Think launches Fabric to squeeze more work from mixed AI chips

Why it matters

Think is betting that AI infrastructure buyers will shift spending toward utilization, cooling and local control. Independent benchmarks and named deployments are still missing, leaving the case for its vertically integrated Riyadh-built stack dependent on company-reported results.

Riyadh AI infrastructure company Think founder Ahmed AlSharif launched Think Fabric on August 31, bundling Think's liquid-cooled AI servers, workload-management software and cluster interconnect into a bare-metal platform designed to run models across different accelerator architectures. Think is demonstrating Fabric at LEAP 2026 in Riyadh, according to its August 31 announcement.

Fabric turns a lesson from AlSharif's career in video games into an infrastructure thesis: expensive, fixed hardware has to be treated as a constraint worth engineering around. According to Think's company biography, AlSharif studied computer games technology at Abertay University before working on PlayStation hardware and software, including PS4 and PSVR architecture. He later held technical leadership roles at Electronic Arts and Meta, then served as chief technology officer of Saudi game developer Sandsoft Games.

When AlSharif introduced Think publicly, he described nearly two decades spent asking how engineers could extract more performance from the hardware available to them. Console development taught him to optimize within a fixed silicon budget. His subsequent work at Meta exposed the other extreme, where large infrastructure estates could contain costly capacity that remained underused. Think's founding argument is that hardware designers and workload-software developers have been separated for too long, leaving nobody responsible for system-wide efficiency.

AlSharif founded Think in Riyadh in the first quarter of 2025. Prototyping began that July, with the cooling system and orchestration software developed together. Ammar Enaya, a former regional executive at Cisco, Aruba Networks, Hewlett Packard Enterprise and Vectra AI, joined as co-founder and chief operating officer in October 2025 to build Think's commercial operation. Enaya previously led Ajlan Tech and holds a computer engineering degree from King Fahd University of Petroleum & Minerals.

One stack from cooling loop to model scheduler

Think Fabric combines three products that Think had been developing separately: Node, ILM and Constellation.

Think's Node product page describes a family of compact, liquid-cooled multi-GPU machines, ranging from a two-accelerator MicroNode with 192 GB of aggregate GPU memory to an eight-accelerator RackNode with 768 GB. Think labels the RackNode specifications as subject to final validation. Each machine ships with Think's software and a high-bandwidth interface for joining other nodes.

Think says Node's sealed cooling loop reduces the need for extensive data-center cooling infrastructure. Think also claims as much as 2.35x the power density of comparable servers with the same GPU count and silicon. That comparison has not been independently benchmarked, and the August 31 release does not provide the test configuration behind it.

ILM, short for Intelligent LLM Management, is the orchestration layer. It places workloads according to available memory, thermal headroom and chip affinity, allowing several smaller models to share one accelerator or larger models to be divided across multiple devices. Think says ILM can schedule inference, fine-tuning and training on the same hardware.

Think reports 92.3% sustained accelerator utilization with between 0% and 3.5% overhead in multi-model testing. Those are Think's measurements rather than results from an independent testing lab. They also describe a selected workload, rather than utilization across a production fleet with changing traffic and model sizes.

Constellation connects the machines through high-bandwidth fiber links and presents their memory and compute as a coordinated pool. Think says ILM supports NVIDIA and Intel accelerators, with other architectures under development. The launch materials do not establish how mixed-silicon workloads perform in production.

The vertical integration bet

Bare-metal AI capacity is already available from providers including Lambda and Voltage Park. Think is taking responsibility for more of the stack: the enclosure, cooling loop, workload scheduler and node-to-node fabric. Think also targets installations that remain on premises or air-gapped, an important sales point for governments and enterprises that cannot send sensitive workloads to shared cloud infrastructure.

Heterogeneous orchestration is becoming a separate area of competition. Callosum, for example, is building systems that divide models and workflows among chips including AWS Trainium, Cerebras and SambaNova hardware. Think's distinction is physical integration. AlSharif is betting that scheduling decisions improve when the same engineering organization controls thermal behavior, server architecture and software placement.

That approach concentrates the execution risk inside Think. Supporting mixed chips requires maintaining drivers, kernels and scheduling logic across architectures that change quickly. Building hardware in Riyadh adds manufacturing, supply-chain and field-support demands to an already difficult software problem. Think's machines are designed, prototyped and manufactured in Riyadh, giving AlSharif tighter control over iteration while requiring Think to scale disciplines that software-only infrastructure vendors can avoid.

Investors are funding the factory work

Think raised more than $8 million in a pre-seed round announced on July 15. RAED Ventures and Wa'ed Ventures co-led the financing, joined by Dhahran Techno Valley's venture arm and strategic angel investors. Think said it would use the capital for hiring, manufacturing, product development and commercial expansion.

The six-week gap between the financing announcement and the Fabric launch shows what the round is meant to buy. Think is moving from individual components and proofs of concept toward a product presented as one infrastructure platform. AlSharif's challenge is proving that vertical integration produces better economics after hardware costs, software licensing, support and deployment are included.

The launch materials name no customers using the complete platform, provide no delivery schedule and include no independent benchmarks. Think Grid, a separate subscription compute service hosted in Riyadh, offers access to the same underlying architecture, but its monthly capacity pricing should not be treated as the purchase price for Fabric hardware.

For AlSharif, Fabric carries console engineering discipline into a market that has spent several years solving capacity shortages by ordering more accelerators. His proposition is straightforward: organizations have already bought plenty of costly silicon, and Think can make each chip produce more useful work. Independent benchmarks and repeatable, named customer deployments will determine whether that proposition survives outside Think's own lab.