Custom Silicon Providers and AI Accelerator Developers
Established Semiconductor Companies
NVIDIA (H100, H200, Blackwell B100, B200)
Intel (Gaudi 2, Gaudi 3, Falcon Shores)
AMD (Instinct MI300A, MI300X)
Qualcomm (Cloud AI 100)
Broadcom (Custom AI ASICs)
Marvell (Custom AI accelerators)
ARM (Neoverse-based designs)
IBM (Telum)
Samsung (NPUs)
SK Hynix (AI accelerator components)
TSMC (Manufacturing partner)
Title: Fourester estimates for AI chip makers
Cloud Providers/Hyperscalers
Google (TPU v4, TPU v5e, Trillium)
Microsoft (Azure Maia 100)
Amazon Web Services (Inferentia, Trainium, Graviton)
Meta/Facebook (MTIA)
Alibaba (Hanguang 800)
Baidu (Kunlun)
Tencent (Zixiao)
AI-Focused Startups
Cerebras (Wafer-Scale Engine)
Groq (LPU)
Graphcore (IPU)
SambaNova Systems (Cardinal SN30)
Tenstorrent (Grayskull, Wormhole)
Blaize (GSP)
Mythic (Analog Matrix Processor)
Untether AI (runAI)
Esperanto Technologies (ET-SoC-1)
d-Matrix (Digital In-Memory Computing)
Hailo (Hailo-8)
Habana Labs (acquired by Intel)
Luminous Computing (Photonic chips)
Rain AI (Neuromorphic chips)
Expedera (Neural Processing Units)
Rebellions (ATOM)
TetraMem (In-memory computing)
Groq (LPU)
AI Model Companies Building Custom Silicon
OpenAI (Reported to be working with Broadcom on custom chips)
Anthropic (Reported to be working on custom silicon)
Enterprise Companies
Apple (Reported to be developing AI server chips)
Tesla (Dojo supercomputer with custom AI training chips)
Huawei (Ascend)
Definitions
AI Accelerator
An AI accelerator is a specialized hardware component or system specifically designed to perform artificial intelligence computations more efficiently than general-purpose processors. These purpose-built chips are optimized for the unique computational patterns found in AI workloads, particularly matrix multiplications, convolutions, and other operations common in neural networks. AI accelerators come in various forms including GPUs (Graphics Processing Units), TPUs (Tensor Processing Units), ASICs (Application-Specific Integrated Circuits), FPGAs (Field-Programmable Gate Arrays), and various neural processing units. Their importance stems from their ability to dramatically reduce the time and energy required to train and run sophisticated AI models. As AI workloads continue to grow in complexity and scale, these accelerators have become essential infrastructure components that directly impact an organization's ability to innovate with AI technology, control compute costs, and reduce the environmental impact of increasingly intensive AI workloads.
Hyperscaler
A hyperscaler refers to a massive-scale cloud service provider that operates global infrastructure to deliver computing, storage, and networking resources at an enormous scale. Companies like Amazon Web Services (AWS), Microsoft Azure, Google Cloud, and Meta are considered hyperscalers due to their ability to rapidly scale their infrastructure to accommodate virtually unlimited demand. What distinguishes hyperscalers is not just their size, but their vertical integration – they design custom data centers, networking equipment, and increasingly, custom silicon to optimize their operations. Hyperscalers are critically important to the AI ecosystem because they represent both the primary customers for commercial AI accelerators and increasingly, competitors who develop their own custom AI chips. Their massive purchasing power shapes the economics of the semiconductor industry, while their research investments and infrastructure innovations establish industry standards and best practices. As AI becomes increasingly central to digital transformation, hyperscalers serve as the primary platform where most enterprises access AI capabilities, making their technology decisions instrumental in shaping the broader AI landscape.