Nvidia widens AI chip lead as AMD gains inference traction

Nvidia remains ahead in AI training GPUs, supported by a larger data-center business and stronger profitability. Its Data Center segment reportedly posted a $75 billion quarter, while adjusted gross margin stands near 75%, underscoring its scale and profitability gap with AMD. CUDA supports millions of developers and extensive industry tooling, creating switching costs that analysts say keep customers tied to Nvidia hardware.

The company’s product stack spans H100 and H200 accelerators as well as Blackwell B200 and B300 GPUs. Blackwell Ultra systems introduced in late 2025 include 288 gigabytes of HBM3e high-bandwidth memory, up from 192 gigabytes on the B200, supporting larger AI models and data sets. Amazon Web Services EC2 documentation lists P6 instances with up to eight Nvidia Blackwell B200 or B300 GPUs, signaling hyperscaler adoption for demanding AI workloads.

AMD is focusing on AI inference and agentic AI from a smaller data-center base. It reported $5.8 billion in data-center revenue in the first quarter of 2026, a 57% year-over-year increase largely driven by MI300 accelerators, while ROCm 7 and related tools aim to improve MI350-class training and inference performance. Broadcom, Marvell and hyperscaler-built accelerators such as Amazon’s Trainium, Google’s TPU and Meta’s designs are adding pressure, but Nvidia’s software moat and scale point to continued leadership in AI training.