Think has introduced Think Fabric, a bare-metal AI compute platform built to run heterogeneous, mixed-silicon infrastructure under a single integrated stack. The platform combines Think’s Node hardware, ILM software, and Constellation clustering system, targeting efficiency constraints in AI infrastructure such as cooling, density, GPU utilization, and infrastructure cost.

Think Fabric is designed as a unified system that pairs high-performance hardware with an orchestration layer, with the goal of giving operators more control and flexibility over how AI accelerators are deployed and used. Think is demonstrating Think Fabric publicly for the first time at LEAP 2026 in Riyadh.

On the hardware side, Node is aimed at high-density AI compute in a compact form factor. Think says Node uses a patent-pending sealed-loop cooling design, with each GPU using a patent-pending high-density cooling topology intended to extract heat from the GPU core and memory more efficiently. Think claims Node can deliver “radically higher sustained compute” and “dramatically higher sustained power density” compared with a typical AI data center server with the same GPU count and silicon, and says the form factor can reach up to 2.35x greater power density than other server manufacturers with like-for-like specifications.

ILM (Intelligent LLM Management system) is Think’s proprietary orchestration software, intended to increase GPU utilization by optimizing workload distribution across AI accelerators. ILM can run multiple AI models on a single GPU, enabling model sharing on GPUs that might otherwise be underutilized, which Think says can reduce the total hardware required.

Constellation connects multiple Think Nodes into a unified cluster using a high-bandwidth fiber optic interconnect for low-latency, multi-node coordination. In Think Fabric, ILM is used for synchronization and local orchestration across nodes. Think also says Constellation is intended to scale AI compute clusters without the footprint and complexity associated with traditional data center infrastructure.

For data center engineers, the most concrete technical differentiator here is the attempt to treat cooling, packaging density, orchestration, and interconnect as one co-designed system instead of discrete layers sourced and tuned separately. But several of the biggest claims—especially around sustained compute, sustained power density, and the 2.35x power-density figure—will matter most when operators can map them to rack-level constraints like facility water strategy, allowable heat rejection, and the practical limits of serviceability in dense deployments.

“This first era of AI has been defined by the rush to scale, but it’s clear that the long-term success of AI will only happen by addressing the issues around efficiency and cost,” said Ahmed AlSharif, Founder and CEO of Think. “The solution to the rising cost of AI infrastructure is not to spend even more or build bigger—it’s to maximise the investments already made, and to lower the cost of AI compute through increased efficiency.”

Think Fabric is being demonstrated at LEAP 2026 in Riyadh, according to Think. More information is available at think-ai.com and Think’s LinkedIn page.

Source: Think.