GMI Cloud offers a free week of Alibaba's Qwen and Wan models to sell its inference API

GMI's first data centers supported Bitcoin mining ^[6]. The company moved into AI cloud infrastructure after clients and investors asked for GPU capacity, according to RuntimeWire ^[6]. In a 2025 interview, founder and CEO Alex Yeh described what came next as a sequence ^[7]:

1. Provide GPUs. 2. Build out the machine-learning stack. 3. Offer API access to customers who do not want to manage the hardware themselves.

Yeh has said customers increasingly want a simple API rather than direct interaction with GPUs, RuntimeWire reported ^[8].

Integration should be the cheap part of trying GMI. The models sit behind an OpenAI-compatible API ^[2], and I'd expect most teams to reach them by changing configuration in a client they already run. Evaluation is where the effort goes. GMI pitches Qwen3.8-Max at jobs with long contexts, Qwen3.8-Flash at work where latency matters and Wan3.0 at making video ^[3]. Those descriptions are GMI's own, and RuntimeWire notes that the promotion does not establish independent performance results ^[3][12]. For "latency-sensitive" to hold for a given product, that product's prompt lengths, concurrency and output sizes would have to resemble whatever traffic GMI had in mind. Seven days of free use ^[2] covers one weekly traffic cycle exactly once. A team can run a harness against all three models in that time, provided the harness is written before the clock starts.

The round behind GMI's infrastructure was mostly borrowed money. Debt made up about 82% of the 2024 Series A ^[16]. Banpu was to provide power, and Wistron planned to co-develop products with GMI ^[10]. In my view, that capital structure explains the move up the stack better than the campaign page does. Debt-heavy financing rewards keeping GPUs busy, and an API can fill them with traffic from developers who would not first arrange a dedicated GPU cluster ^[15].

Alibaba gets another route to developers out of the arrangement ^[14]. RuntimeWire previously reported that Nebius added an Alibaba Qwen model to its hosted inference menu ^[13]. For organizations the clock is longer: GMI's enterprise-credit applications stay open until 27 days after the date of Qwen's post ^[17].