ChatGPT人工智能引领新一轮科技革命：2022年末面世的ChatGPT在23年3月再次迎来GPT-4的突破。最新发布的GPT-4有两大亮点：一是多模态与长内容生成能力，能将一张草图快速编为完整网站，能够在法考、数学考试上取得优异成绩等，成为在知识、技能、逻辑领域的全能“人才”。二是商业化进程加快，本次GPT-4的发布伴随着New Bing的接入，微软随后发布Microsoft365 Copilot，极大提高office的生产力和交互方式，同时GPT-4通过开放API接口，尝试接入更多的商业合作伙伴，以创造出更多商业化的应用。

GPT-4 leads the next phase of the artificial intelligence tech revolution

After OpenAI shook the global information market with the launch of ChatGPT at the end of 2022, its newly released multimodal GPT-4 earlier this month is making waves again. We take a closer look at two key impacts:

•Multimodal: GPT-4’s multimodal and long-form content generation capabilities allow it to quickly transform a sketch into a complete website, for instance, or produce excellent results in law and mathematics examinations. Its all-rounded prowess covers diverse applications of knowledge, skills and logic. 

•Monetization: the commercialization progress accelerated with GPT-4, which can be accessed via Microsoft’s new Bing chatbot. Microsoft went on to release 365 Copilot, a functionality that has greatly raised productivity and interactivity in its Office suite. OpenAI aims to add more partners for GPT-4 through open API to create more commercial applications.

算力为ChatGPT底层瓶颈。目前ChatGPT曾因访问量激增而导致官网瘫痪，暴露了其算力不足的问题。在当今以深度学习为中心的人工智能发展中，AI模型的进步主要依赖于大规模数据的消化，伴随着AI模型的数据量、结构的复杂程度不断增加，模型尺寸呈现指数级增加，但传统服务器已无法满足AI时代下算力超越摩尔定律的翻倍速率，需要CPU与加速芯片组合的AI服务器来满足更高吞吐量互联的需求，AI服务器已成为支撑人工智能发展的重要算力支撑。根据IDC数据，2021年全球服务器市场规模为992亿美元，同比增长9.01%， AI服务器市场规模为156.3亿美元，占整体市场的15.76%，同比增长39.1%，远超服务器行业平均增势，该比重在ChatGPT加持下有望继续成长，如纬颖科技2022年AI服务器出货量已占比20%，我们认为未来数据中心发展核心为人工智能方向。

Computing power capacity a bottleneck in ChatGPT development

AI servers provide key support: the official ChatGPT website has been paralyzed as visits surged, exposing a lack of computing power. Current AI development centers on deep learning, where the AI model progresses by digesting masses of data. As the data volume rises with the structural complexity of the AI model, the model size increases exponentially. In the AI era, traditional servers are no longer able to match the doubling rates of computing power beyond Moore’s Law. Instead, AI servers are needed for higher throughput interconnection, combined with CPUs and accelerator chips. AI servers are the key to computing power support for AI development. 

Growing faster than average servers: IDC data shows that the global market for servers amounted to USD99.2bn in 2021, an increase of 9.01% yoy, of which AI servers accounted for 15.76% at USD15.63bn. The AI server market growth of 39.1% yoy is much faster than the average server market rate. With the advent of ChatGPT, we expect the proportion of AI servers will keep growing. Take Wiwynn for example: its AI servers accounted for 20% of total server shipments in 2022. We believe that AI will lead the way to future data center development.

由ChatGPT快速迭代引出的算力缺口和云平台繁荣，带动服务器终端客户capex进展超预期。微软早在2019年以10亿美金投资OpenAI，2020年买断GPT-3背后基础技术的独家许可，Azure云平台成为其独家云供应商。微软数据披露，从GPT到GPT-3，参数量从1.17亿到1750亿，增长1500倍，庞大的参数量需要算力和服务器持续迭代。同时，2023年3月，Azure首次向B端开放企业级服务，并将ChatGPT技术扩展到Power Platform上，将其打造为开发者们的新“栖息地”。综合来看，GPT-4参数量指数级增长+开放Azure服务带来的客户增量，推动微软capex回升向上。云服务商竞争仍在继续，客户Capex回暖使AI服务器产能有望受益扩张。

Cloud compute power gap lifts end-customer server capex beyond expectations

Microsoft invested USD1bn in OpenAI in 2019 and one year later bought the exclusive license to the fundamental technology behind GPT-3, making its Azure cloud platform the exclusive cloud provider. Microsoft data reveals that the number of parameters from GPT to GPT-3 increased from 117m to 175bn, a jump of 1,500x. The immensity of parameters comes from continuous iteration, which utilizes enormous computing power in servers. Earlier this month, Microsoft Azure launched B2B services for the first time, extending ChatGPT technology to its Power Platform, making it a new habitat for developers. Taken together, exponential rises in GPT-4 parameters and customers brought by opening access to Azure services have pushed Microsoft’s capex back up again. Competition remains intense among cloud service providers and we expect to see a customer capex recovery to drive the expansion of AI server production capacities.