NuAxi
NuAxiSemiconductor trade intelligence

MangoBoost, PFN Partner on Heterogeneous AI Inference Systems

What happened

MangoBoost announced on November 18 that it had signed a memorandum of understanding (MOU) with Preferred Networks (PFN) to collaborate on artificial intelligence (AI) inference infrastructure technologies. The two companies will conduct a proof of concept (PoC) for a next-generation "AI Token Factory" architecture, exploring ways to improve system utilization by increasing AI inference throughput and reducing latency. The proposed architecture combines MangoBoost's data processing units (DPUs) and other solutions with PFN's MN-Core AI inference processors and AMD graphics processing units (GPUs). GPUs will handle the prefill stage, in which AI models process input instructions, while PFN's AI inference chips will perform the decode stage, which generates tokens; MangoBoost's DPUs, key-value (KV) cache pooling storage and prefill-decode disaggregation (PDD) solutions will be used to optimize data transfers and KV cache management. The companies will initially validate the architecture in a simulation environment; once samples of the MN-Core L1000 chip become available, they will expand testing to performance validation on actual hardware, and they will also consider publishing a joint research paper based on the evaluation results. An MOU signing ceremony will be held at SC26, a supercomputing conference taking place in Chicago from November 15 to 20, local time.

Impact and signal

Medium impactImpact score: 50/100

Affected components and suppliers

data processing units (DPUs)MN-Core AI inference processorsAMD graphics processing units (GPUs)KV cache pooling storageprefill-decode disaggregation (PDD) solutions
MangoBoostPreferred Networks (PFN)AMD
AI inferenceMOUheterogeneous computingDPUAI Token Factoryproof of concept

Source and time

Source: The Elec · 11/18/2026 · View original source

Open the NuAxi live feed →

Real-time pricing, supply signals and AI analysis for semiconductor traders