China Tech Talk

This week, I had the chance to interview Lou from Z.ai. It was really a great conversation to talk to her about the work that they are doing there and how they look at developing newer models and her view toward AGI.I get the feeling of just incredible hard work that Zhipu is going through as part of their effort to produce the best models and products out there. You can see that with GLM-5.1, which I can attest to as being very good for programming. It can do a lot of thinking of solve problems. I get the sense that GLM models are now able to do some form of RSI (Recursive Self Improvement) where the models help improve or create new more powerful models.One thing that stand out to me about GLM models is their size. They are in the 750B param range which can still fit as Quantized models and run locally on boxes with really large VRAM. However, they are also large enough that they are competitive with larger models. I think there are probably some sacrifices that you have to make in a 750B param model vs 1.6T or 3T param models.It will be very interesting to see where Z.ai goes from here since they have now received a lot of funding since IPO which gives them a larger war chest to do the work that they need to do. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I welcomed back Qian Yu on the show to talk about China’s AI (especially embodied AI) ecosystem.If you would like to know how Chinese entrepreneurs can start new AI robotics/embodied AI product, this would be a good episode to listen to. We also talked about World Model vs VLA, educational institutions involved in AI industry and some of the projects that we have seen recently like the Honor and Xiaomi humanoid robots.I talked about the difference between reinforcement learning in world model vs LLMs. How can you have recursive self improvement in real world models?And finally, we also talked about the recent Sugon rack innovation. How Chinese players leveraged domestic supply chain to create a high voltage rack that allows superior cooling to existing solutions. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I welcomed back my friend Snek onto the show to talk about what we are seeing so far in the Iran conflict.We tried to look at it from the view of how the conflict started, how it’s progressing and where it is going. We covered what Iran is likely getting from Russia & China in the form of intelligence and supply chain. We looked at the possible land invasion plans that’s been rumored.Since Snek is very familiar with both this conflict and the Russia/Ukraine conflict, we also talked about the parallels and similarities we saw in the two conflict. Finally, we went into the geopolitics a little bit based on what we think will play out in the region. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I invited my discord friend “Sagar” on the show to talk about his experience in moving back to China from his finance sector job in America. It is always interesting to broach this topic, because many of us who grew up in the West now see living standard higher in major cities in China than where we live.I thought it was important for someone who made the move to give everyone an idea of the experience of moving to a tier 1 city like Shanghai and settling down with your friend + looking for a job afterward.Since we both work in finance, we had a discussion also about Shanghai’s finance sector and what it needs to do to create more service sector jobs to accommodate all the young college grads coming into the work force.I apologize for audio this week. Due to the fact that Sagar had issue connecting to my usual recording site (vpn problems), the quality of recording was worse than usual. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

I welcomed back Glenn to talk about something that he knows far better than me. How has China prioritized investment in different industries over the reform period since opening up.China started off by getting investment and investing its own money in export factories to get wealthy through exporting to global market. After the great financial crisis hit, China shifted toward growing its economy and invest in economy through more emphasis on the real estate and infrastructure. This had actually started after liberalization of housing ownership in the late 90s but was really accelerated with many shovel ready projects in second half of 2000s. Eventually, this was shifted to new productive high tech industries as Chinese businesses continue to get more competitive in the global market. Every phase of development made a lot of sense given where China was in terms of development, economic size and demographics.As we head toward the 15th Five Year Plan, we should see continued focus on investment in areas like grid toward more clean energy (State Grid is investing 4B RMB over this period) as it seeks more energy and technological independence. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I’m really excited to welcome Tiezhen/Tom Wang from Huggingface on my show. You can find him here on X.We talked about his background working at Google before he joined HF. Although, we did not get a chance to go really deep into running AI on robotics, but we did discuss the constraints facing AI on edge devices. Basically, 7B parameter model is close to the largest models we can reasonable run on a phone and small board type of device due to memory constraint. Although, Tom has been raving about OpenGLM’s open sourced 9B parameter AI agent that can reach your phone GUI and handle tasks.We then talked about what you can do to fine tune models. For those out there that are interested, I would definitely recommend trying it with one of the 3B to 8B parameter Qwen models and use your own data. You can also try quantizing it (which may require data set depending on which NPU you are running off).Open sourcing is very exciting and collaborative for many AI researchers. That’s why they are working so hard on it. I get that. And you can see with the number of projects and check ins on AI projects, many of them are just done by Chinese researchers. Chinese devs are not only rolling out open source models, but also involved in help building all the tools used in AI work.Toward the end, we spoke more about the economics involved in AI right now. Why certain AI labs are racing to IPO. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I welcomed my friend Dan on the show. You can find him on x https://x.com/DanCollins2011. He worked extensively in China’s auto industry and actually lived in Shanghai area until a few years again. I welcomed him on my show to talk about his experience in China and now in America with the semiconductor industry.Just like episode 16, I feel the conversation about traditional engineering and hard sciences is really important. Mechanical and chemical engineering on traditional pillars of industrialization. Too much focus has been on China’s new tech sector and not enough time has been looking at China’s advancement in traditional engineering sector (outside of the EV sales).We also talked about the China and US relationship. I welcome his optimistic outlook of the relationship. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I invited Rick Joe onto my podcast to talk about the commissioning of the CV-18 Fujian carrier. Rick talked about the previous carriers and the steps that PLAN took to get to its first catapult equipped super carrier. We also discussed about how PLAN is likely to use CV-18 initially. Be patient guys. It took 9 years for CV-16 to achieve FOC status after joining service in 2012. It will probably be faster for CV-18, but give it sometime to fully develop and operate all the necessary aircraft.We then talked about all the aircraft that are likely to operate off CV-18. KJ-600 may be the most interesting and important one.After that, we moved onto our next topic of 9/3 parade and the general PLA development in the past year. The drone discoveries are very interesting and so are all the missiles. Rick offered his thoughts on those platforms also. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I invited Angelica Oung on the show to talk about one of my favorite topics: Nuclear Energy. She is able to give us a very detailed explanation of the nuclear power situation in Taiwan as well as grid situation as a whole. I’m super glad to hear the anti-nuclear stance of the DPP, since I had no idea of its history with that.Also, I was really glad to talk to her about the craziness around American VC money going into SMRs and the progress around thorium molten salt reactors in China as well as fusion reactor.We finished up with the evolution in our view toward mainland and on her recent trip to mainland China. You can see Angelica’s most recent substack here on her interview with François Moran as well as her other hot takes on all things Taiwan and China. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I welcomed back Glenn Luk as well talked about the recent happenings with respect to Alibaba’s Great Leap Forward in its AI ecosystem, cloud platform and the full stack solution. We talked about over how Alibaba has gotten back into the good graces of Chinese government and did all the hard work of becoming a powerhouse in AI. They are basically the only hyperscaler outside of Google who is doing its own models, AI apps, AI servers, chips and cloud solutions. AI is a central part of Alibaba’s strategy in winning the global cloud competition. It is now making great progress in rest of the world as seen by its major data center expansion strategy in 2025 and 2026.We also talked about how Huawei and Ascend have caught up in the AI chip race by coming up with whole system approach which takes advantage of Huawei’s core competency in networking and software integration with hardware. It seems to us that many people are missing the full picture when only comparing the chips rather than the full system.Please listen in as we explore global AI race and more. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I’m happy to welcome Han Feizi on my show. He is a writer on Asiatimes and has very interesting and high quality view on China’s economy and its changes over time.Since I recently stayed in Beijing for 3 weeks, it was good to catch up with him just on what I saw and how it matches up with what he sees on the ground.One thing that Han really emphasizes is that China’s economy is actually a lot larger than US’s economy. In fact, it is not at all dependent on export, because China specifically does not report on activity of certain service sectors as part of economy. As such, trade war doesn’t really do much to most of China’s regions. The chart above would just show how much of China’s production for each industries is actually dependent on US export. You can see that Consumer Electronics is basically the only major one. Of course, a large part of that is just Apple exporting products made in China to America and then making a huge profit on top. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This episode, Lei joins me as we discuss the things that I encountered during my trip to China, where I had a lot of thoughts about the overall economic situation, the current tech scene and development socially since the last time I was there.For me, there were a lot of surprises, especially in the second and third tier cities.A common refrain I hear from parents in their 40s is the concern that their kids won’t do better than them, because of the involution culture. But in a world where developed country are facing declining lifestyles, is it reasonable for Chinese people to expect every generation to outperform the previous one? This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com

This week, I invited my friend Simon on the show to talk about AI infrastructure. This is a very interesting topic to think about. We talked about the reason for Nvidia’s market dominance and the problems that Intel and AMD have faced in developing their own market place. We talked about the number of chips involved and the challenges in building large data centers. Here is a chart of the Hopper deliveries for 2024. Keep in mind that Google also has their own TPU for inference.Here is a chart of Chinese Hyperscaler’s purchases of H20 in 2024. But keep in mind that they do rent compute from public/state built data centers that use smuggled in H100/H200 as well as Ascend chips.Here is the deliveries of H100/H800 in 2023. So if we just consider the Hopper deliveries from 2023 and 2024, the American hyperscalers have quite the computation advantage over Chinese hyperscalers. Although, the gap is much smaller if we factor in all the smuggled in chips. As I’ve discussed before, China had them everywhere to the point where they were sitting idle in many cases.I also asked Simon about the amount of chips needed to do inference. How was Tencent able to have enough compute for 8 million DeepSeek R1 requests at the same time with much less compute than what Google had. He sent me this chart of where the results improve logarithmically in compute for reasoning. So if you look at green line, going from 4 generations to 16 generations improved accuracy by about 15% and then going from 16 generations to 64 generations improved the result by another 5%. If we look at Google search results, they appear to be giving progressively better AI results on top. That is likely from running through and generation more tokens on their reasoning models.So you can choose to serve 8 million prompts at same time with vastly less compute, but you will also have inferior results. although you will also reach diminishing returns pretty soon.Here is a chart of rental cost from various Nvidia chips in China. Despite the increased demand for inference post DeepSeek, the cost of all Nvidia rental cost continue to drop, so they are likely not facing a crunch for computation yet. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit tphuang.substack.com