Super Micro Computer (SMCI) shares plunged 10.03% to $26.63, a decline of $2.97, amid heavy selling pressure. The drop came even as the company announced a new AI storage solution developed through a collaboration with IBM and NVIDIA, a technology designed to accelerate AI inference, improve GPU efficiency, and support expanding enterprise workloads.

Supermicro Partners With IBM and NVIDIA on AI Storage Solution

Supermicro introduced a new KV cache storage platform that combines technologies from three major infrastructure providers. The system integrates Supermicro Petascale all-flash servers, IBM Storage Scale software, and NVIDIA Dynamo, and is built on AMD EPYC 9005 processors with NVIDIA Spectrum-X Ethernet networking.

The collaboration targets growing demand for faster AI inference across enterprise applications and cloud infrastructure. Modern AI systems process larger datasets and longer conversations, increasing pressure on existing memory resources, and companies require storage systems capable of handling complex workloads without reducing processing efficiency. KV cache offloading speaks directly to that constraint: it moves reusable model context from limited GPU memory to shared flash storage, turning cache capacity into a resource that can scale with demand.

Supermicro designed the platform to extend KV cache capacity beyond local GPU memory. This approach allows AI systems to reuse previously processed information instead of repeating calculations. Shared storage also helps organizations expand inference capacity while maintaining consistent application performance.

KV Cache Technology Delivers Up to 56x Faster First-Token Response

Supermicro and NVIDIA tested the technology using a large language model containing 120 billion parameters. Their testing recorded up to 56 times faster time-to-first-token performance with 130,000-token prompts, demonstrating substantial improvements when the system reused cached information. Time-to-first-token measures the wait before a model starts producing output, a latency figure that grows with prompt length when context is not cached.

As input lengths increased from 30,000 tokens to 130,000 tokens during testing, response preparation time rose by only 0.377 seconds with caching, compared with 29.55 seconds without caching. The findings highlighted the technology's potential benefits for applications processing extensive information.

The platform also supports higher GPU utilization by reducing unnecessary computation and processing delays. Its shared architecture allows multiple GPU clusters to access stored context efficiently, and the companies validated the system under heavy workloads and competing network traffic conditions. For operators running many concurrent inference workloads, that combination ties storage architecture directly to GPU productivity and serving capacity.

Supermicro Targets Enterprise AI Infrastructure Growth

The announcement expands Supermicro's infrastructure portfolio as businesses increase spending on AI deployment and inference services. Enterprise applications increasingly rely on coding assistants, automated workflows, and systems that process lengthy documents, all of which require reliable storage capacity alongside powerful computing resources.

Supermicro also offers factory rack integration and deployment services for the new platform. According to the company, these services can shorten deployment periods from weeks to days. IBM contributes storage management capabilities, while NVIDIA supplies inference software and networking technology.

The partnership strengthens Supermicro's position in enterprise AI infrastructure despite the stock's sharp decline. The company said its latest solution addresses computing efficiency, storage scalability, and system reliability across demanding applications. However, the announcement included no financial forecasts or confirmed revenue contributions from the new platform, leaving customer adoption and the platform's commercial contribution to be observed in future company disclosures.

Source: Blockonomi