SCX.ai builds Australian inference service on ASICs

SCX.ai is building an Australian AI inference service for production workloads, using specialised accelerator hardware rather than competing directly with hyperscale AI training infrastructure.

The company is positioning its service between two deployment models: large AI data centres built around GPU clusters, and smaller models running directly on phones or laptops. Its focus is inference - running trained models to generate outputs - from distributed Australian data centres, with customers charged according to the number of tokens they consume.

Local inference

"We know that there is demand in our little business. We've got enough capacity in our business to do fine. We want people to understand there is an Australian alternative. That's our whole goal," said David Keane, Founder and CEO, SCX.ai.

SCX.ai separates AI training from the computing required to operate models once they have been trained. Training workloads are typically handled in environments built around GPU farms, PyTorch and Nvidia's CUDA software stack. SCX.ai is instead focused on production inference capacity: the computing businesses use when AI applications generate text, images or other outputs.

That distinction shapes its hardware choices. SCX.ai uses specialised chips for inference rather than general-purpose GPUs, which means models must be recompiled to run efficiently on its systems. The lack of CUDA on the hardware is not a barrier to inference, but can delay the availability of newly released models.

Keane said that delay can be a couple of weeks. Training, by contrast, would continue in the PyTorch-CUDA environment already familiar to AI researchers.

SCX.ai is not claiming complete independence from the global semiconductor supply chain. Its architecture still relies on advanced chips manufactured overseas. However, owning and operating the hardware in Australia gives customers greater operational control than relying entirely on a foreign-hosted AI service.

Open weights

"They're fully auditable. You can look at them now. Most people don't know how to do it, but there are some researchers that certainly do," said Keane.

SCX.ai's local-inference strategy is also built around open-weight AI models. Keane cited DeepSeek as an example of a model whose weights can be made publicly available and run independently of the company that developed it. Under SCX.ai's model, the software can run on SCX.ai-controlled infrastructure rather than being accessed only through an external provider's endpoint.

Keane contrasted this with proprietary hosted models, using Anthropic as an example. In those services, an application sends a request to the provider and receives generated tokens in return, while decisions about performance, availability and access remain with the provider.

SCX.ai runs Chinese-developed models in secure containers on its own machines, and customers can have their own version of a model. Each inference request therefore does not need to be processed by the original model developer, even when the underlying model was developed outside Australia.

The model still requires specialist expertise. Researchers can inspect model weights and fine-tune models, although most users do not have the skills to do so themselves.

Once open weights are obtained and deployed, the model can operate without sending every request back to its original developer. This supports SCX.ai's aim of providing locally operated inference while remaining part of an international hardware and software ecosystem.

Operating model

"We do charge in tokens. You pay us based on how many tokens you use. That's how we make money. These machines churn away and process the loads, and it's charged based on tokens," said Keane.

SCX.ai does not operate a GPU-as-a-service model in which customers rent specific hardware. Instead, customers send processing requests to SCX.ai's infrastructure and pay according to token consumption.

A request might ask an application to draft an email from an earlier example. SCX.ai's machines process that request and return a stream of tokens, which the customer's application or agent then converts into text, images or another output.

Customers neither buy the specialised hardware nor rent a physical server. SCX.ai owns the infrastructure and sells the inference capacity it produces, with token usage providing the billing unit.

The infrastructure is also designed to avoid some of the requirements of the largest AI campuses. SCX.ai's facilities operate in tier-three data centres, which do not require the same water-cooling systems or very large power requirements associated with major AI training sites.

SCX.ai describes this architecture as modular core: centralised enough to handle workloads too demanding for a phone or laptop, but distributed across smaller facilities rather than concentrated in a single large AI data centre. This allows capacity to be deployed without waiting for purpose-built hyperscale facilities.

The systems use specialised ASIC hardware. The chips are manufactured at TSMC and integrated into dedicated rack systems by a Californian company. The servers use proprietary networking and are designed around the accelerator rather than installing conventional cards in standard general-purpose servers.

Supply limits

That dependence on overseas hardware leaves SCX.ai exposed to geopolitical risks affecting the supply of advanced AI processors. Restrictions on future generations of chips could limit the company's ability to expand or upgrade, even if hardware already owned and installed in Australia remained available for existing workloads.

"Our belief is that that's the level of independence we can do. Any more than that is unrealistic," added Keane.