Review | From Microsoft + OpenAI to NVIDIA! How are overseas AI large-scale model enterprises cultivated?

Huatai Securities believes that the AI big model industry is capital-intensive, talent-intensive, and data-intensive. The key to the success of big model enterprises is how to form a "data-model-application" flywheel. By deeply reviewing overseas basic big model enterprises and sorting out the resource endowments and path choices of each company, Huatai sees:

1. Microsoft&OpenAI's continuous investment in disruptive innovation is the deep reason for their current leadership;

2. Alphabet-C has a deep accumulation of talent and technology, but has not formed synergy in its management structure;

3. Nvidia has formed a moat that other chip companies find difficult to surpass through the CUDA framework;

4. Meta and others are currently fighting back through model open source.

Currently, domestic BIDU-SW, Alibaba, Sensetime, Huawei, and others have actively joined, and the industry presents a competitive pattern of "hundred-model war". Whether a flywheel can be formed is the key to the final victory.

Microsoft&OpenAI: The leader in technology level and product landing

Reviewing the development process of OpenAI, the following characteristics and strategic choices are crucial:

1. High talent density and a firm belief in AGI internally;

2. Grasping the correct technical route without wavering, quickly choosing the Transformer architecture after its birth, and firmly choosing the decoder route among them;

3. Cooperating with Microsoft to solve the problem of insufficient computing power;

4. Promoting product landing and forming a positive feedback loop of model calling, data feedback, and model iteration.

Currently, Microsoft has integrated OpenAI's big model capabilities into its various ToB and ToC products or services such as office software, search, operating systems, and cloud services.

Alphabet-C: Rich technology reserves and a broad AI-enabled business ecosystem, currently accelerating the catch-up of productization and ecologicalization

Alphabet-C has rich AI technology and talent reserves, but its productization has lagged behind OpenAI: in terms of algorithms, it launched the current main LLM basic infrastructure-Transformer in 2017, and milestone big models such as BERT and PaLM-E; in terms of computing power, it has self-developed TPU chips; TensorFlow framework was also launched by Alphabet-C.

Alphabet-C has a rich product ecosystem including search engines, maps, email, office suites, etc., and there is ample room for AI big models to land. Since 2023, Alphabet-C has accelerated its catch-up with Microsoft and OpenAI:

1. Merging Alphabet-C Brain and DeepMind, the two major AI teams, to pool resources;
2. Accelerate the productization and landing of large models. PaLM-2 was released at the I/O conference and has been applied in more than 25 functions and products, strengthening the collaborative ability of the chatbot Bard with Alphabet-C and other external applications.

Meta: A defensive strategy for open-sourcing models to deal with opponents' strong closed-source models

One of Meta's biggest contributions to the AI industry is the open-source deep learning framework PyTorch, which has become one of the most popular frameworks in the field of deep learning due to its flexibility, ease of use, and high performance.

Entering the era of large AI models, Meta has been slow to productize, but has adopted a model open-source strategy as a defensive strategy against competitors' strong closed-source models.

Meta has successively open-sourced large language models LLaMa, CV large models SAM, Dinov2, speech large models MMS, and multimodal large models ImageBind in 2023, promoting the prosperity of the open-source community.

Among them, multiple open-source models have been derived based on LLaMa, and their performance is close to that of Alphabet-C and OpenAI's proprietary models, making them a cost-effective choice for small and medium-sized enterprises and developers.

NVIDIA: A leader in AI chips, AI cloud services, and model foundries with new business models worth paying attention to

NVIDIA is the world's largest GPU supplier, continuously building a moat around CUDA, and forming a unified ecological closed loop with the three hardware components of CPU+GPU+DPU.

The key factors that enable the company to continuously strengthen its leading advantage are:

1. Sensitively capturing changes in terminal demand. From 2010 to 2015, the focus shifted to high-end gaming graphics cards; after 2018, it shifted to data centers and autonomous driving development;

2. The soul figure Huang Renxun leads the decision-making and leads the industry change;

3. Acquisitions to improve the business map.

In terms of business models, the company is deeply bound to TSMC and follows the OEM model, and recently launched AI Foundations, positioning itself as a "supercomputing cloud service + model foundry", which is worth paying attention to as it transforms from hardware to computing power cloud services and MaaS business models.