AMD

AMD (Advanced Micro Devices) is one of the world's largest semiconductor companies, founded in 1969 in Sunnyvale, California, and today headquartered in Santa Clara. Under CEO Lisa Su (since 2014), AMD has gone from struggling x86 second-source to credible cross-stack AI competitor — the only company with at-scale silicon across all three major AI compute surfaces: datacenter GPU, server CPU, and client NPU.

AMD's AI hardware portfolio is built around four product lines. The Instinct datacenter accelerators — currently the MI355X (CDNA 4 architecture, 288 GB HBM3e, generally available Q3 2025) — are AMD's direct response to NVIDIA's Blackwell, with the MI400 series (MI430X / MI440X / MI455X), launched in July 2026 inside AMD's first rack-scale system, Helios — 72 MI455X GPUs per rack — and targeting NVIDIA's Vera Rubin generation, with Microsoft and Anthropic among the launch customers and shipping in the second half of 2026. ROCm is AMD's open, permissively licensed software stack (currently 7.2.2) for AI compute on Instinct and select Radeon GPUs — the deliberate counterweight to NVIDIA's CUDA moat, and the vLLM CI pass rate jumping from 37% (Nov 2025) to 93% (Jan 2026) is the most-cited proof point of its maturation. Ryzen AI brings on-device AI to laptops and mini-desktops via the XDNA 2 NPU; the Ryzen AI Max+ 395 (Strix Halo, January 2025) hits 50 TOPS NPU and 126 TOPS system-wide with up to 128 GB unified memory. And EPYC server CPUs — currently the 9005 "Turin" generation, with Venice (Zen 6, up to 256 cores, H2 2026) — pair full-width AVX-512 with the new x86 AI Compute Extensions (ACE) standard AMD co-authored with Intel via the x86 Ecosystem Advisory Group in April 2026.

Major 2025-2026 customer wins validate AMD's position as the credible #2 datacenter AI vendor: Oracle Cloud Infrastructure went generally available on MI355X (with a 50,000-GPU MI450 commitment starting Q3 2026), Microsoft Azure is running MI300X in production, OpenAI signed a 6-gigawatt multi-year supply deal in October 2025, and Meta and xAI have publicly committed to MI350-class deployments. At roughly 28,000 employees AMD remains a fraction of NVIDIA's scale — but the cross-stack footprint plus the ACE standards play give it a structural position no other challenger has.

On August 6, 2026 AMD announced a definitive agreement to acquire Taalas, a Toronto startup founded in 2023 by former Tenstorrent chief executive Ljubisa Bajic, which etches model weights directly into silicon instead of streaming them from high bandwidth memory. Terms were not disclosed and the deal is expected to close by the fourth quarter of 2026 subject to regulatory approval. AMD plans to sell the technology as system-level solutions alongside its Instinct accelerators, giving it a dedicated inference answer to NVIDIA's own specialization moves — at the cost of the flexibility a general-purpose accelerator retains when models change.