The Artificial Investor - Issue 38: The current status of the AI Cold War
My name is Aris Xenofontos and I am an investor at Seaya Ventures. This is the weekly version of the Artificial Investor that covers the top AI developments of the last seven days.
This Week’s Story: Chinese AI models surpass the most advanced US counterparts
After a week break, we are back with the story about two Chinese AI models, startup’s DeepSeek R1 and Alibaba’s QwQ surpassing the strongest Western model, OpenAI’s o1, in reasoning capabilities. This came as Pony AI, one of China’s leading autonomous-driving companies, debuted in the Nasdaq with a 400-million-dollar IPO that valued the company at 4.6 billion dollars.
This brings us back to the US/China AI Cold War discussion. What country is the most innovative one? Who is the strongest across the different segments of the market? How is this expected to evolve?
If you prefer to listen to this issue, click here for the audio version.
Powered by ElevenLabs
✍️ The Chinese government has been funding AI startups and infrastructure projects through public funding initiatives. The US has responded with bans on exports of chips and other parts of the AI value chain to China, and Trump is waiting to take over and escalate things further with special import tariffs.
How are the two poles of the global AI scene shaping up? Let’s break it down by layer of the Tech stack.
⛏️ AI Infrastructure Layer: Raw materials
Before we go into the hardware layer, let’s have a look at who controls the mining and supply chain of the required raw materials. After all, AI chips are made of something. This something is in short: silicon, dopants and additives (gallium, germanium, arsenic, antimony), metals (copper, aluminum, gold, tin and palladium) and some other materials (tantalum, tungsten, etc).
China has a leading position in sourcing these relevant materials. Specifically, the country:
is the world’s largest silicon producer, accounting for approximately 70% of total global production
produces the vast majority of the main dopants and additives globally, specifically 98% of gallium and 48% of antimony
is the world’s largest aluminum producer, accounting for approximately 59% of global production
covers its copper supply needs with domestic production (c.8% of global volumes)
The US has a strong production of copper only (albeit still 30% below China). In terms of the remaining materials, the United States produces no gallium, less than 2% of the world's refined germanium, relies heavily on aluminum imports and produces annually 5% of the volume of silicon that China does.
This dependency on Chinese-supplied minerals poses significant challenges for the US in the AI value chain.
💾 AI Infrastructure Layer: Hardware
Within the hardware side of the AI Infrastructure Layer we find two main sub-layers: i) pure-play manufacturers (also called Foundries) and ii) end-to-end manufacturers. China seems to lead the former, while the US the latter.
In the Foundry market, China is clearly ahead of the US. In the global top 10 we find three Chinese companies (SMIC, HuaHong and Nexchip) with an aggregate market share of 9%, while on top we find TSMC with 64% of the market, a company from Taiwan, which arguably could be considered within the sphere of influence of China. There is only one American company in the top 10, GlobalFoundries, with 5% of the market. This compares to an aggregate c.80% share for Chinese and Taiwanese companies.
On the end-to-end manufacturing side, the situation is the exact opposite. In the broader semiconductor market, which includes all types of computer chips and end clients, Nvidia owns half of the market, while no Chinese company has made it to the top 15. Zooming into GPUs, the main chips used in AI use cases, Nvidia is the undisputed leader with 80%+ market share. Even within China, Nvidia has a 90% market share (the US company has 3,000 employees based in China after all, or 10% of its workforce), followed by China’s leader, Huawei (only 6% of the market). These are figures that are expected to change in 2024 as US’s chip export restrictions kick in; we shall know their exact impact in the coming months.
However, it’s not just about size. Nvidia is technically superior to its Chinese competitors and with stronger entry barriers. The American leader has the most advanced manufacturing capabilities, as well as a developed software ecosystem that facilitates the implementation and use of its chips at scale. Nvidia released in Q1’23 its flagship AI chip, the H100. The H100 powers X.AI’ recently-built 100,000-chip AI cluster and has a performance of 19.5 PetaFLOPS consuming 700W. Huawei has just launched the most advanced Chinese chip, Ascend 910C, which performs overall worse than the H100 (9.7 PetaFLOPS at 380W) that was released nearly two years ago. This demonstrates that Chinese semiconductors are currently inferior to Western ones and the gap is expected to remain as Nvidia is preparing the launch of its new Blackwell series early 2025.
⚙️ AI Infrastructure Layer: Models
There are four main types of last-generation AI models: closed-source large language models (LLMs), open-source LLMs, large reasoning models (LRMs) and vision-and-large-language models (VLLMs). US companies lead across all fronts, but are very closely followed by the Chinese ones.
The four leading US LLM developers are OpenAI (ChatGPT-4, o1), Meta (Llama 3.1), Google (Gemini) and Anthropic (Claude). In aggregate, the two scale-ups, OpenAI and Anthropic, have collectively raised 25.5 billion dollars of funding and are expected to generate c.5 billion dollars of revenues in 2024. On the other side, we have China’s five leading AI companies, also called the “New AI Tigers”: Zhipu AI, Baichuan AI, Moonshot AI, Minimax and 01.A; as well as an up-and-coming player, DeepSeek. There is no information in terms of the revenues they generate and funding information is incomplete, however, we estimate that the six Chinese companies have raised 3.5 -4 billion dollars in funding, which is one sixth of what their American counterparts have raised.
In terms of model performance, once again we shall say that the published benchmarks are not a perfect way to compare models. Nevertheless, the leader among closed-source LLMs seems to remain OpenAI’s GPT-4, closely followed by the leading Chinese closed-source model, 01.AI’s Yi-Large. Yi-Large has a trillion parameters (compared to an estimated 1.5 trillion for GPT-4), it ranks close to GPT-4 in English-language rankings and surpasses it in Chinese-language ones, and costs only a third of its American competitor. While Chinese models are staying close to the American ones, training data is a key driver of model performance, which is where the US players have an advantage. Globally c.270 million terabytes of data is generated in the Web every day outside China, which is twice the amount of data generated in China daily (c.130 million terabytes).
On the open-source LLM side, depending on the benchmark you look at (IFEval, GPQA or MMLU), Meta’s Llama 3.1 or Mistral’s Large come on the top of the list. Again, a Chinese model stands very close behind them, Alibaba’s Qwen 2.5 72B. The benchmark scores are so close that most humans would not recognise the difference between them.
On the reasoning model side, OpenAI released o1 earlier this year demonstrating to the world that a large Chain-of-Thought (CoT) model can beat humans in reasoning tasks and open up a new model scaling dimension: reasoning iterations during inference. Nevertheless, in November 2024 two Chinese reasoning models were launched, Alibaba’s QwQ and DeepSeek’s R1, which beat o1 in various math benchmarks (you can try DeepSeek R1 here using the “Deep Think” switch, it’s great).
Video generation models are the most sophisticated types of VLLMs. OpenAI previewed Sora in February this year and wowed the world with its capability to generate realistic 10-second videos using just text. Nevertheless, 10 months later Sora is still not released to the public, which has opened up the way to US-based AI challengers, such as Runway, Pika Labs and Luma. On the other side, two emerging video-generation technologies have appeared in China, Kling AI and Hailuo (by Minimax). Kling offers high-resolution outputs (you can try it here) and Hailuo is the most affordable option in the market starting at 10 dollars per month. It seems that we are in a tie here.
☁️ AI Infrastructure Layer: Cloud
On the Cloud Infrastructure side, the US has the clear lead. Amazon AWS, Microsoft Azure and Google GCP collectively generated 63 billion dollars of revenues in Q3 '24 positioning them on the top 3 of the global Cloud market with an aggregate market share of c.70%. The top 3 Chinese Cloud providers are Alibaba Cloud, Huawei Cloud and Tencent Cloud. While they collectively own c.70% of the Chinese market, they represent only 8% of the Cloud market worldwide.
Overall, the US is home to c.2,500 data centers, while China to c.450, less than a fifth. One important consideration here is the expected AI data center bottleneck, energy. While the US leads in current installed nuclear capacity (c.9 GW vs. 53 GW for China), the Asian giant is expanding its nuclear capacity faster. China has 22 reactors under construction (24 GW) and 70+ planned (88 GW), while the US nuclear fleet is aging, with most reactors built between 1967 and 1990.
🖥️ AI Application Layer
The two most advanced AI applications currently are i) LLM chatbots and ii) autonomous vehicles, and the US seems to have the lead in both.
The three main US AI chatbots are ChatGPT (OpenAI), Gemini (Google) and Meta AI. Their estimated aggregate userbase is 830 million monthly active users and over 1 million corporates (although there is a large overlap among the three userbases). The three main Chinese AI chatbots are Ernie Bot (Baidu), Doubao (ByteDance) and Tongyi Qianwen (Alibaba). Their estimated aggregate userbase is 250 million monthly active users, i.e. less than one third of the American chatbot users, and 2.2 million corporates, which is double the amount of businesses using American chatbots. Overall, 9,500 US AI companies have raised a total of 600 billion dollars in the last 10 years, while less than 2,000 Chinese AI companies have raised just over 85 billion dollars in the same period.
We wrote about the acceleration of the robotaxi wave in the 34th issue of the Artificial Investor. The main US players are Waymo, Zoox and Cruise (with Tesla joining them soon), but Waymo is the only one that has achieved real scale. The robotaxi operator is currently providing about 700,000 paid rides per month. The three American players have collectively raised about 25 billion dollars by large companies, such as Google, Amazon, General Motors and Microsoft. The main Chinese players are Apollo Go (Baidu), Pony AI and WeRide, with Apollo Go being the one with the largest scale. The robotaxi operator is currently providing about 275,000 monthly paid rides, which is just over a third of Waymo’s volumes. The three Chinese players have collectively raised about 3-4 billion dollars, which is about a sixth of the amount of their US counterparts. This includes the about 800 million dollars raised in the Nasdaq IPOs of WeRide (October 2024) and Pony AI (November 2024), which have clearly been trying to take advantage of the US investor appetite to fund the robotaxi trend.
🧪 Reseach & Intellectual Property
While the current success of AI chatbots and robotaxis may be lagging indicators of AI competitiveness, the quality of research and the volumes of granted patents may be more of leading ones. China is winning this battle overall.
China is the global leader in AI research publications and is neck and neck with the United States on Gen AI. However, China’s research publications have less impact than U.S. ones, with fewer citations and less private-sector involvement.
China dominates the top 20 patent owners in Gen AI, accounting for 13 of them and four of the top five (the fifth is IBM). Corporations dominate the entire list, but there are three research organizations in the top 20 and they are all Chinese (the Chinese Academy of Sciences, Tsinghua University and Zhejiang University). The most prolific of these research universities, the Chinese Academy of Sciences, ranks fourth overall. Critics claim that these figures are distorted by the fact that i) China has a broader scope of what AI innovations are eligible to be patented, and ii) patents issued by the Chinese Patent Office are of relatively poor quality (measured in terms of percentage of patents also filed in another jurisdiction, which is 4% for Chinese patents vs. 32% for US ones).
🔮 Looking ahead
Wrapping up, the US seems to be winning the AI Cold War currently.
America is home to the number one AI semiconductor company (Nvidia), the top three Cloud providers (Amazon AWS, Microsoft Azure and Google GCP), as well as the leading AI chatbots (ChatGPT, Gemini, Meta AI) and robotaxi business (Waymo). On the other hand, China is better positioned at the base of the hardware value chain, such as the supply of minerals, renewable energy production and pure-play chip manufacturing, while it is catching up on the AI model side.
Driven by higher volumes of research and patents, as well as significant government AI funding, AI subsidies and a huge nuclear energy initiative, we expect China to continue to close the gap with the US. Will the forthcoming 35% China income tariffs to be introduced by the new Trump administration be sufficient to slow down the trend? Stay tuned as we follow the AI Cold War throughout 2025 and beyond.
See you next week for more AI insights.