Amazon and OpenAI resell China's GLM-5.3 and Kimi K3; MiniMax is the cheapest Chinese lab at $13 billion
AWS put Zhipu's GLM-5.3 and Moonshot's Kimi K3 into Bedrock on revenue share days apart, and OpenAI's Codex now resells both through Baseten. The labs are interchangeable, so the platforms keep the spread, and SmartKarma is shorting the cheapest listed lab before its own raise.
Vincent Jiang · 2 min read
Two American platforms now sell Chinese intelligence and keep a cut. AWS put Zhipu's GLM-5.3 into Amazon Bedrock on October 5 and Moonshot's Kimi K3 on October 7, both paid by revenue share on invocation volume, and via the inference firm Baseten both models also run inside OpenAI's Codex, a US coding agent reselling Chinese weights 23. Zhipu's Hong Kong shares jumped more than 7% on the first deal 2. The tollbooth is American. The traffic is Chinese. The lab everyone is pricing, MiniMax, rides the same road and gets the same haircut.
Two camps, one $13 billion question
The argument is about MiniMax (100 HK), last valued at $13 billion, the cheapest listed Chinese lab 1. Smartkarma is short into the M3.1 launch: the model "looks level with tier-one Chinese peers at best," and its release "opens an equity-raise window" 4. The bull case is comparables. Zhipu already trades at $46.6 billion, Moonshot raised at nearly $50 billion after sitting at $4.3 billion at the end of 2025, and DeepSeek is closing a round of about $7.5 billion at roughly $75 billion ahead of a Shanghai STAR listing 1.
MiniMax trades at a quarter of the marks Moonshot and DeepSeek list at
Data
| Latest valuation | |
|---|---|
| MiniMax (listed) | $13B |
| Zhipu (listed) | $46.6B |
| Moonshot AI (private) | $50B |
| DeepSeek (private) | $75B |
Why the platforms set the price
The spread is wide, but Kechuangban Daily names the trap: Chinese models are highly substitutable with each other, which caps their bargaining power over revenue-share terms 3. MiniMax's own M2.1 has sat in Bedrock since February, part of a six-model Chinese push that included DeepSeek, Qwen and Kimi 2. GLM-5.3 scores 45 on the Artificial Analysis index and Kimi K3 44, against 58 for Claude Opus 5.5; MiniMax's M3.1 landed at parity, not leadership 14. When the products are interchangeable, the distributor keeps the spread.
Chinese models in Bedrock still trail the frontier by 13 to 14 index points
Data
| Artificial Analysis index | |
|---|---|
| Claude Opus 5.5 | 58 |
| GLM-5.3 | 45 |
| Kimi K3 | 44 |
The verdict date
Watch two events: Moonshot's Hong Kong float, planned for the first quarter of 2027 with early investor meetings possibly this month, and DeepSeek's STAR process beginning this year 1. If MiniMax raises at a multiple of $13 billion, the comp buyers win. If Chinese models' OpenRouter call share, which passed US models on October 6 3, keeps growing while listed revenue per call stays thin, the short does. Interchangeable models do not make moats. They make toll roads, and Amazon holds the booth.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
