An AI accelerator is a processor designed to speed up the calculations that neural networks depend on – above all, multiplying large matrices of numbers – far more efficiently than a general-purpose computer chip (CPU). The term covers graphics processing units adapted for AI (see GPU), Google’s Tensor Processing Units (TPUs), custom chips designed by cloud companies and startups, and the neural processing units (NPUs) built into phones and laptops (see On-Device AI).
Why special chips
Training and running AI models consists mostly of the same simple operations repeated trillions of times. CPUs are built to handle many different tasks one after another; accelerators instead pack thousands of small arithmetic units that work in parallel, often at reduced numerical precision, which is enough for neural networks and saves energy (see Quantization). Equally important is memory: accelerators sit next to stacks of fast memory (see High-Bandwidth Memory) and are linked by high-speed networks so that thousands can act as one machine (see AI Supercomputer). Performance is commonly measured in operations per second (see FLOP).
The main players
Nvidia dominates the market – its data center revenue reached $107.6 billion in the quarter ending October 2026, up 110 percent from a year earlier. Its data center GPUs – from the V100 and A100 to the H100, the Blackwell generation and the Vera Rubin systems that began shipping in 2026 – power most of the world’s AI, helped by its CUDA software ecosystem (see CUDA). AMD competes with its Instinct line.
Google began using its own TPUs internally in 2015 and announced them in 2016; it offers several generations through its cloud, trains its Gemini models on them, and in 2026 split its eighth generation into separate chips for training and for inference. Anthropic agreed in October 2025 to use up to one million TPUs and in April 2026 signed for several gigawatts of next-generation TPUs from 2027. Amazon designs Trainium chips, now used for both training and inference – Trainium3 launched in December 2025, and Anthropic runs Claude on more than a million Trainium2 chips – alongside its older Inferentia line; Microsoft and Meta have developed their own as well. OpenAI announced in 2025 that it would co-design custom accelerators with Broadcom. Startups pursue radical designs: Cerebras builds a single chip the size of a silicon wafer, and Groq built chips specialized for fast inference – in December 2025 Nvidia took a non-exclusive license to Groq’s technology and hired its founder and president, while Groq continues as an inference cloud (see AI Antitrust). In China, where US export controls restrict access to the most advanced chips, Huawei’s Ascend line has become the main domestic alternative as Nvidia’s China sales collapsed. Critics point to “circular” deals in which chipmakers invest in the AI companies that buy their chips.
Outside China, almost all advanced accelerators are made by TSMC, mostly in Taiwan, using machines from the Dutch company ASML (see EUV Lithography); Huawei’s chips come from China’s SMIC. This concentration makes the supply chain a matter of geopolitics (see Compute Governance and Chip Smuggling).
Trade-offs
Specialized chips are faster and more efficient for the tasks they are built for, but riskier: AI methods change quickly, and a chip designed for today’s models may suit tomorrow’s less well. General-purpose GPUs have stayed dominant partly because they are flexible. Large companies design their own accelerators mainly to reduce dependence on Nvidia and lower costs; whether that will dent Nvidia’s lead is one of the central questions of the industry (see Moat).
Why it matters
Accelerators are the engines of the AI boom. Who designs and builds them, and who is allowed to buy them, shapes which companies and countries can train frontier models – making these chips one of the most strategic products in the world economy.
Written with Claude, made by Anthropic, which trains and runs Claude on Nvidia GPUs, Google TPUs and Amazon Trainium chips, has committed tens of billions of dollars to compute from Google, Amazon, Microsoft and Broadcom, and counts Google, Amazon, Microsoft and Nvidia among its investors.
Related
GPU · CUDA · High-Bandwidth Memory · AI Supercomputer · EUV Lithography · Export Controls · Hyperscaler
Sources: Norman P. Jouppi et al.: In-Datacenter Performance Analysis of a Tensor Processing Unit (arXiv) (2017) · Norm Jouppi (Google Cloud): Google supercharges machine learning tasks with TPU custom chip (May 19, 2016) · Anthropic: Google and Broadcom partnership for compute (April 6, 2026) · Anthropic: Expanding our compute partnership with Amazon (April 20, 2026) · Amazon: Trainium3 UltraServer (December 2, 2025) · Tom’s Hardware: Google splits its TPU into two chips for the first time (April 2026) · Groq: Groq and Nvidia enter non-exclusive inference technology licensing agreement (December 24, 2025) · Tom’s Hardware: Huawei expects $12 billion in AI chip revenue as Nvidia’s China market share hits zero (May 5, 2026) · Anthropic: Expanding our use of Google Cloud TPUs and Services (October 23, 2025) · OpenAI: OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators (October 13, 2025) · Nvidia: NVIDIA Announces Financial Results for Third Quarter Fiscal 2027 (November 17, 2026)
Leave a Reply