Comparison of Generative AI Strategies Among Chinese and American Cloud Providers

The development of generative AI is still in its early stages, and the technical routes have not yet converged. A significant divergence currently exists: should large models be ecologically open or vertically integrated?

By Wu Junyu

Edited by Xie Lirong

Generative AI has been racing for a year now, and the core participants in this arms race are cloud computing vendors. In the past year, global cloud vendors have been reshaping their technical architecture around generative AI into three layers: infrastructure, model platforms, and application ecosystems. The logical relationship among these three layers of technical architecture is that the infrastructure provides the computing power required for training and inference of large models centered around AI chips; the model platform integrates self-developed, third-party, or open-source large models to improve application development efficiency; and generative AI applications need to select models for development, directly facing business needs.

While laying out the three-layer technical architecture, the strategies of cloud vendors vary. A major divergence is whether large models should be ecologically open or vertically integrated.

A typical case of the ecological open route is Amazon Web Services (AWS). Its large model hosting platform, Amazon Bedrock, integrates over 20 selected large models from 7 leading model companies worldwide. In simple terms, Bedrock is like a supermarket for models. Amazon believes that no single model can be suitable for all scenarios. A multi-model approach can meet customer needs and also provide Amazon with more opportunities for computing power consumption, encouraging partners to expand the ecosystem. This can drive the industrial flywheel and create a virtuous cycle.

On April 23, AWS updated the Bedrock platform. The update logic is to allow enterprise users to use large models on Bedrock with low barriers, high efficiency, and low risk. There are three important releases: first, the introduction of Meta’s new open-source model Llama 3, AWS’s self-developed Titan series models, and a new model from Cohere; second, the simplification of user import, customization, fine-tuning, and evaluation functions for models; third, the provision of tools to control model behavior and avoid harmful, erroneous, or risky content.

A typical case of the vertical integration route is Microsoft Azure. Although Microsoft also integrates some third-party models, its flagship model is OpenAI’s GPT-4. The benefit of vertical integration is that Microsoft covers the four layers of computing power, platforms, models, and software, resulting in significant short-term revenue growth. Microsoft’s intelligent cloud has already rebounded in revenue growth due to generative AI in the second half of 2023.

A typical case of the middle route is Google Cloud. It ranks third in the global cloud market share, following AWS and Microsoft Azure. Google Cloud has developed three self-developed large models, the most among its peers, and also has one open-source large model. It has even integrated some third-party models. Google’s strategy is to hedge bets across multiple fronts, pursuing self-developed, open-source, and open routes simultaneously.

The development of generative AI is still in its early stages, and the technical routes have not yet converged; continuous trial and error are still required. Each company’s ecological route is in the exploratory stage, and there are no clear boundaries or absolute right or wrong. Considering the subsequent market development trends, all parties do not rule out further adjustments or even changes to their routes.

Currently, the landscape of generative AI layouts among Chinese cloud vendors is also taking shape. Similar to AWS’s ecological open route is Alibaba Cloud. Similar to Microsoft Azure’s vertical integration route is Baidu Smart Cloud. Other cloud vendors are finding their strategies based on their strengths, with Huawei Cloud utilizing domestically produced Ascend AI chips to expand its territory, and Tencent Cloud focusing on selling SaaS applications that have been transformed by generative AI. Telecom operator clouds are still hoarding AI computing power on a large scale.

Comparison of International Cloud Vendors’ Strategies

Generative AI is becoming a growth engine for cloud vendors.

From 2021 to 2022, the U.S. cloud market experienced relatively sluggish growth. The growth rates of AWS, Microsoft Azure, and Google Cloud all declined. However, in the second half of 2023, the U.S. cloud market has rebounded. Microsoft has benefited from generative AI, leading to a significant rebound in revenue growth. However, the cloud is a game of ecosystems. The cultivation cycle of ecosystems often lasts 3 to 5 years. Factors such as the number of partners, fairness, and openness of the ecosystem typically determine the growth potential.

According to data from international market research firm Gartner in July 2023, the top five global public cloud market players are: AWS (40.0%), Microsoft Azure (21.5%), Alibaba Cloud (7.7%), Google Cloud (7.5%), and Huawei Cloud (4.4%).

Image description: The generative AI technology architecture of AWS

AWS is the largest cloud vendor in the global market share. Ecological openness has always been its strength. This continues into the generative AI business. AWS has formed a complete and open generative AI strategic layout around infrastructure, model platforms, and application ecosystems.

The key to infrastructure is to provide powerful, stable, and low-cost computing power for model training/inference. AWS has deployed NVIDIA AI chips and two self-developed AI chips (training chip Trainium and inference chip Inferentia). The key capability of generative AI at the infrastructure layer is managing computing clusters of tens of thousands of nodes. The difficulty lies in the fact that during the training of tens of thousands of nodes, graphics cards, networks, and systems may fail at any time. Once interrupted, it not only costs a lot of time but also wastes expensive computing power. Currently, AWS has the ability to support over 100,000 Trainium 2 chips for parallel training. This can effectively save model training time and improve computing power utilization efficiency.

The key to the model platform is ecological openness, expanding the model selection space, and lowering the barriers to model use. To this end, AWS’s Bedrock model platform integrates over 20 selected large models from 7 leading model companies worldwide. The scope of cooperation exceeds that of other cloud vendors. This is because developing generative AI applications requires selecting suitable models while comprehensively considering factors such as parameter scale, accuracy, performance, and price. A person from a Chinese SaaS software company stated in March this year that his company has used 11 domestic and foreign models across different businesses. This is because each model excels in different business areas, and no single model can be suitable for all scenarios; they must each play their respective roles.

To further lower the barriers to model use, Bedrock provides functions such as user import, customization, fine-tuning, and evaluation of models. Enterprises can customize models integrated into Bedrock, which can reduce operational costs and accelerate application development. One of the risks of using generative AI is model hallucination (the behavior of AI making nonsensical statements). To address this, Bedrock also provides tools for enterprises to remove privacy-sensitive or harmful information.

Additionally, the key to a thriving application ecosystem is prosperity. Currently, enterprises have deployed a large number of SaaS applications. Generative AI applications are lighter, more fragmented, and atomic than SaaS applications. They serve as functional components of SaaS applications and can be invoked in chats, dialogues, and searches. This is similar to how mini-programs exist within super apps like WeChat and Alipay. They are rich in functionality and efficient.

To this end, AWS provides two product services. One is the AI assistant Amazon Q, which has basic functions such as QuickSight (data reporting), Connect (intelligent customer service), Supply Chain (supply chain management), and CodeWhisperer (code generation), allowing enterprises to customize their own AI assistants. The second is to open compatibility with mainstream SaaS applications; Amazon Q has over 40 connectors compatible with popular data sources, covering well-known SaaS applications such as Google Drive, Microsoft 365, Salesforce, and ServiceNow. When enterprises use Amazon Q, they can connect the data of these SaaS applications, linking internal networks, knowledge bases, and process documentation.

The core principle of AWS’s generative AI layout is to enlarge the cake—doing well in computing power, model platforms, providing basic applications, connecting data interfaces, and setting platform rules. It maintains clear boundaries with SaaS companies, AI model companies, and other partners.

Ecological fairness and openness is a historical tradition of this company. Since its inception in 2006, a number of well-known SaaS companies such as Salesforce and ServiceNow have naturally formed ecological cooperation around AWS. The tacit understanding between AWS and its partners is that AWS provides basic cloud services and core applications, while upper-layer industry applications are jointly constructed by partners. Each partner plays its role, and as they grow, they will consume underlying computing power, leading to revenue growth for AWS. This creates a virtuous cycle of industrial flywheel. Therefore, conflicts of interest between AWS and its partners are relatively few.

Unlike AWS, Microsoft follows a vertical integration route. This is particularly evident in the model platform and application ecosystem layers.

At the model platform level, Microsoft has prioritized resources for OpenAI’s GPT-4—it is the default option for the Microsoft Azure model platform. The Microsoft Azure model platform primarily supports four types of models: OpenAI’s GPT-4, Microsoft self-developed models, models from strategic partner Mistral AI, and some open-source models. For Microsoft, it can simultaneously gain revenue from computing power and models.

At the application ecosystem level, Microsoft’s flagship application is the Copilot assistant, which directly targets enterprise customers. Copilot is deeply integrated with Microsoft’s own software (such as the Office suite, Teams collaboration platform, GitHub code platform, Dynamics 365 development platform, and Power BI reporting tool). Enterprises must pay for AI functionalities. With the Copilot assistant, Microsoft has increased the revenue potential of these software products.

Microsoft is excelling across all four layers of computing power, platforms, models, and software. This strategy has led to significant revenue and profit growth. In 2023, the revenue growth rate of Microsoft’s intelligent cloud rebounded by 5.4 percentage points compared to its lowest point that year; the operating profit margin rebounded by 5.5 percentage points compared to its lowest point that year.

Unlike AWS, Microsoft is a typical software company. Microsoft excels at developing its flagship software products, and its historical success came from high revenues and profits from Windows, Office, and other flagship products. The logic behind Microsoft’s successful transformation to the cloud post-2010 is the integration of its traditional software business, self-developing, acquiring, and investing in a number of software products to form its own cloud ecosystem.

Software gross margins typically reach 60%-80%, approximately 20% higher than IaaS (Infrastructure as a Service) computing resources. Microsoft can achieve high revenues and profits on its own, thus it is relatively less dependent on partners. However, this also means that its ecosystem can be somewhat exclusive. Microsoft’s self-developed and invested software businesses can easily lead to direct competition with other SaaS/PaaS software partners.

Google Cloud’s generative AI layout involves multi-pronged bets, adopting a strategy of self-developed, open-source, and open approaches simultaneously. It has developed multiple closed-source large models, open-source large models, and invested in several generative AI startups.

At the model platform level, Google’s closed-source large model Gemini competes with OpenAI’s GPT-4. Google’s open-source large model Gemma competes with Meta’s Llama. Google’s model platform also supports third-party models such as Anthropic’s Claude 3. In the application ecosystem layer, Google’s self-developed generative AI applications focus on traditional areas of strength such as video, voice, text, and translation.

Google has adopted a broad investment strategy in generative AI startups. It is a shareholder of Anthropic, with investment amounts only lower than AWS (AWS invested $4 billion, while Google invested $2.3 billion). Google has also invested in large model companies such as Character.AI, large model companies like AI21 Labs, the model open-source community Hugging Face, and video model companies like Runway.

A Chinese software industry investor commented in January this year that Google Cloud currently has a significant gap in market share compared to AWS and Microsoft Azure. As a follower, it faces considerable growth pressure. In the past few years, Google Cloud has achieved rapid growth through price wars. In 2023, Google Cloud began to adjust its strategy, achieving profitability, but its revenue growth rate is gradually slowing down.

He further explained that generative AI is an opportunity for Google Cloud to catch up and return to growth. The development of generative AI is still in its early stages, and technical routes have not yet converged; continuous trial and error are still necessary. The benefits of multi-pronged bets are that they allow for multiple trials. Moreover, one of Google’s investment requirements is that the invested companies must train their models on Google Cloud. This will lead to computing power consumption, thereby generating revenue. The challenge is to find a delicate balance, as there are contradictions in business models between self-developed and open-source models, and certain conflicts of interest between self-developed and third-party models.

Comparison of Chinese Cloud Vendors’ Strategies

The Chinese cloud market has been in a downturn for the past three years, with market growth continuously slowing. Generative AI is also seen as an important engine for growth in China.

International market data from IDC shows that the overall market size of China’s public cloud services (IaaS + PaaS + SaaS) reached $19.01 billion in the first half of 2023, a year-on-year increase of 14.7%. IDC data also indicates that the compound annual growth rate of the generative AI market in China over the next five years is expected to be 55.1%. The growth rate of generative AI far exceeds that of the public cloud market, prompting major cloud vendors to actively deploy AI.

Alibaba Cloud’s layout is relatively close to AWS, with a strategy of open ecology and model open-source, while also adopting a broad investment strategy. This can stimulate the consumption of underlying computing power for Alibaba Cloud, thereby driving revenue growth.

Alibaba Cloud ranks first in China’s cloud market share and revenue scale. Over the past three years, Alibaba Cloud has experienced significant pain points, with low revenue growth. The company’s core task during this period has been to optimize government and enterprise projects, improve revenue quality, and undergo multiple rounds of management adjustments. Currently, there are high expectations within Alibaba Cloud for generative AI as a growth engine. One judgment is that Alibaba Cloud’s revenue growth rate will gradually recover due to generative AI in 2024.

At the model platform level, Alibaba Cloud aligns with AWS’s thinking, providing enterprises with ample choice. Currently, Alibaba Cloud has self-developed the Tongyi series of closed-source models and has released three open-source models from the Tongyi series. It has also established the Modao community, learning from the Hugging Face open-source model community. Alibaba has invested in several top five generative AI unicorns in China, including Zhipu AI, Lingyi Wanyi, Baichuan Intelligence, MiniMax, and The Dark Side of the Moon. These companies are training large models on Alibaba Cloud and providing services externally.

Alibaba Cloud management has publicly stated that large models are still in the early stage of rapid technological evolution. The choice between open-source and closed-source models should be left to developers. How to balance the contradictions between open-source and closed-source business models? An Alibaba Cloud insider stated that open-source and closed-source are upstream and downstream relationships. Open-source is upstream in technology, mainly aimed at community participation in research and development iterations, expanding the scale of enterprise users, and ensuring technological leadership over peers. Closed-source is downstream, aiming for commercialization. In his view, open ecology and model open-source are aimed at forming a virtuous cycle of “the stronger the model, the more applications, the broader the user base, and the greater the computing power.”

Baidu Smart Cloud has adopted a similar vertical integration route to Microsoft Azure. Baidu’s investment in generative AI is aggressive. This is because Baidu Smart Cloud has gaps in revenue scale and market share compared to the top tier, and generative AI is an opportunity for it to narrow this gap.

Baidu’s goal is to achieve revenue growth through the sale of computing power, platforms, models, and applications, similar to Microsoft. Baidu has indeed reaped the first wave of generative AI dividends. In the fourth quarter of 2023, the revenue growth rate of Baidu Smart Cloud rebounded by 11 percentage points compared to its lowest point that year.

At the model platform level, Baidu focuses on its self-developed Wenxin series models. In April of this year, Baidu’s founder, Li Yanhong, openly expressed opposition to open-source large models. A Baidu Smart Cloud executive once stated that open-source software code is public, and community developers can participate in accelerating software iterations, helping companies reduce R&D costs. However, open-source models are black boxes, and no one knows the algorithms, parameters, or data. Developer participation does not significantly aid model iteration. Another reason is that training and inference costs for open-source models are very high. Developer participation cannot reduce R&D costs; it only increases computing power costs.

In the application ecosystem layer, Baidu is exploring generative AI applications with several software vendors and has established an AI application store. A Baidu Smart Cloud insider stated that China’s SaaS software industry has always been weak, and software development has primarily relied on labor outsourcing services. The downside is high costs and low efficiency. To address this, Baidu has launched the Qianfan AppBuilder application development platform, allowing enterprises to develop lightweight, fragmented, and atomic generative AI applications. This lowers the development and usage barriers.

Huawei Cloud has its own self-developed Ascend AI chips. This is its greatest confidence in laying out generative AI. Currently, there is a shortage of AI computing power in China. Chinese tech companies, including Alibaba, Tencent, ByteDance, and Baidu, are all using Huawei’s Ascend AI chips.

In October 2023, the U.S. Department of Commerce cut off advanced AI chips to China. After Chinese companies were unable to procure NVIDIA A100/A800, H100/H800 chips normally, the most realistic domestic alternative is Huawei’s Ascend. In 2023, Huawei’s Ascend AI chip production capacity is between 300,000 and 400,000 units. A Huawei Cloud executive stated that some Internet cloud vendors are using Huawei Cloud’s computing power to train AI large models.

Tencent Cloud excels in upper-layer SaaS applications. After launching the Hunyuan model in 2023, several self-developed SaaS applications (Tencent Meeting, WeChat Work, e-signature, AI code assistant, etc.) are undergoing AI upgrades. Currently, these SaaS applications are also being sold externally.

Telecom operator clouds (Tianyi Cloud, Mobile Cloud, Unicom Cloud) have all released self-developed large models. They are currently stockpiling AI computing power on a large scale. For instance, China Mobile’s capital expenditure on computing power is expected to increase by 28.0% in 2024, reaching 47.5 billion yuan. Investment in computing power will tilt towards AI computing power and domestic computing power. The scale of China Mobile’s capital expenditure on computing power even exceeds that of Alibaba, Tencent, Baidu, and other vendors.

A common viewpoint is that Chinese cloud vendors face two challenges in their generative AI layout: first, computing power, which will limit model capabilities; second, applications, as domestic software companies are generally small in scale and have yet to complete their cloud transformation.

In terms of computing power, due to the cut-off of high-end AI chips, the current model training costs for Chinese cloud vendors are relatively high. However, a positive signal is that domestic AI chips like Huawei’s are accelerating their deployment. A senior executive from a Chinese SaaS company stated that the capabilities of every major model vendor in China are rapidly improving. While no single large model is as strong as GPT-4, they can use MOE (Mixture of Experts, a model design strategy that combines multiple expert models for better performance) to leverage the strengths of each model. This way, their combined capabilities do not have a significant gap compared to GPT-4.

In terms of applications, there are opportunities as several software companies are using generative AI to narrow the gap. Currently, in software development projects, AI code platforms can significantly reduce the scale of human resources. This can enable the software industry to transform and upgrade from being labor-intensive to being technology-intensive.

Many cloud vendor insiders predict that in 2023, Chinese enterprises are generally “stockpiling computing power” and “rolling out models.” Due to high computing power costs and intense model competition, the trend of “rolling out models” may slightly decline in 2024. The focus will shift to “rolling out applications.” Regardless of whether it is “stockpiling computing power” or “rolling out models,” the ultimate goal is to develop applications and implement them in actual business.

Generative AI Restructuring Cloud Computing

Many technical experts from cloud vendors believe that as generative AI accelerates its landing, the embryonic form of a new generation of cloud computing is also beginning to emerge.

Around 2010, cloud computing gradually replaced hardware-centric traditional IT. Now, generative AI is reconstructing cloud computing. The changes are unfolding around three layers of technical architecture:

At the IaaS (Infrastructure as a Service) layer, intelligent computing power (GPU chips, etc.) is growing at a rate far exceeding that of general computing power (CPU chips). General computing power and intelligent computing power generally cannot be directly compared through EFlops (floating-point operations per second). However, data from the China Academy of Information and Communications Technology in 2023 shows that in 2022, the global scale of general computing power was 440 EFlops, while intelligent computing power was 451 EFlops. Intelligent computing power is growing at over 50%, far exceeding that of general computing power.

At the PaaS (Platform as a Service) layer, a MaaS (Model as a Service) platform has emerged above PaaS platforms, significantly improving the application development efficiency based on model invocation. The R&D costs of PaaS are high, and the sedimentation period is long; SaaS companies typically take 3 to 5 years to develop PaaS platforms. However, MaaS is more flexible, and invoking models to develop applications is simpler and more direct, and it will take on more of the development platform’s work in the future.

At the SaaS (Software as a Service) layer, a number of generative AI applications are integrated into SaaS applications, becoming functional components. They are lighter, more fragmented, and more atomic. They can be invoked in chats, dialogues, and searches through AI assistants. SaaS applications and generative AI applications are like the relationship between WeChat and mini-programs. Some single-function SaaS applications will be eliminated, while those that incorporate generative AI applications will see an increase in their paid value.

Many cloud computing technical experts believe that this round of innovation in cloud computing technology will last more than five years. In the short term, the gains and losses of individual cities and pools are not significant.

In 2023, Microsoft has taken aggressive actions, leveraging generative AI to seize the initiative. In the model and application layers, Microsoft has delivered impressive results. In 2024, AWS is expected to use its comprehensive and robust strategy to steadily advance in infrastructure, model platforms, and application ecosystems. Its strategic layout is even more comprehensive.

AWS and Microsoft Azure have been rivals since 2010, engaging in a 12-year battle during the cloud transformation era. Now, with the onset of AI transformation, a new competition has begun. The comprehensive competition across the three dimensions of computing power, models, and applications will ultimately determine the victor.

AWS’s advantage is that it holds over 40% of the global cloud market share, providing a more stable foundation. Its ecosystem is also larger, with broader cooperation relationships with SaaS software vendors and AI model companies. AWS’s growth potential is stronger. For partners, AWS serves as fertile ground for innovation, and as AI model startups continue to grow, more generative AI applications will flourish within SaaS software. This ecosystem will become more prosperous, and AWS’s revenue will subsequently increase.

Since 2024, structural adjustments have also begun to emerge in the Chinese cloud market, and a new round of competition is brewing. Generative AI is the focal point of this competition.

Internet cloud vendors and telecom operator clouds are currently switching roles. The strategic adjustments made by Internet cloud vendors over the past three years have reached their conclusion. Each company’s revenue growth rates and profit levels have significantly improved. Telecom operator cloud strategies are slowing down, beginning to lower growth expectations and reduce losses. Alibaba Cloud hopes to end the past three years of sluggishness with ecological openness and generative AI; Huawei Cloud aims to further expand its territory with domestic chips. Baidu hopes to achieve catch-up through vertical integration and generative AI. Telecom operators aim to hoard more intelligent computing power to expand market share.

Regardless of the strategy taken, the competition in the Chinese cloud market, like in the international cloud market, will be a long-distance race.

Leave a Comment Cancel reply