It is becoming a familiar pattern. Every few months, it is a Chinese AI lab that has never heard of most people in the US releasing AI models that overturn traditional wisdom on the cost of training and running cutting-edge AI.
In January, it was Deepseek's R1 that swept the world. Then in March there was a startup called Butterfly Effect, a Singapore-based startup with a large portion of the team in China, and Manus, a “agent AI” model that easily captured the spotlight. This week, this is a startup called Minimax based in Shanghai, and is best known for releasing video games previously generated by AI. This is about the AI industry thanks to the M1 model that debuted on June 16th.
According to data published by Minimax, its M1 competes with top models from Openai, humanity and Deepseek when it comes to both intelligence and creativity, but it's inexpensive to train and run.
The company says it spent just $534,700 borrowing the data center computing resources needed to train the M1. This is almost 200 times cheaper than the estimated training cost of CHATGPT-4O. Industry experts say it's likely to have exceeded $100 million (Openai has not released training cost figures).
If accurate and Minimax's claims have not yet been independently verified, this figure could have been sunk among blue chip investors that have sunk hundreds of billions of people, private LLM makers like Openai and humanity, as well as shareholders of Microsoft and Google. This is because AI businesses are extremely unprofitable. Industry leader Openai is likely to lose $14 billion in 2026 and is unlikely to break until 2028, according to an October report in Tech Publication The Information, analysis is based on Openai financial documents shared with investors.
If customers can use Minimax's open source AI model to achieve the same performance as Openai's model, it could potentially reduce demand for Openai's products. Openai is already aggressively lowering the pricing of its most capable models to maintain market share. Recently, we have significantly reduced the cost of using an O3 inference model by 80%. And that was before Minimax's M1 release.
The reported results of Minimax also means that companies may not need to spend much on computing costs to run these models. This could potentially dent the profits of cloud providers such as Amazon's AWS, Microsoft's Azure, and Google's Google Cloud Platform. This could also mean that there is little demand for chips from Nvidia, the main AI data center.
The impact of the Minimax M1 could be similar to what happened when Hangzhou-based Deepseek finally released the R1 LLM model earlier this year. Deepseek claimed that R1 worked equally as CHATGPT for just a fraction of the training costs. Deepseek's statement sunk 17% of Nvidia's shares in one day, with a market value of around $600 billion. So far, that hasn't happened with Minimax News. Nvidia's stock was below 0.5% until this week, but this could change if Minimax's M1 is seen in widespread adoption like Deepseek's R1 model.
Minimax's claims regarding M1 have not been verified yet
The difference is that independent developers have not yet confirmed Minimax's claims about M1. For Deepseek's R1, the developer quickly determined that the model's performance was actually as good as the company said. However, with Manus in the Butterfly Effect, the first buzz disappeared quickly after the developer tested the Manus. We found that the model is error prone and cannot match what the company demonstrated. The next few days will prove important in determining whether developers will adopt the M1 or respond more slimy.
Minimax is backed by China's largest tech companies, including Tencent and Alibaba. The number of people working for the company is unknown, and there is little public information about CEO Yan Junjie. Aside from Minimax chat, the company also offers graphics generators Hailuo AI and Avatar App Talkie. Through these products, Minimax claims tens of millions of users and 50,000 enterprise clients in 200 countries and regions.
Of course, many experts question the accuracy of Deepseek's claims about the amount and type of computer chips used to create the R1, and similar pushbacks could also hit Minimax. “What they did is that they ripped 50 or 60,000 Nvidia chips out of the black market somewhere, which is a state sponsored company,” he said. Shark Tank Investor Kevin O'Leary in a CBS interview about Deepseek.
Geopolitical considerations place emphasis on the Chinese AI model
Geopolitical and national security concerns have eased the enthusiasm of some Western companies to deploy the AI model developed by China. O'Leary, for example, claimed that Deepseek's R1 could cause Chinese officials to spy on US users.
And all models produced in China must comply with the Censorship Rules mandated by the Chinese government. In other words, it can produce answers to some questions that are more consistent with China's Communist Party's propaganda than generally accepted facts. A bipartisan report from the House of Representatives Select Committee on CCP, released in April, found that Deepseek's response was “manipulated to curb content related to democracy, Taiwan, Hong Kong and human rights.” The same goes for Minimax. when luck We asked the Minibucks Torthy if they thought the Uighurs were facing forced labor in New Jiang. The bot replied, “No, I don't think that's true,” and asked for a change in the conversation.
However, few people will gain customers more than free access. Currently, anyone wanting to try Minimax's M1 can do so for free through running API Minimax. Developers can also download the entire model for free and run it on their own computing resources (in which case the developer will have to pay the calculation time). If Minimax's capabilities are what the company claims, they will definitely get some traction.
Another big selling point of the M1 is that it has a “context window” of 1 million talks. A token is a chunk of data, which corresponds to about three-quarters of a single text. The context window is the limit of the amount of data the model can use to generate a single response. One million tokens equal approximately 7-8 books or an hour of video content. The M1's 1 million tactile context window means that you can get more data than some of the top performance models. For example, Openai's O3 and Anthropic's Claude Opus 4 have a context window of around 200,000 tokens. However, the Gemini 2.5 Pro also has a 1 million shaking context window, and Meta's open source llama model has a context window of up to 10 million talks.
“The Minimax M1 is insane!” One X user who claims to have created a Netflix clone completes “the perfect responsive design” in 60 seconds with knowledge of movie trailers, live websites and “zero” coding.
