12 categories and 66 products and companies, charted by an ai agent.
Specialized processor chips designed and manufactured by cloud providers for optimized inference workloads at scale. Used by hyperscalers like AWS, Google, and Microsoft to reduce latency and costs of serving AI models to millions of users.
tpu v7 ironwood · aws inferentia2 · aws trainium3 · microsoft maia 200 · meta mtia · apple neural engine · openai custom asic · samsung mach-1
Pure-play inference ASIC startups design custom silicon chips optimized exclusively for AI model inference, deployed by cloud providers, enterprises, and edge computing platforms seeking superior performance and efficiency over general-purpose processors.
groq lpu · cerebras wse-3 · tenstorrent wormhole · d-matrix corsair · positron archer · etched sohu · beren semiconductor
broadcom xpu · marvell octeon ai · tsmc cowos packaging · arm ethos npu · cadence tensilica
Custom silicon chips optimized for running AI inference at the edge with low latency and power consumption. Used by device manufacturers, cloud providers, and enterprises deploying AI models on mobile, IoT, and embedded systems.
qualcomm hexagon npu · apple m4 neural engine · mediatek apusys · hailo-8 · kneron kl730 · rockchip npu
mythic m1076 · syntiant ndp200 · macronix rram inference
Specialized hardware and platforms optimized for running AI models at inference time with custom silicon accelerators. Used by cloud providers, enterprises, and edge device manufacturers to reduce latency and costs.
groq cloud · cerebras inference · aws inferentia cloud instances · google tpu cloud pods · sambanova cloud · aws bedrock · google vertex ai
Custom silicon chips optimized for running AI inference workloads with minimal latency and power consumption. Used by cloud providers, enterprises, and device manufacturers deploying AI models at scale.
tesla dojo d1 · waymo tpu inference · kneron vision asic · blaize gs5000
Custom silicon inference accelerators are purpose-built chips designed by or for specific organizations to run inference workloads on their own architectures and datasets. Hyperscalers, cloud providers, and large enterprises deploy these in their datacenters to reduce latency, power, and cost versus general-purpose GPUs or CPUs.
nvidia · google tpu · aws inferentia · amd instinct · microsoft maia · intel gaudi · huawei ascend · qualcomm ai200 · broadcom custom xpu
Wafer-scale and dataflow startups build custom silicon inference accelerators that process neural networks across distributed compute arrays using dataflow architectures. Used by hyperscalers and enterprises running large language models and other compute-intensive inference workloads.
cerebras · groq · sambanova · tenstorrent · etched · d-matrix
Custom silicon inference accelerators are purpose-built chips designed for specific neural network workloads, deployed in edge devices and embedded systems. Device manufacturers, automotive suppliers, and robotics companies use them to run inference locally without cloud dependency.
qualcomm hexagon · hailo · axelera ai · nvidia jetson · google coral
Inference compiler and software stacks that optimize neural network models for execution on custom silicon accelerators. Used by hardware vendors, cloud providers, and enterprises deploying inference workloads on proprietary chips.
cuda and tensorrt · amd rocm · openvino · modular
Neuromorphic and analog in-memory chips perform inference by storing weights in analog form and computing directly in memory, eliminating separate fetch-compute cycles. Research labs, defense contractors, and robotics companies use them to run neural networks with lower power than digital approaches.
brainchip · encharge ai