AMD — Advanced Micro Devices

Advanced Micro Devices (AMD) is a fabless semiconductor company that designs server and client CPUs, datacenter AI accelerators, GPUs, FPGAs, and data-processing units. It is the principal merchant-GPU competitor to Nvidia for AI training and inference at scale, and for practical purposes the only one with material share. Under CEO Lisa Su, who joined in 2014, AMD developed a server-CPU line (EPYC) and an AI accelerator line (Instinct MI).

Overview

Snapshot

Financials and market

Business and products

AMD's portfolio spans datacenter accelerators, CPUs, GPUs, and programmable logic:

• Instinct MI300X / MI325X / MI350X — datacenter AI accelerators.
• EPYC server CPUs (Zen 4, Zen 5) — widely deployed in hyperscaler CPU fleets.
• Ryzen client CPUs.
• Radeon consumer and professional GPUs.
• Xilinx FPGAs and adaptive SoCs — acquired 2022 for $49B; serve networking, telecom, edge inference, and datacenter acceleration markets.
• Pensando DPUs — acquired 2022.

In market position, AMD is the second-largest supplier of merchant AI accelerators behind Nvidia and ahead of other merchant vendors; the first or second supplier of server CPUs by share, gaining against Intel; and the leading FPGA supplier through Xilinx, ahead of Altera/Intel.

AI accelerators and the compute supply chain

The Instinct MI family is AMD's direct competitor to Nvidia's datacenter GPUs: the MI300X (2023), MI325X (2024), and MI350X (2025–2026) line up against Nvidia's H100, H200, and Blackwell parts. AMD said on July 20, 2026 that Helios, its first rack-scale AI system and a direct rival to Nvidia's rack-scale offerings, will ship later in 2026 to Microsoft, Meta, OpenAI, and other customers at an estimated $5 million-plus per rack; an expanded Azure deployment pairs MI455X GPUs with Venice CPUs and Pensando networking (Source: cnbc.com). CEO Lisa Su formally unveiled Helios at the company's Advancing AI conference on July 23, 2026, calling it the industry's "highest performance AI rack," naming OpenAI, Meta, Oracle, Anthropic, and Microsoft among customers, and projecting the AI accelerator market would reach roughly $1.4 trillion by 2030 (Source: techcrunch.com). Two days later, on July 22, 2026, AMD and Anthropic announced a strategic partnership under which Anthropic will deploy up to 2 gigawatts of Instinct MI450-series GPUs in Helios rack-scale systems beginning in the first half of 2027, AMD will make a strategic equity investment of up to $5 billion in Anthropic, and Claude will be used to accelerate ROCm software development (Source: newsroom.amd.com; cnbc.com). AMD's architectural approach has been to provide larger HBM capacity per package than the corresponding Nvidia part, trading some raw compute for memory capacity, a profile suited to inference of large models. SemiAnalysis detailed the MI455X on July 25, 2026 as the first 2-nanometer datacenter accelerator, carrying 3,470 mm² of logic silicon on a 5.5×-reticle CoWoS-L module, 20 PF FP8 against Nvidia Rubin's 17.5 PF, and 12 HBM4 stacks for 432 GB per package and 23.3 TB/s of bandwidth against 8 stacks and 288 GB for Nvidia and Google. The analysis named unstable internal development clusters and vLLM parity below the 90% target as the principal execution risks to the Anthropic deployment (Source: semianalysis.com). See Inference Economics and Token Pricing.

AMD shares Nvidia's upstream dependencies. All leading-edge AMD silicon is fabricated at TSMC on advanced nodes (the MI300 series on N5/N4 variants) and uses TSMC's CoWoS advanced packaging. AMD buys high-bandwidth memory from SK Hynix, Samsung, and Micron, competing directly with Nvidia for the constrained HBM supply documented in (Source: Raw Sources/SemiAnalysis - CoWoS and HBM Supply Chain.md). Because both companies depend on TSMC fabrication, ASML lithography, and SK Hynix/Samsung HBM, a Semiconductor Supply Chain disruption affects AMD and Nvidia symmetrically.

AMD has guided to roughly $5B+ in datacenter GPU revenue in 2026, small against Nvidia's run rate of more than $100B but, in the company's framing, the first material share taken from Nvidia in the segment.

Taalas acquisition

AMD announced a definitive agreement on August 6, 2026 to acquire Taalas, a Toronto startup founded in 2023 that builds processors tailored to a single model's weights by finalising a small number of a chip's metal layers after the model is fixed. Vamsi Boppana, senior vice president of AMD's Artificial Intelligence Group, said the technology "strengthen[s] our AI portfolio by delivering differentiated inference performance and efficiency," and AMD said it will integrate the technology into its accelerator roadmap alongside Instinct GPUs. Taalas had raised $219 million in total. Its first product, unveiled in February 2026, hard-wires Meta's Llama 3.1 8B and is claimed by the company to run at 17,000 tokens per second per user — a vendor figure achieved partly through a custom 3-bit quantisation that Taalas concedes degrades output quality. The deal remains subject to customary closing conditions and regulatory approvals, with no closing date given and no price disclosed (Source: unite.ai). Model-specific silicon trades the generality that makes a GPU reusable across model generations for throughput on one fixed set of weights, which places the bet on inference volume for stable models rather than on training flexibility. See Inference Economics and Token Pricing, Etched.

AMD has also released open-weight models trained on its own accelerators. An open-model review published August 2, 2026 recorded Instella-MoE-16B-A3B-Think, a 16-billion-parameter mixture-of-experts model with 3 billion active per token trained on Instinct cards, with base, supervised-fine-tuning, MidTrain and DPO checkpoints all released (Source: interconnects.ai). See Open-Weight Frontier Models.

Software stack: ROCm

ROCm (Radeon Open Compute) is AMD's alternative to Nvidia's CUDA software stack. The maturity gap with CUDA has historically been the binding constraint on AMD's AI competitiveness, since developers are predominantly fluent in CUDA rather than ROCm. From 2024 to 2026 AMD increased ROCm investment and collaborated with hyperscalers, with Microsoft, Meta, and Oracle each deploying MI300-class fleets. The gap with CUDA narrowed over this period but remained, by AMD's own account, real.

Geopolitics and export controls

AMD's MI300-class accelerators are subject to the same performance-threshold-based export controls as Nvidia's. AMD created China-specific SKUs, such as the cut-down MI308, that have faced the same licensing denials as Nvidia's H20. AMD's role as a second source for AI accelerators is relevant to US AI supply resilience, given that a single-supplier market concentrates risk in one vendor.

Financial performance

For the first quarter of 2026, AMD reported revenue of $10.3 billion, up 38% year over year, and guided to 46% growth in the second quarter. On the same earnings call, Lisa Su told analysts that AMD's data-center CPU total addressable market had doubled to $120 billion by 2030. Shares rose 16% after-hours on the print, lifting the company's market capitalization to roughly $580 billion (Source: theinformation.com).

History

AMD was founded in 1969. Lisa Su became CEO in 2014, when the company was financially distressed; over the subsequent decade it returned to competitiveness as a fabless design house with a leadership-class server-CPU line (EPYC) and an AI accelerator line (Instinct MI). The Xilinx (FPGAs and adaptive SoCs) and Pensando (DPUs) acquisitions, both in 2022, broadened the portfolio into programmable logic and data-processing units; the Xilinx deal was valued at $49B.

Relationships

• contradicts: Nvidia & TSMC — AI Compute Infrastructure — primary competitor, though dependency structures overlap.
• depends-on: Semiconductor Supply Chain, SK Hynix — HBM Leader, Samsung Semiconductor, Micron Technology.
• related: Intel, Specialty Inference Hardware — Cerebras, Groq, SambaNova, Export Controls (AI), Compute Governance, AI Race Dynamics.