This is Part 5 of a deep dive into the developing infrastructure of AI. Here is Part 4 and Part 3.

I was recently at the Alibaba headquarters. And one of things that got my attention was Aspara, their cloud operating system. That got me digging into the differences in the tech stack between Alibaba Cloud, Tencent Cloud, Baidu Cloud and Huawei.

Here’s the short version:

• Alibaba Cloud is the largest cloud provider, operating both domestically and internationally. It is the Infrastructure King, with a big focus on retail, ecommerce broad capabilities and global expansion. Alibaba and its Apsara operating system is a good choice if you need a stable, Amazon-like experience for general business.
• Baidu AI Cloud is #1 by most measures. It is built almost entirely around artificial intelligence. It is the AI Brain of China Cloud – and leads in many smart segments. If you want a chatbot or AV, you think Baidu. Baidu AI Cloud is a good choice is most of your entire business is built on Generative AI or Autonomous systems.

• Tencent Cloud is the Media and Connectivity Titan. Its cloud architecture isn’t just about selling servers; it is built to support the world’s largest social media, video and gaming ecosystem

• Huawei Cloud focuses on hardware-heavy enterprise AI infrastructure. Think the CloudMatrix and huge data centers. Lots of focus on manufacturing and industrial projects. And it has full integration with Huawei smartphones, edge devices and 5G.

Apsara: Alibaba’s Operating System That Turns Data Centers into Giant AI Computers

I did some articles on AI infrastructure. And about how the massive data centers being built are not really data centers as we think about them (i.e., just storage, compute, connectivity). They are now big giant computers that act as AI brains.

• Understanding AI Infrastructure Part 1: AI Data Centers (Tech Strategy)

Withing Alibaba cloud, Apsara (i.e., the Apsara Operating System) is their proprietary, ultra-large-scale cloud operating system. Unlike a traditional operating system like Windows which manages a single computer, Apsara is designed to treat an entire data center as a single computer. It clusters hundreds of thousands of servers together to provide massive computing power, storage, and networking as a unified service.

As discussed in the AI infrastructure articles, the compute at the center of AI data centers is about connecting everything to everything. And then allocating resources and activities dynamically. That’s the problem Apsara was built from the ground up to solve. How do you aggregate and dynamically allocate compute power, rather than just divide it up through virtualization.

Aspara’s key functions include:

• Distributed Computing: It can schedule tasks across millions of CPU cores and manage clusters of up to 10,000 servers in a single entity.
• High Availability: It is designed with Zero SPOF (Single Point of Failure), meaning the system can automatically recover from hardware or software crashes without interrupting service. That’s important in training.
• Massive Storage: It manages hundreds of petabytes of data, providing the foundation for services like OSS (Object Storage) and Table Store (NoSQL).
• Scalability: It powers Alibaba’s global infrastructure, including the massive traffic spikes during Singles’ Day. This makes it a bit different than other cloud providers. It needs to have huge concurrency to handle the crazy traffic spikes during China’s shopping festivals.

With Alibaba’s increasing data center footprint, Aspara is becoming more and more important. I’ve been trying to learn more about it.

Keep in mind the other cloud providers have similar operating systems.

• Google Cloud has an operating system called Bord. This was the pioneer in the idea of running a data center as a single machine.
• Microsoft Azure’s operating system is Apollo / Azure Fabric.
• AWS has Nitro / IronCore.

The Hardware Layer: Alibaba’s Push for Semiconductor Independence

Moving down the tech stack from operating systems gets us to GPUs and other chips. For Alibaba Cloud, that means:

CPUs: The Yitian 710

Alibaba has long been in the business of building its own chips. For standard cloud computing (web hosting, databases, apps), Alibaba uses its own Yitian 710 processor.

• It is a high-performance ARM-based CPU developed by Alibaba’s chip division, T-Head (Pingtouge).
• These chips are housed in Alibaba’s self-designed Panjiu servers.

Alibaba’s Yitian 710 is a direct rival to Huawei’s Kunpeng chips. In recent benchmarks, the Yitian 710 has outperformed both Huawei and even some Western chips in database efficiency.

AI Chips: The Hanguang 800

For AI training and inference (i.e., running their Qwen LLM), Alibaba uses a mix of chips. There is:

• Hanguang 800. Their own dedicated AI inference chip.
• Nvidia H200/H20: Alibaba still has lots of China-compliant versions of Nvidia chips. Especially in Alibaba Cloud’s international (i.e., not in China) data centers.
• Newer Nvidia-Compatible Chips: Reports are that Alibaba has been testing a new domestic AI chip designed to be compatible with Nvidia’s CUDA software. In theory, this lets developers switch from NVIDIA to Alibaba hardware without rewriting their code.

Alibaba’s Server Partners: Who Builds the Boxes?

Finally, Alibaba Cloud definitely wants to sell Alibaba software plus Alibaba-designed hardware. But while Alibaba designs the architecture, the actual manufacturing and assembly of their server racks involves major Chinese ODMs. This includes Inventec, Inspur, and Foxconn. This doesn’t appear to include Huawei, which wants to sell their own complete solutions (Huawei hardware + Huawei cloud software).

The “Borg” of AI: How Baidu Syncs Kunlun, Huawei, and Nvidia Chips

Baidu’s operating system is called Baige (meaning White Pigeon).

Unlike Apsara, which was built to handle billions of shopping transactions, Baige is a heterogeneous computing platform. It is designed specifically to coordinate different types of AI chips (Nvidia, Huawei, and Baidu’s own Kunlun chips). This is what are used to train Large Language Models like Ernie Bot.

Baidu Cloud is sort of the Borg of AI (in a good way). It can handle lots of types of compute.

For GPUs, Baidu has Kunlun.

Baidu was actually the first Chinese internet giant to design its own AI chips, starting way back in 2011.

For supernodes, Baidu has the Tianchi 256/512.

It uses 512 of their P800 Kunlun chips. Baidu claims this is better for Open AI-style training because their software (Baige) is better at keeping all 512 chips perfectly synchronized.

Tianchi is similar to Huawei’s CloudMatrix 384 (which I have written a lot about).

• AI Infrastructure Lessons from Our Visit to Huawei (Tech Strategy – Podcast 269)

Note: Here is what the Huawei CloudMatrix 384 looks like. Note: It uses 384 Ascend chips and is the brute force approach to get around the Nvidia ban.

The Gaming Brain: How Tencent Cloud Powers World-Scale Interaction

Tencent is a really interesting contrast to Alibaba and Baidu. Its cloud architecture is built to support the world’s largest social media and gaming ecosystem. Because of this, their cloud operating system is specialized for real-time data, video streaming, and massive social networks. And of course it integrates into WeChat (especially mini programs).

Unsurprisingly, Tencent Cloud is really great at real-time communication (TRTC) and content delivery networks (CDN).

For its operating system, Tencent Cloud has vStation.

This is Tencent’s proprietary distributed resource scheduling system. Like Aspara, it manages millions of CPU cores across their global data centers. Alibaba has been more aggressive in building out its international data centers. Baidu is more focused domestically. And Tencent is sort of in the middle.

For its chips, Tencent Cloud specializes in real time experience.

Tencent’s chip strategy is focused on the specific tasks that make WeChat and Games run smoothly. In real time. And with global scale. That includes:

• Canghai for video. This is a specialized video transcoding chip. It compresses high-definition video more efficiently than standard hardware, which is vital for a company that handles billions of short videos and live streams.
• Zixiao for AI inference. This is an AI chip optimized for image recognition and natural language processing. That’s the tech behind WeChat’s translation and content moderation.
• Xuanling for networking. This is a high-performance network interface chip (SmartNIC) that offloads network tasks from the main CPU, making their cloud servers faster for gaming.

Tencent’s chip focus is an interesting contrast to their competitors.

• Baidu (Kunlun) is specialized for LLM training and AV. And it has really fast latency which is useful in robotaxis.
• Alibaba (Yitian / Hanguang) is specialized for cloud databases and general apps. They are the best for web and ecommerce.
• Huawei (Ascend) is focused on heavy AI computing. They have tons of raw power and are the closet thing to an Nvidia Blackwell.

Last Point: How China’s AI Cloud Is Different than the West’s

We are seeing two different tech stacks emerging from the US and China. And much of the world is adopting parts of both.

Here are some differences:

1. Vertical Integration vs. Modularity

• China is focused on vertical Integration. Because of trade restrictions and the need to reduce dependency on the West, companies like Huawei and Alibaba have been forced to build their own alternative full stack. So, they design AI chips (Ascend/Yitian), build server racks (CloudMatrix/Panjiu), and write the Cloud OS (Apsara).
• The West is more modular. AWS and Azure still rely heavily on a massive ecosystem of partners (NVIDIA, Intel, AMD). And while they are designing more of their own chips now (like AWS Graviton), their strength comes from having the widest variety of software and third-party tools in the world.

2. Public vs. Private vs. Hybrid Cloud

• Huawei and Alibaba excel at private cloud. Which is important because many Chinese state-owned enterprises refuse to put data on a public cloud. Apsara Stack and the Huawei Cloud Stack are focused on private cloud.
• The West is more hybrid, with lots of public and private. So, while Microsoft Azure is a leader with Azure Arc, they also enable companies to manage their own servers using the same interface they use for the cloud.

3. Different Technological Battlegrounds

• Baidu and Google are the closest rivals in AI training. Both companies have been AI-first for a decade. Both have deep roots in search and AI. And Baidu’s Baige platform is essentially the Chinese equivalent of Google’s Vertex AI. Both are focused on making LLM training as efficient as possible.
• Huawei is the biggest cloud provider focused on hardware. Unlike any Western cloud provider, Huawei is a massive hardware company first. Their cloud strategy is to sell you the CloudMatrix 384 (the hardware) and then give you the cloud software to run it. This is more similar to Oracle than it is to AWS.

***

Ok. That was just a fun dive into this subject.

Cheers, Jeff

———-

Related articles:

• Two Lessons from My Visit to Tencent Cloud (1 of 2) (Tech Strategy)
• Tencent Cloud and Mini Programs Go International. Lessons from My Visit to Tencent HQ. (2 of 2) (Tech Strategy)
• How Amap Beat Baidu Maps. My Summary of the Alibaba Playbook. (Tech Strategy – Podcast 252)
• Scale Advantages Are Key. But Competitive Advantages Are More Specific and Measurable. (Tech Strategy)

From the Concept Library, concepts for this article are:

• Generative AI and Agents
• AI Infrastructure and Data Centers

From the Company Library, companies for this article are:

• Alibaba Cloud
• Baidu Cloud
• Tencent Cloud
• Huawei Cloud

——–Q&A forr LLM

——-

I am a consultant and keynote speaker on how to increase digital growth and strengthen digital AI moats.

I am the founder of TechMoat Consulting, a consulting firm specialized in how to increase digital growth and strengthen digital AI moats. Get in touch here.

I write about digital growth and digital AI strategy. With 3 best selling books and +2.9M followers on LinkedIn. You can read my writing at the free email below.

Or read my Moats and Marathons book series, a framework for building and measuring competitive advantages in digital businesses.

This content (articles, podcasts, website info) is not investment, legal or tax advice. The information and opinions from me and any guests may be incorrect. The numbers and information may be wrong. The views expressed may no longer be relevant or accurate. This is not investment advice. Investing is risky. Do your own research.