ADATA’s TRUSTA wants to reduce AI deployment costs by using DRAM and SSDs alongside GPUs

ADATA’s enterprise storage brand TRUSTA has introduced a new AI memory solution designed to reduce the pressure on GPU memory during AI inference and fine tuning. The product is called AI Scaler Extended Memory, and it aims to let enterprises deploy large AI models across GPU memory, system DRAM, and high speed SSD storage instead of relying only on expensive GPU VRAM.

The idea directly targets one of the biggest problems in AI infrastructure. Large language models often need huge amounts of memory to run efficiently, and that usually pushes companies toward high end GPUs with large VRAM pools. Those GPUs are expensive, hard to secure, and in high demand because of the rapid growth of AI workloads.

TRUSTA says its AI Scaler Toolkit can extend model deployment beyond GPU only systems. In practical terms, that means parts of the model workload can use system memory and SSD storage as part of a broader memory hierarchy. The goal is not to make SSDs as fast as GPU memory, but to make AI infrastructure more flexible and less dependent on buying multiple costly accelerators.

The main pitch is lower cost for AI inference and fine tuning

According to TRUSTA, the solution can reduce deployment costs by more than 50 percent in model inference and fine tuning scenarios. The company says some inference workloads that would normally require multiple GPUs can be optimized to run on one GPU combined with expanded system memory.

That could be useful for enterprises that want to run AI on premises because of privacy, compliance, cost control, or data location requirements. Cloud AI can be convenient, but many companies still prefer or need local infrastructure for sensitive data and predictable long term costs.

TRUSTA is also positioning the toolkit as free and open source. It is not tied to one fixed hardware configuration, so companies, researchers, and developers can configure systems based on their own budgets and performance needs.

Model support is broad at launch. The toolkit supports popular model families including Llama, Qwen, Mistral, Mixtral, GPT OSS, DeepSeek, Phi, and Gemma. It also supports agent focused applications such as OpenClaw, NemoClaw, and Hermes Agentic, which shows TRUSTA is targeting newer AI agent workflows, not only basic chatbot inference.

The hardware side is also important. TRUSTA is showing its TD7P51 ECO PCIe Gen5 enterprise SSD at Computex, with capacities up to 15.36TB and support for U.2, E1.S, and E3.S form factors. The SSD includes Flexible Data Placement technology, which is meant to improve reliability and stability through smarter data placement.

This launch also signals a bigger shift for ADATA. The company is no longer presenting itself only as a memory and storage supplier. With AI Scaler, it is trying to move into software and hardware integrated AI infrastructure, where storage, DRAM, and GPUs work together as part of one deployment strategy.

The real test will be performance. Using DRAM and SSDs can reduce cost, but GPU memory remains much faster. The success of this approach will depend on how well the software moves data, how much latency it adds, and which AI workloads benefit most.

Still, the direction makes sense. AI demand is growing faster than GPU supply, and many companies cannot simply buy their way out of the memory problem. TRUSTA’s approach gives enterprises another option: use the memory and storage already available in the system more intelligently, and reduce dependence on expensive GPU only configurations.