inference platform, and 120 million yuan in new investments for the promotion and application of the big model. Each BGBU increased investments in productizing the big model by over 600 million yuan.

The interim report indicates that with the support of major national research projects, iFLYTEK's Spark big model will maintain an industry-leading position in the coming years, and there will be no significant increase in self-funded computing power investments. In addition, iFLYTEK is actively building up its core talent pool, increasing its technical staff by 507 in the first half of the year while improving quality and efficiency in 2023.

3. Domestic models should quickly copy the commercialization answer

According to the interim report, iFLYTEK's revenue and gross profit in the first half of the year both increased significantly compared to the same period last year, with gross profit growth outpacing revenue growth.

The interim report shows that iFLYTEK generated revenue of 9.325 billion yuan in the first half of the year, an increase of 18.91% year-on-year, with second-quarter revenue of 5.678 billion yuan. Gross profit for the first half was 3.748 billion yuan, an increase of 19.08% year-on-year. Sales collections for the first half totaled 9 billion yuan, an increase of 1.5 billion yuan or 20.04% year-on-year compared to the same period last year.

Behind these figures lies the gradual implementation of iFLYTEK's commercialization strategy. As iFLYTEK's understanding of big model industrialization is at the forefront of the industry, it is gradually gaining momentum in competition.

Currently, based on strategic direction, market space, competitive advantage, and team strength, the first batch of seven strategic focus businesses and three exploratory businesses have been identified. The seven strategic focus businesses include key products in sectors such as education, consumer goods, and smart cars that have the potential to reach billions of yuan in revenue. The three exploratory businesses focus on exploring the application of the Spark big model in both the C-end and B-end markets, as well as exploring operational models for the Spark big model in the healthcare sector.

iFLYTEK's current priorities for commercializing its big model lie in three main areas: accelerating large-scale implementation in existing scenarios such as education, healthcare, and automotive; enabling external capabilities through the big model; and actively investing in exploring C-end applications represented by the Spark App, while also focusing on transformative innovation opportunities based on integrated big model hardware and software products.

Against this backdrop, in the first half of 2024, education products and services generated revenue of 2.86 billion yuan, an increase of 25.14% year-on-year; healthcare business revenue was 228 million yuan, up 18.8% year-on-year; the open platform generated revenue of 2.345 billion yuan, a 47.92% increase year-on-year; smart hardware revenue was 900 million yuan, a 56.61% increase year-on-year; and automotive business revenue was 350 million yuan, up 65.49% year-on-year.

According to a report released by the China Chamber of Commerce Industry Research Institute in March 2024, titled "Forecast Research Report on the Market Prospects of China's AI Big Model Industry in 2024," the market size of China's big model industry grew from 1.5 billion yuan in 2020 to 7 billion yuan in 2022, with a compound annual growth rate of 116.02%. In 2023, it was approximately 14.7 billion yuan, with companies like Baidu, Tencent, and Huawei vying for a share of the market. Analysts at the China Chamber of Commerce Industry Research Institute also predict that the size of China's big model industry will reach 21.6 billion yuan in 2024. In fact, the new era of AI infrastructure has only just begun.

(Image from Shutterstock, based on the VRF Agreement)

In the era of the Industrial Internet, there are countless specific connectivity tools available, and cloud services are becoming more standardized and simplified. It is no longer a problem for enterprises to improve efficiency in a specific field by introducing digital tools.

However, in the long run, these are not the core of the Industrial Internet, as current demand and supply are fragmented. In the future, as customer demand grows, this fragmentation will become increasingly apparent in the process of industrializing big models.

Therefore, big model developers and operators need to focus on industry models and operational efficiency. Only by truly finding and implementing relevant strategies, like iFLYTEK, will there be a future for the industrialization of big models.

*The cover image in this article is from Shutterstock, based on the VRF Agreement.