calculate

Perception

cognition

consciousness

Specific scenarios, universality check, switching scenarios requires retraining

Universal

Bringing cost reduction and efficiency improvement to the industry

Domestic AI technology is only 1-3 years behind the United States, with huge development potential

Reasons for the widening gap

2025

upstream

midstream

downstream

Upstream chip

Upstream optical module

Midstream algorithm layer, large model

Downstream landing scene

The training chip has a large computing scale, mainly GPU

Inference is to use the trained model to infer using new data, which requires less calculation.

The current market mainly comes from training, but the demand for inference has begun to increase explosively.

Category 4: Versatility gradually decreases (flexibility), but efficiency gradually increases

field programmable gate array

Global revenue of US$12.5 billion in 25 years, with an annual compound rate of 11%

China

Overseas Xilinx and Intel dominate, while domestic manufacturers emerge.

Process manufacturing and gate-level scale are important indicators for evaluating product performance.

Integrated circuits customized for specific purposes, high performance and low energy consumption

In 20217, the world is 170 billion yuan, with an annual compound rate of 9%

There is no obvious leading manufacturer yet

The cycle for planning, infrastructure, and equipment to be equipped with optical modules is 2 years. In 2024, there will be a concentrated explosion of demand for optical modules brought by AI.

Optoelectronic co-packaging technology assembles the switching chip and optical engine together in one slot

Traditional: pluggable optical module

Low power consumption, large bandwidth and high speed

As the scale of data centers continues to expand, the demand for optical communication bandwidth and transmission rate increases, and the traditional model is not cost-effective.

USD 5.4 billion in 2027 (global)

It will be gradually commercialized in 2024

overseas

domestic

At present, large models are mainly oriented to NLP, such as GPT-3

Learn a common language so that the model has language understanding and generation capabilities

development challenges

It is difficult to have a major outbreak in the short term

Text-image fusion, GPT-4

Mostly small models

Those with advanced progress include

Tencent and Baidu have layouts in three types, and Huawei has one more large scientific computing model

ecological difference

next 2-3 years

Pay attention to the upstream manufacturers that cooperate with these three companies

For consumer needs: games, office, film and television, search

The four stages of game launch: project establishment - research and development - distribution - commercialization

AIGC: Simplify the two major links of R&D and distribution, shorten the R&D cycle, and reduce production costs.

R&D end

Issuing side

Currently saving more than 50% of manpower and time

ROE will increase by 20% in the future

Dachang

Office: Kingsoft

Bloomberg

Flush, Oriental Fortune

other