Use Case: How Small Startups Can Capitalize on AI Large Language Models KellyOnTech

AI large language model (LLMs) startups have been gaining popularity in recent years and are known for their high valuations. Among them, the most famous OpenAI is valued at $83 billion in 2024, accounting for nearly 17% of the entire AI market of $500 billion.

AI LLMs Startups Dance with the Tech Giants

The large language model (LLMs) startups in China and abroad have chosen to cooperate deeply with tech giants.

OpenAI

Founded in 2015, OpenAI and Microsoft have cooperated deeply and have entered the third stage since 2019,2021 and 2023. In addition to deploying OpenAI’s models in Microsoft’s products, it has also launched Azure OpenAI services, which enable developers to build cutting-edge AI applications by directly accessing OpenAI models. In 2023, OpenAI’s annual revenue reached US$1.6 billion, a year-over-year increase of 700%, compared with US$200 million in 2022.

Anthropic

Anthropic is a relatively new company founded in 2021 by former OpenAI employees. It emphasizes aligning AI behaviour comply with human intentions, and its developed Claude series of models are also widely praised.

In March this year, Amazon completed a $4 billion investment in Anthropic, which uses Amazon web services (AWS) as its primary cloud provider for mission-critical workloads, including safety research and future fundamental model development. In addition, Anthropic will use AWS Trainium and Inferentia chips to build, train, and deploy its future models, and has a long-term commitment to provide AWS customers around the world with access to its future foundation models on Amazon Bedrock, AWS’s fully managed service that provides secure, convenient, and broad industry access.

Minimax

Minimax was established in 2021. The founding team came from SenseTime, and investment institutions include Alibaba, Tencent and other giants.

Minimax adopts a “take all” approach, making many To-C products and To-B products, including the self-developed role-playing AI chat application Glow, the self-developed general large language model ABAB, and the corresponding solutions to enterprise customers. Minimax is also one of the eight LLM companies that passed the “Interim Measures for the Management of Generative Artificial Intelligence Services”.

Moonshot AI

Founded in 2023, Moonshot AI mainly develops To-C products. For example, gaming application “Coax Simulator”, which is connected to the large language model of Moonshot AI. This game is quite representative. It designs various interactive scenarios for couples. Users can define the dialogue scenes and roles by themselves, and AI characters simulate real dialogues. In October 2023, the company lauched KIMI AI, the world’s first intelligent agent that supports 200,000 Chinese characters.

Alibaba invested a combined total of about $800 million in it in fiscal year 2024, accounting for 36% of preferred shares, along with Alibaba’s computing power support as part of the investment.

Why do AI large language model companies in China and abroad choose to cooperate deeply with large companies?

In addition to access to relatively abundant capital, one of the reasons why LLM startups are tied to well-known companies is because the computing power cost of training large language models is high.

According to Lambda Labs, OpenAI used 1,023 A100 GPU to train ChatGPT. If you factor in storage, power consumption, and site rental and operation and maintenance cost, according to estimates by software industry experts, the operating cost of ChatGPT is between $100,000 to $700,000 per day, or between $3 million to $21 million per month.

How can small startups seize the opportunity of AI large language models? The answer is to focus and find a breakthrough.

How Can Small Startups Seize the Opportunity of Large Language Models?

Here I use the AI Cantonese large language model project introduced by the International Youth Entrepreneurship Council as an example.

We know that there are over 6,000 human languages in the world. However, many languages will disappear because of the limited language resources available for computational processing and analysis.

Although Cantonese is one of the most widely spoken languages in the world, with over 86 million speakers globally covering Hong Kong, Macau, Guangdong Province, Southeast Asia and overseas China Towns，it is a low resource language.

How difficult is it to let AI learn Cantonese?

Take the “water culture” of Cantonese as an example.

Chatting is called “blowing water”, commissions is called “pumping water”, gossips is called “boiling water”.

What are the application scenarios of the AI Cantonese large language model?

AI Cantonese large language model can be used for meeting records, social listening, customer support, news summaries, translation, call centres, education and digital marketing.

The startup VOTEE AI team has found its niche in the Cantonese language segment and seized the opportunity of the AI large language model. In June this year, it just released the enterprise version, supporting on-cloud (API), on-premises (commercial use license), and even on-device (community)deployment.

In addition, the company launched the Hong Kong Artificial Intelligence Research Community to further improve the Cantonese corpus collection, annotation and community modeling, which is strongly supported by the Hong Kong government and university research institutes.

I hope you can fully use your own strengths, look for an alternative way to find your own niche in areas that large companies have not yet entered.

中文版