How ai& is Building a Sovereign AI Inference Stack in Japan

Artificial intelligence infrastructure is increasingly becoming a systems engineering problem.

Enterprises need compute, model serving, application deployment, security and data governance to work together. The location of that infrastructure is also becoming more important as organisations place sensitive business data into large language model workflows.

ai& is approaching that problem by building a vertically integrated AI stack based in Japan.

What sovereign AI means for ai&

The central idea is control over where inference runs and where data is processed.

ai& owns and operates infrastructure in Japan and uses that infrastructure to serve AI models through its inference platform.

For customers using the standard service, inference requests remain within Japan. The company also operates a zero-data-retention policy and states that customer inputs and outputs are not used for model training.

That gives organisations a clearer infrastructure boundary than using an API where the underlying compute location may be less relevant to the application developer.

Matching hardware to the workload

One of the more technically interesting parts of the ai& architecture is heterogeneous compute.

The company operates accelerator infrastructure from multiple hardware vendors rather than standardising the entire service around one processor family.

Its current inference material highlights NVIDIA and AMD hardware, while the wider ai& infrastructure also includes Tenstorrent systems.

The engineering objective is to choose the most appropriate architecture for a specific workload and optimise how that workload moves through the inference stack.

This can include separating different stages of model execution across different hardware where that improves throughput or utilisation.

ai& has also publicly documented work comparing NVIDIA, AMD and Tenstorrent systems for video inference, demonstrating how different architectures can provide different performance, quality and cost characteristics.

Why ai& is focused on inference

ai& currently focuses primarily on delivering finished inference rather than renting raw accelerator capacity.

That distinction matters.

The user sends a model request and receives generated tokens. ai& handles the hardware scheduling, inference software and underlying infrastructure.

The company is also expanding into model training and post-training, but inference remains the main production service.

This reflects how AI deployment is changing. As agentic applications generate more requests and longer context windows, inference can become one of the largest ongoing costs in an AI product.

Reducing the cost per useful output therefore has a direct effect on whether an application can scale economically.

Using existing developer workflows

Infrastructure optimisation is less useful if developers need to rebuild their software around it.

ai& addresses this through an OpenAI-compatible API.

Applications already making compatible API calls can move to ai& by changing the service endpoint and API credentials.

Streaming, tool calling and structured outputs are supported, allowing existing agents, coding tools and production applications to continue using familiar interfaces.

Enterprise customers can also request custom environments when they need a specific model or isolated deployment.

Build.io extends the stack

The other part of the ai& strategy is application deployment.

In April 2026, ai& acquired Build.io, a Platform as a Service provider designed to take applications from source code to production infrastructure.

This expands the stack above inference.

A development team can build an application, deploy it through Build.io and connect it to AI models running on ai& infrastructure without managing each infrastructure layer independently.

For ai&, that supports the broader goal of controlling and optimising more of the path between the accelerator and the final AI application.

Where sovereign AI becomes useful

The approach is particularly relevant where organisations need predictable data residency, control over model choice or an alternative to closed AI APIs.

Japanese enterprises are an obvious target because workloads can remain inside the country.

The same architecture can also appeal to international organisations that want clearer visibility into where inference happens and how their data is handled.

As AI workloads become larger and more autonomous, these infrastructure decisions will increasingly sit alongside model quality when engineers choose how to deploy production AI.

ai& is betting that inference performance, cost, hardware flexibility and sovereignty can be engineered together within the same stack.