AMD

AMD makes processors for most of the places software runs, and in most of them it is the main alternative to a larger rival.

In the Data Center segment, EPYC server CPUs compete with Intel's Xeon and have led it on performance per socket for several generations, while Instinct GPUs compete with NVIDIA for AI training and inference. AMD is now selling those pieces together: in 2025 it bought ZT Systems' design team to build full AI racks, and its Helios rack, due in 2026, combines Instinct GPUs, EPYC CPUs and AMD's Pensando networking.

In Client and Gaming, Ryzen processors compete with Intel Core in desktops and laptops, Radeon graphics cards with NVIDIA's GeForce, and the semi-custom business supplies the chips in PlayStation 5 and Xbox Series consoles. Embedded is mostly Xilinx, bought in 2022: FPGAs and adaptive SoCs for cars, telecom, aerospace and industry, plus embedded versions of EPYC and Ryzen.

The technology that ties it together is the chiplet. AMD builds processors from several small dies in one package, mixing TSMC's newest process for compute with cheaper ones for I/O, and stacking dies on top of each other (3D V-Cache in Ryzen and EPYC, compute on I/O in Instinct). It reuses the same Zen CPU cores and graphics designs across servers, PCs, consoles and embedded parts, so one architecture serves many markets. Owning no factories means AMD's products are as good as TSMC's processes, without the cost of building them.

The software layer matters more each year. ROCm, AMD's open-source GPU software, is how it competes with NVIDIA's CUDA, and AMD has bought several software teams to improve it.

In CPUs, switching costs are low by design: EPYC and Ryzen run the same x86 software as Intel, so a customer can move either way with a new server or laptop. What keeps customers is performance per dollar and per watt, generation after generation, and the platform: server makers build boards around EPYC sockets and PC builders around AM5. That means AMD has to keep winning on the product; it cannot count on lock-in.

In AI GPUs, the pull runs the other way. NVIDIA's CUDA holds most AI developers, and the cost of leaving it, rewriting kernels and retuning models, is what AMD has to overcome. Once a customer has done that work and built a cluster on ROCm, as the largest buyers have, it has a reason to keep buying each new Instinct generation.

The stickiest products are the semi-custom chips and FPGAs. A console uses one chip for its whole life, and backward compatibility ties the next console to the same architecture. An FPGA design built in Vivado and qualified for a car or an aircraft stays for a decade; moving to Altera or Lattice means redesigning and requalifying the product.

Revenue was $34.6B in fiscal 2025, up 34.3%, after $25.8B in fiscal 2024 and $22.7B in fiscal 2023. Ten years earlier, in fiscal 2016, it was $4.3B, with a gross margin of 23.2%. Over the five years from fiscal 2020, revenue grew about 28.8% a year, driven by EPYC, the Xilinx acquisition in 2022 and, from 2024, Instinct GPUs.

Gross margin was 49.5% in fiscal 2025, up from 44.5% in fiscal 2020 as data center chips grew in the mix. Operating margin was lower, 10.7% ($3.7B of operating income), because the Xilinx and other acquisitions add large amortisation charges, and because AMD spends heavily on R&D to release new CPUs and GPUs every year.

As a fabless company, AMD spends little on equipment: capital spending was $974M, 2.8% of revenue. Operating cash flow was $7.7B and free cash flow $6.7B, 19.4% of revenue. Equity pay was $1.6B, 4.7% of revenue; after it, free cash flow was $5.1B, 14.7%. AMD ended fiscal 2025 with about $6.5B more cash and short-term investments than debt.

AMD is turning from a chip supplier into a supplier of full AI systems. Instinct GPUs now ship on an annual cadence, and the Helios rack, due in 2026 with the MI450, puts AMD's GPUs, CPUs and networking in one rack-scale product to compete with NVIDIA's NVL72 systems. OpenAI's agreement to deploy six gigawatts of AMD GPUs, starting with MI450, is the first commitment on that scale. The ZT Systems design team builds the racks; AMD sold the manufacturing part of ZT to Sanmina and relies on partners to assemble them. Software comes with it: ROCm releases more often, and AMD bought teams in inference software and co-packaged optics in 2025.

In CPUs, AMD keeps moving EPYC and Ryzen to newer TSMC processes and more stacked cache. In PCs, Ryzen AI chips with larger NPUs and integrated GPUs aim at laptops that run AI models locally.

The pattern is the same in each market: AMD competes as the alternative to a larger incumbent, Intel in CPUs and NVIDIA in GPUs, and grows when its product is ahead.

• NVIDIA's lead in AI. CUDA, NVLink, networking and full-rack systems give NVIDIA years of head start and most of the market. AMD has to match the hardware and close the software gap at the same time, and the Helios rack is its first full answer.
• ROCm's maturity. Every model and kernel that runs worse or later on AMD costs a sale. Software engineering at this scale is new for a company that grew as a hardware maker.
• Intel and Arm in CPUs. Intel's 18A chips aim to close the gap with EPYC and Ryzen, while cloud companies' own Arm chips (Graviton, Axion, Cobalt) take workloads that once ran on x86.
• Dependence on TSMC for every leading product, in Taiwan, competing for the same wafers and packaging capacity as NVIDIA and Apple.
• Export rules that limit AI GPU sales to China and can change with little notice.
• Mature markets. Consoles follow a cycle, PCs grow slowly, and the embedded FPGA business swings with customers' inventories.