AMD gains ground on Nvidia, but revenue gap keeps widening

AMD is becoming a serious alternative to Nvidia for AI infrastructure, supported by competitive inference performance and major customer agreements. OpenAI and Meta have each signed deals covering up to 6 gigawatts of AMD GPU deployments, while Oracle plans a 50,000-GPU MI450 supercluster. These commitments span multiple years and depend on purchasing milestones, rather than representing capacity already installed.

The financial gulf continues to widen. AMD reported quarterly Data Center revenue of $6.72 billion, compared with Nvidia’s $89 billion. Both businesses more than doubled year over year, but Nvidia added substantially more revenue. The comparison is imperfect: AMD’s segment includes server processors, while Nvidia’s includes networking and other infrastructure products.

AMD’s strongest technical progress is in inference. Its MI355X beat selected Nvidia B200 and B300 configurations on GPT-OSS-120B in MLPerf Inference v6.1. ROCm software improvements increased server throughput by 38% without hardware changes, while a 72-GPU configuration achieved 95% scaling efficiency. These results demonstrate workload-specific competitiveness, not universal leadership or consistently lower production costs.

Nvidia retains advantages in frontier-model training, distributed infrastructure and its established CUDA ecosystem. AMD must demonstrate reliable training at scale and convert deployment plans into recurring revenue. Its customer agreements also carry significant equity incentives: OpenAI and Meta each received warrants to purchase up to 160 million AMD shares at $0.01 per share, subject to milestones. Repeated adoption by Microsoft, Oracle and Meta strengthens AMD’s position, but does not establish parity with Nvidia.