AMD

AMD (Advanced Micro Devices, Inc.; NASDAQ: AMD) is a semiconductor company whose AI portfolio combines Instinct GPUs, EPYC server CPUs, ROCm software and Helios rack-scale systems. Led by chair and CEO Lisa Su, it has expanded through systems acquisitions and collaborations with AI developers. In August 2026, AMD announced an agreement to acquire Taalas, extending its accelerator roadmap toward specialized inference silicon.[9][13][2]

On August 6, 2026, AMD reached a definitive agreement to acquire Taalas, subject to customary closing conditions and regulatory approvals. Founded in 2023 and headquartered in Toronto, Taalas develops specialized AI inference silicon. AMD described its technology as optimizing inference dataflows to reduce compute and memory bottlenecks, and planned to integrate it into the accelerator roadmap and develop system-level solutions alongside Instinct GPUs.[2]

Taalas co-founder and CEO Ljubisa Bajic described the approach as “building the hardware around the model.” Vamsi Boppana, senior vice president of AMD’s Artificial Intelligence Group, positioned the technology within a platform that lets customers choose compute for different AI workloads.[2]

AMD and OpenAI announced a multiyear, multigeneration agreement for 6 gigawatts of Instinct GPUs on October 6, 2025. Their hardware and software collaboration began with MI300X and continued with MI350X; the agreement extended it to MI450 and future generations, including rack-scale AI solutions. The announcement scheduled the first 1-gigawatt MI450 deployment to begin in the second half of 2026.[3]

AMD issued OpenAI a warrant for up to 160 million common shares. Vesting was tied to deployment and purchase milestones, AMD share-price targets, and OpenAI technical and commercial milestones. The first tranche was to vest with the initial 1-gigawatt deployment, with further tranches as purchases scaled toward 6 gigawatts. CFO Jean Hu said the partnership was expected to deliver tens of billions of dollars in AMD revenue.[3]

In its August 4, 2026 earnings announcement, AMD reported launching Helios and the Instinct MI400 family, including MI455X, introducing sixth-generation EPYC CPUs, and releasing ROCm.ai as a developer experience for building, deploying and optimizing on AMD platforms. It also announced an Anthropic partnership for up to 2 gigawatts of MI450 GPUs in Helios racks, with multiyear collaboration using Claude to optimize Instinct and ROCm development. A separate Cerebras collaboration aimed to combine Helios with the Wafer-Scale Engine for ultra-low-latency inference serving.[13]

AMD completed its acquisition of ZT Systems on March 31, 2025. The transaction added rack-scale design and customer-enablement capabilities to AMD’s CPU, GPU, networking and ROCm portfolio. ZT’s design teams joined the Data Center Solutions business unit led by Forrest Norrod. Former ZT president Doug Huang became senior vice president of Data Center Platform Engineering, leading design and customer enablement; founder and former CEO Frank Zhang became senior vice president of ZT Manufacturing to help lead the manufacturing divestiture. Both reported to Norrod.[6]

The manufacturing divestiture closed on October 27, 2025, when AMD sold ZT Group Int’l to Sanmina. Sanmina issued AMD’s selling subsidiary 1,151,052 shares and paid, or would cause to be paid, $2.4 billion in cash, subject to purchase-price adjustments. The seller was also eligible for up to $450 million in additional cash if post-closing conditions were met. AMD’s subsequent earnings release identified the disposed operation as ZT Systems’ manufacturing business.[7][13]

Before these rack-scale systems and inference agreements, AMD’s December 2023 MI300 launch established two distinct accelerator designs and expanded the ROCm software ecosystem used to deploy them.[4]

AMD announced MI300X and MI300A availability on December 6, 2023. MI300X was a CDNA 3 GPU accelerator for workloads including LLM training and inference, with 192 GB of HBM3 memory and 5.3 TB/s peak theoretical memory bandwidth. The eight-accelerator Instinct Platform provided 1.5 TB of HBM3 in an industry-standard OCP design, allowing server partners to incorporate the accelerators into existing AI offerings.[4]

MI300A combined CDNA 3 GPU cores, Zen 4 x86 CPU cores and 128 GB of HBM3 in a single package, with unified memory and cache resources. This CPU/GPU APU design served HPC and AI workloads, distinguishing it from the GPU-only MI300X and the multi-accelerator Instinct Platform.[4]

AMD paired the launch with ROCm 6, adding support for FlashAttention, HIPGraph and vLLM. It invested in software through the acquisitions of Nod.AI and Mipsology and partnerships with Lamini and MosaicML. AMD described Lamini as running LLMs for enterprise customers and MosaicML as enabling LLM training on Instinct through ROCm with zero code changes.[4]

AMD collaborated with IBM Research and Red Hat on a Triton-based attention backend for vLLM. In March 2026, the vLLM team at IBM Research credited the three organizations with developing and upstreaming the backend, while identifying IBM Research and Red Hat AI as its initial developers. The backend ran the same source code on NVIDIA, AMD and Intel GPUs, depended only on PyTorch and Triton, and was the default attention backend on AMD GPUs running ROCm.[5]

The kernel grouped query heads sharing a key/value head to increase cache reuse, and grouped multiple query tokens into Q blocks to improve matrix-tile utilization. For decode workloads, where a single query token limits that approach, parallel tiled softmax split traversal of the KV cache across kernel instances. A second kernel reduced their partial results, trading greater parallelism against extra launch overhead. The team also designed persistent kernels with fixed launch grids and dynamic work assignment from GPU metadata to improve CUDA graph reuse; those pull requests were still pending in the March 4 account.[5]

The contributors reported late-2025 end-to-end latency tests using Llama 3.1 8B, batch size one, 500 input tokens and varying output lengths. On MI300, they reported approximately 5.8× improvement over earlier implementations; on H100, long decode requests reached 100.7% of FlashAttention 3 performance. These were results for the tested backend and workload.[5]

AMD’s CPU, GPU and adaptive-computing portfolio predates its recent AI systems expansion. The ATI acquisition added graphics technology in 2006; Zen-based Ryzen and EPYC processors and the Radeon Instinct MI25 followed in 2017, and Xilinx added adaptive-computing products in 2022.[10][9][8]

Advanced Micro Devices was incorporated in Delaware on May 1, 1969, and became publicly held in 1972.[9]

AMD acquired ATI Technologies in October 2006, combining its microprocessor business with ATI’s graphics, chipsets and consumer-electronics technology. The announcement valued the transaction at approximately $5.4 billion, excluding assumed equity awards. AMD financed the cash portion with existing funds and new debt, including $2.5 billion in term loans. It also announced Fusion, a design initiative for x86 processors integrating CPU and GPU at the silicon level.[10][11]

AMD’s fiscal 2007 report recorded a $5.604 billion ATI purchase price, including vested equity awards and acquisition costs, and said the financing term loan had been fully repaid in 2007. That year it recorded approximately $1.3 billion in acquisition-related goodwill impairment and $349 million in developed-technology impairment. AMD attributed the charges to reduced profitability and revenue forecasts.[12]

AMD formed GlobalFoundries through transactions with Mubadala Tech and WCH on March 2, 2009. It transferred its remaining GlobalFoundries shares on March 4, 2012, while continuing a wafer-supply relationship. By fiscal 2017, third-party foundries fabricated silicon wafers for all AMD products.[9]

In the third quarter of 2016, AMD completed a public offering of 115 million common shares at $6 per share, producing approximately $667 million in net proceeds after underwriting discounts and expenses. Its 2016 convertible-note issuance produced a further $782 million in net proceeds; debt repurchases and credit-line repayments offset much of the financing inflow.[9]

In 2017, AMD launched its first-generation Ryzen desktop CPUs using the Zen x86 architecture. EPYC 7000 server processors followed in June with up to 32 Zen cores. AMD also launched Radeon Instinct MI25 for data-center machine learning, deep learning and AI. Its product development included CPU/GPU integration through APUs and parallel GPU computing beyond graphics. AMD reported approximately 8,900 employees as of December 30, 2017.[9]

AMD completed its all-stock acquisition of Xilinx on February 14, 2022, following the October 2020 announcement. The acquisition added FPGAs, adaptive SoCs, AI engines and software expertise. Former Xilinx CEO Victor Peng became president of AMD’s newly formed Adaptive and Embedded Computing Group, responsible for FPGA, adaptive SoC and software roadmaps.[8]

For the quarter ended June 27, 2026, AMD reported revenue of $11.5 billion, GAAP operating income of $2.0 billion and GAAP net income of $2.3 billion. Data Center revenue was $6.7 billion, up 107% year over year, and represented 58% of company revenue. That segment included EPYC server CPUs and other data-center products as well as AI accelerators. Client and Gaming generated $3.8 billion, and Embedded generated $977 million.[13]