AMD

AMD is a fabless semiconductor company — it designs chips but outsources manufacturing to TSMC. The company has undergone a dramatic transformation under CEO Lisa Su, evolving from a struggling Intel competitor into the world's second-largest AI accelerator company with $34.6B in annual revenue and a market cap exceeding $450B.

The business operates across four segments. Data Center is the growth engine — $16.6B in FY2025 revenue (+32%), encompassing Instinct GPU accelerators and EPYC server CPUs. Client (PC processors) contributed roughly $7.5B. Gaming (console/PC GPU) is declining. Embedded (acquired via Xilinx) provides stable FPGA revenue. Data Center now represents nearly 50% of total revenue and is growing fastest.

The GPU story has inflected. AMD's Instinct MI300X established credibility in 2024-2025, and the MI350 is now ramping with meaningful performance improvements. But the real transformation happened in early 2026: Meta announced a partnership to deploy up to 6 gigawatts of AMD GPUs using custom MI450-based accelerators optimized for Meta's workloads — a multi-year, multi-generation commitment that is the largest GPU deal in AMD's history. Oracle committed to 50,000 MI450 GPUs in Helios rack-scale deployments starting Q3 2026. OpenAI selected AMD as a preferred partner for H2 2026 training and inference. Eight of the top ten AI companies now use Instinct GPUs in production.

The MI450 architecture (codename "Altair"), paired with the Helios rack-scale design, AMD's next-gen EPYC "Venice" CPUs, and Pensando "Vulcano" networking, represents AMD's first full-stack AI infrastructure solution — competing not just at the chip level but at the system level. The Meta custom GPU effort is particularly notable: AMD customized the MI450 architecture for Meta's specific workloads without a full ASIC tape-out, blending the economics of a merchant GPU with the optimization of custom silicon.

AMD's EPYC server CPU continues taking share from Intel — now at approximately 35-40% of the server CPU market, up from 15% in 2019. EPYC is the "other AI story" within AMD: every AI server needs a host CPU, and EPYC's performance advantage over Intel Xeon means AMD captures revenue on both the GPU and CPU in many AI deployments.

The ROCm software stack — AMD's answer to NVIDIA's CUDA — remains the critical strategic variable. CUDA's 20-year ecosystem creates deep switching costs. But ROCm is maturing rapidly: it's fully open-source, PyTorch and major ML frameworks now support it with improving optimization, and AMD expects to approach CUDA parity in core capabilities by end of 2026. The Meta and OpenAI commitments suggest the software gap is narrowing to the point where it's no longer a dealbreaker.