Multiverse Computing and Qualcomm Collaborate to Deliver AI Models That Run Up to 93% Faster on Dragonfly Accelerators

Real-time emergency medical reporting (compressed LLM):

• Up to 93% faster response times
• Up to 44% higher throughput
• Up to 45% reduction in memory usage
• Up to 21% reduction in power consumption
• No loss in accuracy

Confidential financial document RAG chatbot:

• Up to 35% faster response times
• Up to 54% higher throughput
• Up to 45% reduction in memory usage
• Up to 14% reduction in power consumption
• No loss in accuracy

For the medical use case, the results were independently corroborated by multiple outlets, confirming the 93% faster response time, 45% memory savings, and 21% power reduction . Multiverse Computing noted that running on the newer Dragonfly AI200/AI250 accelerators directly is expected to yield even better results .

How It Works

Multiverse Computing optimizes AI models before deployment, shrinking the compute and memory each model requires. This frees up headroom on existing hardware, letting data center operators serve more inference requests or run more models concurrently without adding new accelerators or expanding infrastructure .

Broader Industry Shift

This collaboration reflects a fundamental pivot in the data center industry toward efficiency-first AI infrastructure. Key signals of this shift:

• Qualcomm's DNA in efficiency: Qualcomm built its leadership in mobile and IoT, where maximizing performance within strict power and thermal constraints is essential — principles now directly applied to AI data centers .
• Hardware-level innovation: Qualcomm's Dragonfly AI250 introduces High Bandwidth Compute (HBC) Gen 1, delivering 133 TB/s of effective memory bandwidth per card (18x more than AI200), directly attacking the inference memory bottleneck that drives much of AI's cost and energy waste .
• Scaling without proportional cost: As AI adoption accelerates, operators need to scale deployments without a proportional increase in hardware, energy, and cost. Combining efficient hardware with optimized models offers a practical path to doing so .
• Industry validation: Both companies explicitly framed the collaboration around "breakthrough improvements in performance, cost, and energy efficiency at scale," with Qualcomm's Dino Flore stating that "customers need solutions that deliver high performance with greater efficiency" as AI adoption accelerates across enterprise and cloud infrastructure .

The partnership is a concrete example of the industry moving away from a "more GPUs" mindset toward a performance-per-watt and total cost of ownership focus, where software-hardware co-optimization is the primary lever for sustainable AI scaling.