AMD

What AMD is

AMD is a semiconductor company that designs processors for data centers and supercomputers. Its AI business centres on the Instinct line: data center GPUs such as the MI300X, and accelerated processing units (APUs) that combine CPU cores and GPU compute units in a single package^[1][7]. These are the chips AMD sells into the machines that train and run large AI models.

How the portfolio was assembled

AMD's position in the data center rests partly on its own product line and partly on acquisitions.

On 14 February 2022, AMD completed its acquisition of Xilinx in an all-stock transaction^[3]. Under the terms, Xilinx stockholders received 1.7234 AMD shares for each Xilinx share, plus cash in lieu of fractional shares^[4].

On 19 August 2024, AMD announced it would acquire ZT Systems, which it described as a hyperscale AI infrastructure provider, for $4.9 billion in cash and stock^[6]. AMD said it intends to keep ZT Systems' design and customer teams and to look to sell its manufacturing division^[6]. In other words, the system-design know-how and the customer relationships stay, while part of the assembly business is meant to go.

On the product side, AMD announced the Instinct MI300 series on 6 December 2023, with the MI300X offering 192 GB of HBM3 memory^[1].

What the chips look like

HBM3 is high-bandwidth memory: layers of memory stacked and placed close to the GPU, connected through a very wide interface, which is how it delivers high bandwidth at relatively low power.

The MI300X carries 192 GB of it, and an eight-MI300X Instinct platform provides 1.5 TB of HBM3 in total^[1]. Memory capacity is one of the numbers that decide what an AI system can do. Model weights, and the intermediate state that has to be cached during a long conversation, all have to sit in GPU memory, so the more memory a single chip has, the fewer chips a given model needs.

The MI300A takes a different route. Each one combines 24 Zen 4 CPU cores and 228 CDNA 3 GPU compute units with 128 GB of unified HBM3 memory^[7]. Because the CPU side and the GPU side read and write the same pool of memory, data does not have to be shuttled back and forth between separate memories.

AMD says that El Capitan, a supercomputer powered by Instinct MI300A accelerated processing units, is the fastest supercomputer^[7].

The OpenAI partnership

On 6 October 2025, AMD said that it and OpenAI had signed a multi-year, multi-generation partnership to deploy 6 gigawatts of AMD GPUs, starting with the AMD Instinct MI450 in 2026^[2]. TechCrunch reported that AMD would supply 6 GW of compute capacity to OpenAI in a chip deal worth tens of billions of dollars^[5].

Data center capacity is usually described in terms of power, because how much electricity a site can deliver sets how many racks of accelerators it can run. A gigawatt is one billion watts, so 6 GW describes a very large electricity footprint.

Where AMD sits

Read together, these announcements show AMD's offering getting broader: from a single chip such as the MI300X or MI300A, to a full eight-chip platform with 1.5 TB of memory, to multi-year capacity commitments measured in gigawatts. On the roadmap, the Instinct MI450 in 2026 is the next step on that line^[2].

For a general reader, three numbers are worth watching in this kind of news: how much memory a single chip has (192 GB on the MI300X, 128 GB of unified memory on the MI300A)^[1][7], how much memory a whole system has (1.5 TB on an eight-chip platform)^[1], and how much power a customer commits to (6 GW from OpenAI)^[2]. The first two set how large a model can fit; the third sets how far that scale can be spread.