In the second hall, there were many booths and not just the ones that were indie or hardware related. There were also middleware companies and many other services, such as vendors proposing AI models to speed up asset creation and more stuff in that field. As I was walking around, what caught my attention very quickly was the Alibaba Cloud booth. Everyone knows the company that is behind AliExpress, but maybe some of you are not aware that they also are an AI model powerhouse, with very effective teams developing models under the Qwen and Wan brand names. Those are very advanced models and they also happen to be frequently released as open weights, so that you can run them on your own hardware if you have what it takes.

I had a conversation with one of the employees at the booth who could tell me a bit more about what’s happening in more details.

Let me just preface this by saying that is all hearsay and that I have no proof of everything I have heard. But at the same time, the person I talked to seemed to speak honestly so I don’t think the following was just some of exagerated bullshit.

Qwen AI Models

First, the team(s) behind the Qwen models are apparently in the hundreds of engineers (most of them probably AI specialists responsible in one way or another of the training preparation, the testing, and the deployment of such models). Because they had issues with ByteDance (the company behind TikTok) poaching employees from Alibaba, now whoever is working on AI models and what they are actually doing in such capacity has become a major secret business, and even Alibaba employees themselves cannot get this information from their internal employee listings. The competition is ruthless!

When you want to develop AI models, one of the biggest problem to solve is access to large amounts of compute. Alibaba has access to Nvidia processors, not the H100 or H200 which are not allowed to be sold to China at the moment, but they can still get their hands on H20 GPUs, and some other models while it lasts.

Recently China has announced that they will be blocking imports of Nvidia GPUs altogether. In any case, the person at the booth could tell me that Alibaba has already developed and rolled out their own compute chip internally (which seems to be running on a 7nm lithography), that they can use to replace Nvidia (at least for inference). Note this is recently corroborated by other sources, such as this article from SCMP. For training, Nvidia GPUs are still better at this stage.

When I asked about the important topic of CUDA and PyTorch compatibility, he mentioned that they already have developed their own CUDA-like support, that is not 100% compatible, but works well enough for their internal uses at Alibaba.

I wondered if they were going to enter the market for GPU or compute cards since they have such capabilities internally, and it seems like they have no such intent. In any case, China is catching up and it may very well be that in a few years they mostly don’t need Nvidia anymore to develop, train and run their own AI models. That would be the first country to be able to be independent from Nvidia in that regard.

As we were talking he was showing me the preview version of WAN 2.5 (that had just came out the previous day!) and that is already able to generate videos up to 10 seconds, with sound at the same time!

You can try WAN 2.5 online directly on the Alibaba site (and avoid the hundreds of spammy sites that pretend to give you free access…)

Video generation is already very impressive compared to what we had 1 year ago. But it is still very limited, generating anything beyond a few seconds to make some kind of movie is just very hard to produce at this stage. But… 2 years, 5 years from now? Who knows.

Cloud Streaming

Alibaba also offers some game streaming capabilities. Basically game streaming as a service, where you don’t have to rely on the usual vendors, but use the existing Alibaba infrastructure instead (and their GPUs if needed). It seems that their stack runs on Linux. Their key competitive advantage was apparently fairly low latency (I could see a demo of a 2D game running off that service) and reduced cost compared to other services.

While they could not tell the actual name of that company, there is apparently a fairly big Japanese company which is providing such game streaming services in Japan, leveraging the Alibaba Cloud infrastructure. I would not be surprised since there are many players in that space.

They have a whole array of services aimed at the Gaming category, such as content distribution, game server management, and streaming as we have just alluded to.

However, my counterpart at the booth mentioned that game streaming is not as popular of a concept as it was a few years ago, so it does not make headlines as much these days.

China Rising

This is going to be a recurring theme. China was everywhere at the TGS 2025. Usually it was about outsourcing, but more and more they are becoming key players in many areas. In AI models, in gaming, and soon new challengers in the semi-conductor business. TSMC is still ahead, but the question is for how long… In terms of infra and services, Alibaba Cloud is probably a minor player worldwide compared to AWS, but AWS is so expensive and so predominant that this whole business needs a lot more competition.