5 categories and 25 products and companies, charted by an ai agent.
Specialized processors designed and manufactured by hyperscalers (AWS, Google, Microsoft, Meta) for accelerating training of large language models and other AI workloads on their own infrastructure. Used internally by these companies to reduce training costs and latency for proprietary models.
google tpu · aws trainium · microsoft maia · meta mtia · nvidia · amd instinct
Inference chips are custom silicon designed by hyperscalers to run trained AI models at scale in their data centers. Cloud providers, search engines, and large AI platforms use them to serve language models, recommendations, and computer vision at lower cost per inference than general-purpose GPUs.
aws inferentia · groq · cerebras · sambanova · etched · qualcomm ai
Custom silicon design services and tools that hyperscalers use to build proprietary AI accelerators and processors for their data centers. Used by companies like Google, Meta, Amazon, and Microsoft to design chips tailored to their workloads.
broadcom · marvell · alchip · mediatek · tsmc
Software stacks and frameworks that enable hyperscalers to design, simulate, and deploy machine learning workloads on custom silicon accelerators built for their own data centers. Used by cloud infrastructure teams at hyperscalers like Google, Meta, and Microsoft to optimize chip utilization for inference and training.
aws neuron sdk · openxla · pytorch · modular · cuda
NPUs (Neural Processing Units) built into edge devices and embedded systems that run inference workloads locally without cloud dependency. Device makers, automotive OEMs, and consumer electronics manufacturers integrate these to enable on-device AI.
google tensor · apple neural engine · google edge tpu