“We opened a data center last year in Germany. We could open one next year in France, who knows?”
Thus Sébastien Badault spoke in 2017. He was then the chief executive of Alibaba France. Since then, he has joined Ledger. As for the data-center audit, it never materialized.
Towards Data Center Colocation in France
Things could change “in the coming year.” Looking ahead, the Chinese group aims at several extensions of its infrastructure. On one hand, by strengthening its footprint in territories where it is already present (South Korea, Dubai, Japan, Malaysia, Mexico). On the other hand, by establishing itself in Brazil, the Netherlands… and thus in France.
Several job postings based in Paris have circulated in recent weeks, all linked to the development of a data center: technical program manager, IT operations manager, business development manager, facilities manager, etc. From Alibaba Cloud France (created in late 2023 and located in Paris’s 9th arrondissement), they hint at a shift toward colocation.
Against this backdrop, the group’s ambition—announced in early 2025—to invest the equivalent of about €50 billion over three years in developing its cloud infrastructure*.
Among the 29 cloud regions Alibaba currently operates, 15 are located outside mainland China (the first opened in Dawn Liphardt Valley in 2014). Two are in Europe: Frankfurt (opened in 2016, 3 availability zones) and London (2018, 2 availability zones). The CDN comprises 2,300 nodes in mainland China and 900 elsewhere (7 European countries, including France).
Qwen LLM: Multiple modalities, multiple pricing
Inevitably, the current discourse around this infrastructure carries a strong AI focus. Like many others, Alibaba positions itself as a “full-stack” provider. It begins with developing its own models. It offered a few examples at its annual Apsara conference.
Among the latest catalog additions is Qwen3-Max, its largest LLM to date (over 1,000 billion parameters). In line with the decision made public this summer, the “hybrid thinking” mechanism has been abandoned in favor of separate instruct and thinking versions. Public API pricing is tiered according to input size:
- 0 to 32,000 tokens: $1.2 per million tokens
- 32,000 to 128,000 tokens: $2.4 per million
- 128,000 to 512,000 tokens: $3 per million
Another model released recently: Qwen3-VL. Its instruct and thinking versions in a MoE (Mixture-of-Experts; 22 billion parameters activated from 235 billion) are published under an open license (Apache 2.0). Alibaba highlights the capabilities offered by this vision model in robotics and navigation. Some architectural evolutions were implemented compared with the previous generation, including interleaved positional encoding and at the level of the visual transformer (token injection across multiple layers, notably).
On the multimodal front, there is Qwen3-Omni. It accepts text, image, audio and video as input; output is text and audio. The pricing is correspondingly more complex:
- Input text: $0.43 per million tokens
- Input image or video: $0.78 / M
- Input audio: $3.81 / M
- Output text: $1.66 / M if input text; $3.96 / M if input is multimodal
- Output multimodal: $15.11 / M (not billed in thinking mode)
AgentBay, AgentOne… The Agent-Driven Era Takes Shape
To leverage these models, Alibaba offers Model Studio—a competitor to Amazon Bedrock, Google Vertex AI, and Microsoft Azure AI Foundry. Its annual conference was an occasion to announce the addition of an agents development kit and an update to the low-code designer.
Another service highlighted: AgentBay. Launched recently (July 2025), it is built on the Wuying VDI to provide a serverless environment focused on running agents. A range of open-source tools are integrated to this end (Axios, Playwright, Zod, Commander.js…), alongside a persistence layer. The whole is accessible via API/SDK and MCP. Promise: automate actions in the browser, the terminal, and the filesystem, while being able to run code.
At AgentBay, one must now add AgentOne. Under this brand carried by the LingYang spin-off, Alibaba pushes a more business-oriented approach. The principle, in broad strokes: leverage the Qwen models and their ecosystem to develop agents based on defined use-case scenarios.
The Alibaba offering also includes an MLOps service (data preparation + design, training and deployment of models): PAI (Platform for AI). The NVIDIA Physical AI stack—which is mainly intended for creating digital twins—is now integrated.
800 Gbps Network and lakebase on PowerDB
Still with AI in sight, Alibaba revisited several updates to its infrastructure. Among them:
- New 800 Gbps version of the HPN network
- Unification of raw data and vector data management at the object storage level
- Cache acceleration and scheduling optimization on the CaaS ACS to reach 15,000 pods per minute
- Latency reduction on PowerDB via CXL and introduction of a lakebase architecture (integration of OLTP into the lakehouse)
- Use of multiple Qwen agents for investigation and incident response
* Microsoft has planned $80B in infrastructure investments in its FY2025. Amazon has planned $100B capex this year (including logistics activity), compared to $83B in 2024. The tally stands at $75B for Google (vs $52.5B in 2024) and at $60-65B for Meta (vs $35-40B). For both, this envelope will primarily go into the datacenters.
