What Are Investors in Foundation Models Actually Betting On?

Several things happened in China's foundation model sector at the start of 2026. In private markets, Moonshot AI, the company behind the Kimi chatbot, closed two rounds in under two months. After a $500 million Series C, it raised more than $700 million in February in a round co-led by Alibaba and Tencent, doubling its valuation from $4.3 billion to over $10 billion. That made it the fastest Chinese AI company to become a decacorn. On the commercial side, less than a month after K2.5 launched, revenue over about 20 days had passed the whole of 2025. In public markets, Zhipu AI and MiniMax listed in Hong Kong in January 2026, and their shares kept climbing. Zhipu is up more than 220% since February, and MiniMax rose more than 190% in its first two trading days. On February 12, Zhipu released GLM-5 and, the same day, raised prices on its Coding plans by 30% or more. It was the first time a Chinese model company had raised prices sharply on its own initiative, and the market read it as a sign of confidence in commercialization. MiniMax released its M2.5 coding model that day too. The investment logic behind all this can be taken apart in two ways. The raw-material argument In the AI stack, a foundation model company is not one product at the application layer. It is the most upstream supplier of raw material, and in the history of any industry that position holds some of the highest value density. The barriers are technology, data, compute, and engineering systems built up over years, and none of them can be made up quickly. So there were never many candidates to invest in. For cloud providers this is a practical strategic problem. They can commit now to building their own models and accept the cost and uncertainty, or they can accept that they will pay a token tax to someone else later. Both paths have a price, but the longer they wait, the smaller the window for building in-house and the higher the cost of depend