Dowson Tong, CEO of Tencent Cloud and Smart Industries Group in an exclusive conversation with BTV's Stephen Engle talks about the company's investment shifts from being AI consumer agents to enterprise, why it is prioritizing compute for internal model training and products and the competition on chatbot space.
The focus right now on the build out of AI globally, not just China, is about profitability and the amount of CAPEX that is going into building out the data centers and there's some cracks forming and concern on the sustainability and when there will be ROI, return on investment uh in this uh, you know, CAPEX spree. How long from your vantage point, do you see this in China and globally, this elevated levels of capital raising through bond issuances which Tencent has done as well as spending to build out the offerings.
Mhm. Yes, the uh CAPEX has been uh pretty sizable uh for pretty much most players. Um, but the demand is also very strong. I think uh right now uh if you look at the five-layer stack as uh Jensen uh like to describe, majority of the profits shift downward, right? So, if we can um uh get the same compute at a lower cost and maybe the equation would work uh just fine. Especially uh uh uh as we see uh agents being deployed in corporate environments, as well as uh uh you know, consumers using AI more and more in their daily life. Uh we do believe um the value created by AI uh is tremendous. Um so uh I'm sure we'll figure uh figure out a way to make the economics work.
Do you feel though that Tencent is sort of playing catchup to some of the other larger players, whether it's ByteDance or Alibaba? I mean, Pony Ma has talked about it as well as we're on board the ship right now, but maybe the ship has been a bit leaky. Where do you think you've been leaky in the development? Because there you've also talked about you maybe in the early days did not prioritize the hiring of top-level engineers and the like. Where are you right now? How would you assess how you're trying to capitalize on the AI buildout?
Well, um as I mentioned, there are multiple layers that we play uh in the stack. So, um I think we were uh pretty strong at the infra level. Um there was a time I think uh uh uh because of the lack of compute, our uh we didn't allocate enough for uh model training of Hunyuan. Uh and also there were some uh team dynamics that we were trying to sort out um I would say uh uh two years, three years ago. So, um uh but this year, uh uh in fact, we're seeing a lot of improvement uh with the Hunyuan model, as you uh might see in uh some of the benchmarks uh out there. Uh Hunyuan 4 uh preview has uh captured a lot of um um uh interest and also being used quite extensively. Uh compared to uh roughly half a year ago when Pony uh made the uh statement of a leaky ship, I think we've come a long way.
And uh, How so? How so? He he talked about the ship being leaky.
Where was it leaky and where have you fixed those, where have you patched the holes?
Uh, well, when I talked about the models, uh uh I think uh Hunyuan models have come a long way uh since uh the beginning of the year.
This is the multimodal family of offerings, right? It's like Yuanbao runs on Hunyuan. And you're having Hunyuan 4 coming out by the end of this year, I believe.
So, I think we need to look at individual layers uh differently. So I talked about the model part uh earlier, but if you look at the application part, uh definitely we invested quite a bit in Yuanbao uh especially uh last year when we uh start to uh ramp up uh the user base. Um but unfortunately, I think the uh the consumer um uh uh usage of uh uh the agentic products or chatbot-based um uh products uh uh is quite challenging uh with a lot of competitors, uh throwing a lot of uh uh money in it and I think uh uh in general, we're seeing uh some uh shift of investment over to uh agents uh in the enterprise space, such as Work Buddy.
Right.
Um so that's uh where we have been investing uh this year, uh starting uh in March, I would say. Um and after that, I think we see a lot of uh uh user adoption um with the latest uh models, LLMs, uh from the open weight uh model players, uh in combination with uh Work Buddy, uh integrated with um uh document editing capabilities, uh browser use capabilities, and uh uh uh uh and we're able to build um a much user, much more user-friendly uh agentic products uh for our users. So, I think uh we're seeing um uh different sub-segments playing uh in different directions uh in the uh AI application space.
