Alibaba

3 Lessons from My Interview with Dongliang Guo, Head of International at Alibaba Cloud (Tech Strategy)

Posted on by

I was recently at the new Alibaba Xixi campus in Hangzhou. And I had the opportunity to sit down with Dongliang Guo, who is VP of International Business for Alibaba Cloud.

This was a big deal (for me at least). There are few companies that are going to be as instrumental in the future of China and Asia as Alibaba Cloud.

It is already arguably #1 for infrastructure as a service (IaaS) in the region. And it is the platform on which new GenAI tools are being created and adopted across the region. Everyone in the West talks about Google Cloud, AWS and Azure. But in Asia, it’s Alibaba Cloud (and a few others).

Plus, Dongliang is an interesting guy. And fun to chat with.

A quick clarification. Dongliang’s group, the international business of Alibaba Cloud, includes the non-China markets. So that’s Southeast Asia, Japan, and South Korea (all very important). And also, some select international markets, such as the Middle East (compelling), Europe and Africa. But this business unit also covers multinationals operating in China, which are in an interesting situation. They must reconcile their China businesses with international operations and standards.

That said. Here are my three take-aways / lessons from our discussion.

Take-Away 1: Alibaba Cloud is Focused on the Infrastructure and Model Layers. It Is Differentiating its GenAI with an Open-Source Ecosystem (i.e., ModelScope) and Model Building Tools.

In September 2024, Alibaba Cloud released Qwen 2.5, its updated suited of large language models. And they got pretty great reviews. In particular, Qwen Coder is getting a lot of attention. Developers are paying close attention to what Alibaba Cloud and their updates are coming rapidly.

For GenAI, Alibaba Cloud is definitely focused on the infrastructure and model layers. And not as much at the tooling and app layers (yet). And there is an interesting interplay between foundation models and infrastructure.

Pre-GenAI, cloud-based infrastructure was about offering compute and storage services with greater scalability, elasticity, and efficiency. That was the standard pitch for cloud services versus on premise computing.

However, when you switch to GenAI, you start thinking about training and inference. When training foundation models, the focus is cost (i.e., efficiency) and stability. The stability part is important and Alibaba Cloud appears to be emphasizing this.

Basically, training models is expensive. And it takes longer and becomes more expensive if there isn’t stability in the infrastructure. If your system is failing 8-9x a day while training, you have to spend a lot longer in the training phase. So, Alibaba Cloud talks a lot about its stability (hardware and software). And it provides monitoring and responsiveness services for this phase.

But increasingly the focus in on training and inference.

And Alibaba Cloud emphasizes how its clients can allocate resources flexibly between training and inference. For the training phase, the scope and requirements are fairly predictable. But in the inference phase, things can be pretty unpredictable. How much compute is required for ongoing performance?

So, Alibaba Cloud is integrating the two activities and giving clients’ flexibility between them. You can allocate resource as needed between them.

All of this is interesting. And I’ve been keeping an eye on the adoption of Alibaba Cloud’s infrastructure and models by both clients and developers. There’s a lot of hype in this sector. But I can tell a winner when I see one. And I like the adoption.

But I like their strategic positioning more.

The key differentiating factor for Alibaba Cloud (in GenAI) is that it is committed to building an open-source ecosystem. Alibaba Cloud does have some closed-source LLMs, but there has been a big strategy decision to go open source.

As mentioned in September, Alibaba Cloud announced open sourcing its Qwen2.5 model series. In addition, they made a strategic choice to offer an open-source platform, called ModelScope. That is very different than what we see at OpenAI and Microsoft Azure. I’m a big believer in open source as the foundation. It’s why Android is just a lot more widely used than iOS.

This ties into their second differentiating factor, which is Model Studio, a generative AI development platform. This is where clients and such can build and customize on their foundation models. They are moving fast here.

That’s how I view Alibaba Cloud right now.

  • A lot of focus on infrastructure and models.
  • Solid differentiation in these as services.

Ok. Next take-away.

Take-Away 2: Southeast Asia Cloud Computing is Being “Fixed”. And Also Reshaped for GenAI.

Southeast Asia is clearly a big focus for Alibaba Cloud internationally. I asked about their approach to the region and Dongliang used the phrase “reshaping cloud computing”. That’s pretty interesting. I would call it fixing and reshaping.

Cloud computing is not new to SE Asia as a service. There are data centers everywhere. And this has historically been CPU-based computing that cloud services made scalable, flexible, and efficient. That is standard.

However, a lot of the cloud clients in the region never really got these benefits. A lot did not get the cost benefits. Maybe because staff weren’t let go from their on-premise activities. Maybe because cloud resources weren’t decreased with decreased volumes.

I asked Dongliang about how Alibaba Cloud is different in Southeast Asia and he talked a lot about working with clients to basically fix their cloud practices in the region. So, the companies got the benefits of cloud. “Fixing” has been a sizeable opportunity for cloud in Southeast Asia. And this has a lot to do with bringing best practices from China.

There is also a lot of “reshaping” of the cloud and tech architecture. GenAI requires a different tech stack. There’s no way around it. It requires a big shift from CPU to GPU-based compute. It requires adding the model layer, which is completely new. And there is a massive increase in the amount of data and throughput.

So “reshaping” is the other big opportunity in the region. And this is pretty common in most of the world at the moment.  The fixing situation is more of a developing economy problem.

Take-Away 3: GenAI Has Surprising High Impact Use Cases Everywhere.

I asked Dongliang about GenAI adoption and he surprisingly mentioned gaming as the first adopter. That was interesting. Specifically, he mentioned game design as an interesting use case and first mover.

In the development of video games, the major characters are initially developed by artists in the master design. That’s where a lot of the creativity is. But then these characters need to be extended into lots of different versions (different clothes, situation, weapons, etc.). Well, it turns out GenAI is really good at this step.

Dongliang also mentioned some interesting, high value use cases in pharmaceuticals and finance / banking. Such as using GenAI to create regulatory filings and other required documentation in pharmaceutical research and operations. These require highly trained professionals and LLMs can really increase their productivity.

The banking examples were similar. You get a lot of impact when you deploy GenAI productivity tools into document intensive activities that require advanced expertise.

However, the biggest adopters were ecommerce companies and other internet companies in the region. Not a big surprise. They are digital natives so adopting new tools is quite straight forward.

For them, I found the customer facing use cases the most interesting. He mentioned the use of shopping assistants by ecommerce companies. I have looked into this previously at Lazada (LazzieChat) and JD. These can potentially have the most direct impact on revenue and customer satisfaction. But they also have risks (they can annoy customers).

Other important ecommerce use cases are in tools for merchants and brands. Especially in analyzing customers and creating content. Creating content with GenAI at large scale lets you personalize communications. That can be a powerful move. It can also enable you create product listings more efficiently.

Dongliang mentioned a really interesting example of this in industrial product SKUs. The example he mentioned was screws, which come in a huge number of types and sizes. And online there is not much information on each individual screw available. It would be ridiculous to gather and input lots of information for each type of screw (its size, composition, etc.).

However, GenAI can do this based on an image of a screw. It can complete all the product information for each individual screw. And this has some interesting immediate benefits. First, the information is now in the listings so that improves search results. Second, it improves recommendations. And third it improves reports and intelligence.

Finally, Dongliang mentioned a use case that I have been thinking a lot about. That is about how management gathers information and insights from customer conversations and other types of unstructured data.

Typically, managers make decisions using a combination of reports and their expertise / experience. Reports increasingly come from management dashboards, which is mostly from structured data (data warehouses not data lakes). They look at these. Combine this with their own expertise and make the decision.

However, this is sometimes not enough to make a decision or to understand what is really going on. In these cases, management will often call others to discuss. Or ask for additional information. Or talk to customers. Or maybe even rebuild and add to the dashboard.

The problem is that a lot of key information is in unstructured data. It is in conversations with customers. In conversations with staff about operations. There is a lot of intelligence in conversations and exchanges. And it turns out GenAI is really good at this type of unstructured data. It can summarize tons of discussions with customers (by phone calls, emails, chatbots, etc.). This type of unstructured data is a great source of information. And it can complement management expertise and their dashboards. As Dongliang said, there are lots of insights in conversations. And the truth is often hidden in unstructured data.

***

So those are my three take-aways. It was a pretty great conversation.

Last Point: Why Chinese is Better for NLP. And Why Alibaba Cloud is Different.

I asked how Alibaba Cloud differentiates itself from AWS and Google Cloud in Asia. Dongliang mentioned two ways.

First, they bring best practices from China. This was my earlier point about using best practices from China to capture the benefits of cloud service.

Second, he mentioned that Alibaba Cloud is better at GenAI for Asian languages.

And that was really interesting. Because I had heard that Alibaba Cloud (based on Chinese) was better for other Asian languages. But I never really understood why. Dongliang explained it to me.

First, businesses in Southeast Asia want to operate in their own languages. Businesses in Thailand operate in Thai. In Vietnam, in Vietnamese. Staff want to speak their own language. Not English. That is what their businesses units operate in. That is what their IT staff operate in.

And it turns out Alibaba Cloud works better when being adapted to other Asian languages. Natural language processing (NLP) basically converts written words into tokens. And this happens more efficiently in Chinese than in English. For example, English words are made up of lots of letters and then spaces between each word. That is a certain amount of information in a certain number of characters. However, Chinese characters each represent one word, with no spaces in between. There is a lot more information in each Chinese character.

And then you convert these into token. One and a half Chinese characters can be one token. And only 0.7 English words can be one token. Basically, in Chinese more information can be captured in one token. Dongliang mentioned that in English, 100 words typically results in about 143 tokens. But in Chinese, 100 characters results in about 67 tokens. So, you need fewer tokens for the same information, which is cheaper.

And Korean, Japanese, and other Asian languages are similar to Chinese. So, when you build a Japanese language model based on Qwen you are much more efficient. And Alibaba Cloud has been highlighting examples of Japanese teams creating Japanese foundation models based on Qwen.

Anyways, I thought that was pretty cool.

***

Ok. That’s it for this article. Really fun stuff.

Cheers, Jeff

———

Related articles:

From the Concept Library, concepts for this article are:

  • Generative AI
  • Cloud Services

From the Company Library, companies for this article are:

  • Alibaba Cloud

———

I write, speak and consult about how to win (and not lose) in digital strategy and transformation.

I am the founder of TechMoat Consulting, a boutique consulting firm that helps retailers, brands, and technology companies exploit digital change to grow faster, innovate better and build digital moats. Get in touch here.

My book series Moats and Marathons is one-of-a-kind 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.

Leave a Reply