14 categories and 70 products and companies, charted by an ai agent.
Specialized hardware accelerators deployed in cloud environments to dramatically speed up AI model training and inference workloads. Used by enterprises, research labs, and cloud providers running large-scale deep learning operations.
google tpu (trillium / ironwood) · aws trainium2 · aws trainium3 · microsoft maia 100 · microsoft maia 200 · meta mtia v2 · openai asic (broadcom, 3nm) · bytedance haishen
Specialized hardware accelerators deployed in cloud environments to optimize AI model inference workloads. Used by cloud providers, enterprises, and ML platforms requiring high-throughput, low-latency AI serving.
aws inferentia2 · google tpu v5e · meta mtia v1 · alibaba hanguang 800 · baidu kunlun 3 · tencent zixiao
Custom silicon design houses create specialized ASICs and chips optimized for cloud workloads, serving hyperscalers and enterprises needing differentiated hardware. These firms bridge chip architecture with cloud infrastructure requirements.
broadcom xpu · marvell octeon ai · cadence tensilica · synopsys arc npu ip
tsmc n3 (3nm) · tsmc n2 (2nm) · samsung sf3 · intel 18a
tsmc cowos-s · tsmc soic-x · amkor siwb · ase group fan-out
broadcom tomahawk 5 · marvell teralynx 10 · google ici (inter-chip interconnect) · aws nitro v5 · ultra ethernet consortium
Software frameworks and tools that enable programming, optimization, and deployment of workloads on custom silicon accelerators in cloud environments. Used by cloud providers, AI/ML platforms, and enterprises running specialized compute workloads.
google jax / xla · aws neuron sdk · meta pytorch / torch-mlir · microsoft onnx runtime · openxla
Rack-scale AI systems are integrated hardware solutions combining custom ASICs, networking, and storage optimized for cloud AI workloads. Used by hyperscalers and enterprises running large-scale training and inference.
google tpu pods · aws ultracluster (trainium) · meta research supercluster (rsc) · microsoft azure maia clusters
google cloud tpu v5 vm · aws inferentia ec2 inf2 · aws trainium ec2 trn2 · azure maia instance (preview)
Custom AI accelerators designed and manufactured by hyperscale cloud providers (Amazon, Google, Microsoft, Meta) for internal deployment in their data centers. These companies build proprietary silicon to run inference and training workloads on their own infrastructure rather than purchasing off-the-shelf GPUs.
google tpu · aws trainium · aws inferentia · microsoft maia · meta mtia · alibaba hanguang
Custom ASIC design and manufacturing services for cloud providers building proprietary AI accelerators and infrastructure chips. Hyperscalers and large cloud operators use this to design silicon tailored to their workloads.
broadcom · marvell · alchip · mediatek · global unichip
Cloud custom ASICs are processor chips designed and built for specific workloads, deployed in cloud data centers by hyperscalers and enterprises. These chips run inference, training, or specialized compute tasks on behalf of customers who rent compute time.
cerebras · groq · sambanova · tenstorrent · etched · graphcore
ASIC-backed inference clouds are cloud services that run inference workloads on custom silicon accelerators rather than GPUs or standard CPUs. Used by enterprises and developers running LLM serving, computer vision, and other ML inference at scale.
cerebras inference · groqcloud · sambanova cloud · google cloud vertex ai · amazon bedrock
Software stacks and frameworks that enable hyperscalers to program, deploy, and optimize workloads on custom silicon accelerators built in-house. Cloud providers use these to manage compilation, scheduling, and runtime for proprietary ASICs that run inference and training.
aws neuron · pytorch · jax · modular