What is AMD in AI? Instinct GPUs, explained simply

In plain words

AMD, short for Advanced Micro Devices, is a company that designs computer chips. Many people know it from laptops and gaming PCs: its Ryzen processors run ordinary computers. It does not make AI models or chatbots.

In AI, AMD matters for one product line: Instinct data-centre GPUs. Modern AI is trained and run on GPUs, chips that do huge numbers of small calculations at once. Nvidia is the market leader, and AMD is the main other supplier of the same kind of chip.

A chip is only useful with software that tells it what to do. Nvidia's is called CUDA; AMD's open equivalent is ROCm. AI developers mostly write code in common tools that can run on either, so the same model can often run on AMD or Nvidia hardware.

Everyday example

Two makers of the same kind of engine

Think of two companies that both make truck engines. Transport firms long bought mostly from one, so drivers and mechanics learned that brand. A second maker builds engines that fit the same trucks and carry the same loads, and gives buyers a real choice. In AI, Nvidia is the familiar brand and AMD is the second engine maker.

Where the example breaks: Swapping a truck engine is fairly simple. Moving AI software between chip makers can take real work, because some tools and tricks are written for one brand.

How it actually works

AMD's place in AI rests on three pieces: the chip, the memory around it and the software.

Designed by AMD, made by TSMC

AMD has had no factories since spinning them off as GlobalFoundries in 2008 to 2009. It designs the chips; contract factories make them, mainly Taiwan's TSMC, which has made its most advanced CPUs and GPUs since AMD moved its 7 nm production there in 2018.

Lots of memory per chip

Each Instinct GPU sits beside stacks of high-bandwidth memory. Big memory matters because a model's parameters must be read again for every word it writes. AMD has competed by fitting more memory per chip; it says its newest MI455X has 50% more than Nvidia's comparable chip.

Racks, not single chips

Large models are trained in data centres on thousands of GPUs. AMD now sells whole racks, called Helios, combining Instinct GPUs, its EPYC processors, networking and ROCm, ready to install as one unit.

Open software

ROCm is AMD's open software stack. Popular AI tools such as PyTorch and JAX can be installed on it, which lets most model code move between AMD and Nvidia GPUs with fewer changes than rewriting it, though performance tuning still differs.

Founded

1 May 1969, by Jerry Sanders and colleagues from Fairchild Semiconductor

Based in

Santa Clara, California

Led by

Lisa Su, chief executive since October 2014

Graphics roots

Bought the Canadian graphics-chip maker ATI in 2006, the start of its GPU business

Big AI companies also design their own chips, such as Google's TPU. See AI accelerators.

Questions people ask

What does AMD do in AI?

It designs the chips AI runs on. Its Instinct GPUs train and run large models in data centres, and its EPYC processors run the servers around them. It also supplies ROCm, the software that lets AI code use its GPUs. Its Ryzen processors power many ordinary laptops and desktops.

AMD vs Nvidia: what is the difference for AI?

Both make data-centre GPUs for AI. Nvidia is the market leader, and its software, CUDA, is what most AI code was first written for. AMD offers chips of the same kind, often with more memory per chip, and its own open software, ROCm. Several big AI labs now buy from both.

Does AMD make its own chips?

No, it designs them. AMD spun off its factories in 2008 to 2009, and since 2018 its most advanced processors and GPUs have been made by TSMC in Taiwan. AMD also uses other contract makers for some products.

Which AI companies use AMD chips?

OpenAI, Meta, Microsoft and Oracle were running AMD's first Helios rack systems as of October 2026. OpenAI agreed in October 2025 to deploy six gigawatts of AMD GPUs over several years, and Anthropic signed a deal for up to two gigawatts in July 2026.