Chips they run
The accelerators in their fleets.
Nvidia Blackwell
Blackwell is Nvidia's generation of data centre GPUs announced in March 2024, after Hopper. It includes the B200 GPU, the GB200 superchip that pairs two B200s with a Grace CPU, and the Blackwell Ultra parts B300 and GB300 of 2025
Nvidia Hopper
Hopper is the Nvidia GPU generation that powered the first wave of the generative AI boom. The H100, announced in March 2022 and shipped from late that year, became the most sought-after chip in the world in 2023
Nvidia Ampere
Ampere is the Nvidia generation of 2020, and its data centre chip, the A100, trained many of the models released between 2020 and 2023. It also includes the A10, A30 and A40 cards for inference and graphics
Other Nvidia GPUs
Beyond its flagship data centre chips, Nvidia sells many other GPUs that clouds rent out: the L40S and L4 of the Ada Lovelace generation, RTX workstation and gaming cards, the T4 for inference, and older data centre parts such as the Volta V100 of 2017 and the Pascal P100 of 2016. These cards fill the long tail of AI work, where a single cheaper GPU is enough.
AMD Instinct
AMD Instinct is AMD's line of data centre accelerators and Nvidia's main GPU rival. The MI250X powers Frontier and LUMI; the MI300A, which joins CPU and GPU in one package, powers El Capitan; and the MI300X of December 2023, with 192 GB of memory, opened the line to AI clouds
Google TPU
The Tensor Processing Unit (TPU) is Google's own AI chip. Google began using TPUs inside its data centres in 2015, announced them in 2016 and has rented them to others through Google Cloud since 2018
AWS Trainium and Inferentia
Trainium and Inferentia are the AI chips that Amazon designs through Annapurna Labs, the Israeli chip firm it bought in 2015. Inferentia, for running models, arrived in 2019; Trainium, for training, was announced in 2020
Intel Gaudi
Gaudi is Intel's line of AI accelerators, designed by Habana Labs, an Israeli start-up that Intel bought in December 2019 for about $2 billion. Gaudi 2 appeared in 2022 and Gaudi 3 in 2024
Cerebras wafer-scale
Cerebras makes the largest chips ever sold: each Wafer-Scale Engine is a whole 300 mm silicon wafer, instead of the many small chips usually cut from one. The WSE-2 of 2021 had 2.6 trillion transistors; the WSE-3 of 2024 has 4 trillion and 900,000 cores
Groq LPU
The language processing unit (LPU) is the chip made by Groq, a company founded in 2016 by Jonathan Ross, who earlier helped start Google's TPU project. Groq sells it almost entirely as a service, GroqCloud, which serves open models through an API
Other and custom chips
Beyond Nvidia, AMD and the chips with their own pages here, many companies and countries run their own accelerators. SambaNova sells its RDU; Microsoft designs Maia, Meta designs MTIA, and Tesla built the D1 for Dojo