NVIDIA/AMD Inference Outlook

Jensen Huang is telling investors the next wave of artificial intelligence spending is inference — the step where trained models answer real user requests at scale. (nvidianews.nvidia.com) Inference is the serving side of artificial intelligence: a model is already built, and chips generate tokens for chatbots, agents, search tools, and software copilots every time a user clicks send. In NVIDIA’s February 25, 2026 earnings release, Huang said “computing demand is growing exponentially” and called this moment an “agentic AI inflection point.” (nvidianews.nvidia.com) NVIDIA is making that case with current numbers. The company reported fiscal fourth-quarter revenue of $39.3 billion and fiscal 2026 revenue of $130.5 billion, with data center revenue of $35.6 billion in the quarter and $115.2 billion for the year. (nvidianews.nvidia.com) AMD is in the same argument because it has been pitching its Instinct graphics processors and ROCm software as lower-cost options for running large models after training is done. AMD reported full-year 2025 data center revenue of $16.6 billion, up 32% from 2024, driven by EPYC central processors and Instinct graphics processors. (ir.amd.com) The market chatter centers on cost per token, which is the price of producing each chunk of model output. NVIDIA said in February that Grace Blackwell with NVLink delivers “an order-of-magnitude lower cost per token,” while AMD has pointed to its own MLPerf Inference submissions on Instinct MI325X and MI355X systems to argue it can compete on serving workloads. (nvidianews.nvidia.com) (rocm.blogs.amd.com) That matters in 2026 because enterprise buyers are moving from one-time model training runs to repeated, always-on usage inside products and internal tools. Huang said “enterprise adoption of agents is skyrocketing,” and NVIDIA guided for another sequential revenue increase in the April quarter. (nvidianews.nvidia.com) (finance.yahoo.com) The valuation part of the debate is simpler than the hardware part. As of April 2026, CompaniesMarketCap listed NVIDIA at about 38 times trailing earnings and AMD at about 81 times, which undercuts the claim that AMD is obviously the cheaper stock on a trailing price-to-earnings basis. (companiesmarketcap.com 1) (companiesmarketcap.com 2) There is also a scale gap. NVIDIA’s fiscal 2026 data center revenue of $115.2 billion is roughly seven times AMD’s 2025 data center revenue of $16.6 billion, which is why investors still treat NVIDIA as the incumbent in artificial intelligence infrastructure. (nvidianews.nvidia.com) (ir.amd.com) AMD’s bull case rests on share gains, not current leadership. The company’s investor relations site lists new enterprise partnerships with Meta, Nutanix, Hewlett Packard Enterprise, and Tata Consultancy Services in early 2026, all aimed at expanding deployment of AMD graphics processors and rack-scale systems. (ir.amd.com) The thread running through both companies is the same one Huang is pushing: if artificial intelligence tools are used more often, the bottleneck shifts from building models to running them cheaply and fast. The next earnings reports from AMD and NVIDIA will show whether that inference demand is widening the field or reinforcing NVIDIA’s lead. (nvidianews.nvidia.com)