The results are in. After reviewing this year’s Ad Astra AI Infrastructure Competition entries, our panel of advisors selected AIRSYS LiquidRack™ as the 2026 winner, alongside three honorable mentions. Participants pushed the boundaries of what’s possible, showcasing ambitious innovation ahead of the 2026 AI Infra Summit.
.png)
AI is accelerating the demand for innovation in data center cooling. Higher rack densities require more efficient cooling, and existing technology wasn’t designed to meet those demands. LiquidRack™ addresses that challenge. Rather than relying on centralized cooling infrastructure, it integrates liquid cooling directly into the rack, making it easier for operators to deploy AI capacity, expand over time, and get more value from the power they already have.
That approach delivers measurable results. LiquidRack's patented spray cooling technology transfers heat up to three times more efficiently than conventional approaches, allowing more power to be dedicated to compute. The system achieves a PUE below 1.02, eliminates water consumption, and uses approximately 80% less dielectric fluid than immersion cooling, reducing costs and environmental impact. Its modular design also makes it practical for both new builds and existing data centers, allowing operators to adopt high-density AI infrastructure without overhauling an entire facility.
LiquidRack™ has already been validated through customer deployments across multiple industries and geographies, demonstrating that organizations don't have to choose between efficiency, scalability, and sustainability. By simplifying liquid cooling while improving energy efficiency and reducing infrastructure complexity, LiquidRack™ gives operators a practical way to support next-generation AI workloads.
Airsys's LiquidRack™ paired a clear innovation with the strongest real-world evidence in the pool, backed by validated deployments across multiple industries, not just projections. In a field tackling one of AI infrastructure's biggest bottlenecks, cooling density, LiquidRack stood out as ready to deploy today.
Up to 3x more efficient heat transfer with a PUE below 1.02 and zero water consumption.
Deployment across industries and geographies show real-world impact and operator trust.
Modular, scalable, and ready for the next wave of AI infrastructure growth.
Ad Astra submissions were evaluated by industry leaders with deep experience across AI infrastructure, data centers, enterprise technology, and emerging computing architectures.
TechArenna Advisory
Raejeanne brings more than three decades of leadership across silicon, data center infrastructure, and hyperscale cloud, including senior roles at Intel, Flex, and AWS. Her experience scaling technologies from engineering innovation to global adoption brings a valuable real-world perspective to evaluating the next generation of AI infrastructure.
TechArenna Advisory
Laura brings deep experience in corporate development, strategic partnerships, and scaling technology businesses, with leadership roles at Intel and Ampere and current operating experience as Co-Founder of MisaLabs.ai. Her perspective across capital strategy, partnerships, and AI business growth brings a valuable commercial lens to evaluating which infrastructure innovations have the potential to succeed at scale.
The Rafay Platform is software for operating an AI data center through one control plane and one API framework. It unifies Kubernetes, virtual machines, SLURM, bare metal, GPUs, and AI services, replacing fragmented tools with a consistent way to provision, govern, observe, and monetize infrastructure. Operators can apply multi-tenancy, access controls, quotas, policy enforcement, usage metering, and automation across the entire environment.
Rafay extends this operating model into inference through Token Factory, which turns GPU capacity into production AI services exposed through standardized APIs. Providers can deliver models on demand, meter token consumption, enforce tenant-level controls, and create usage-based services instead of selling only raw GPU hours. Rafay’s observability capabilities add unified telemetry across GPUs, servers, storage, networks, Kubernetes, and applications, with AI-assisted triage and automated remediation. In production environments, Rafay has improved GPU cluster efficiency by 25–40%, accelerated multi-cluster deployment by 50–70%, reduced platform engineering effort by 30–50%, and supported 3–5 times more tenants or clusters without increasing headcount. Rafay belongs on the shortlist because it connects infrastructure, inference, monetization, and operations in one scalable AI Data Center platform.
Although AI has transformed computing, much of the underlying data infrastructure still reflects an earlier era of enterprise IT. Separate storage systems, databases, event streams, vector stores, and data movement tools continuously copy and transform the same information, adding latency, cost, and operational complexity.
The VAST AI OS replaces that fragmented architecture with a single, shared data layer that enables files, objects, tables, vectors, events, and AI workflows to operate on the same data. VAST eliminates traditional infrastructure tradeoffs between performance, scalability, resilience, and simplicity while supporting millions of GPUs from a unified platform. Today, organizations use VAST to power more than 3 million GPUs worldwide, including some of the world's largest AI factories, where a single system serves over 1 billion CUDA cores.
By collapsing multiple infrastructure layers into one platform, customers reduce operational complexity, accelerate AI training and inference, and keep expensive GPU resources productive instead of waiting for data. As AI shifts toward continuous inference and persistent agentic workloads, the industry's competitive advantage will increasingly depend on the data layer. VAST has redefined that layer as a shared operating platform for AI rather than another storage system, making it foundational infrastructure for the next generation of AI computing.
For decades, AI infrastructure innovation focused on endpoints. Faster CPUs, then GPUs, then domain-specific accelerators. The network stayed passive: faster, but unaware of what runs on it. Production AI clusters run at 30 to 50 percent model FLOPs utilization, and the missing capacity goes to congestion, synchronization stalls, and fabric-induced idle time.
Cornelis defines the active compute fabric: an open architecture spanning scale-up and scale-out, with programmable compute throughout. Where legacy interconnects move data passively between endpoints, an active fabric works on the data. Congestion-free delivery is the foundation. Native collective acceleration is the differentiator. In-network compute is the destination.
CN6000 brings the next step of this evolution into production. The multi-protocol SuperNIC is engineered for 800 Gbps and carries Omni-Path, RoCEv2, and Ultra Ethernet on a single adapter. Cornelis modeling shows AllReduce collectives completing roughly 24 percent faster than standard Ethernet, cutting training time for a 250-billion-parameter model on a simulated 10,000 GPU cluster by about 13 percent. That same advantage carries into disaggregated inference and MoE expert routing. General availability is Q4 2026.
Each generation lets the network take on more work: full performance from the hardware you have, ready for the software you are yet to build.
