NVIDIA
NVIDIA designs and sells graphics processing units, AI accelerator hardware, and the CUDA parallel computing platform that underlies virtually all major AI frameworks. Founded in 1993, the company has repositioned as the dominant supplier of compute infrastructure for artificial intelligence. Its core products include the H100/H200 (Hopper) and B100/B200 (Blackwell) data center GPUs and DGX server systems. Data Center now constitutes the overwhelming majority of revenue.
Value-Chain Position
NVIDIA sits at the compute infrastructure layer — the physical substrate on which AI models are trained and served. Foundation model developers such as Anthropic and OpenAI rent or purchase NVIDIA GPU clusters; cloud providers (AWS, Azure, Google Cloud) resell NVIDIA capacity as managed compute. The CUDA software ecosystem ties developer workflows and toolchain investment to NVIDIA hardware, creating switching costs that persist even when competing chips offer comparable raw performance.
Financing and Valuation
NVIDIA is publicly traded on NASDAQ (NVDA). For fiscal year 2025 (ending January 2025), NVIDIA reported total revenue of approximately $130.5 billion, up approximately 114% year over year, with Data Center revenue of approximately $115.2 billion (per SEC filings, February 2025). Q3 FY2026 (October 2025) reported revenue of approximately $57 billion with Data Center at approximately $51.2 billion, up approximately 66% year over year. As of late May 2026, market capitalization was reported at approximately $5.1–5.2 trillion. NVIDIA generates substantial free cash flow and requires no external financing.
Investment Thesis
Bull case. Structural demand for GPU compute continues to grow faster than supply, and NVIDIA has no credible rival at scale near term. The CUDA moat is a two-decade developer ecosystem. Hyperscaler AI infrastructure capex plans from Microsoft, Google, Amazon, and Meta — each committing hundreds of billions through 2026 and beyond — translate directly into NVIDIA GPU orders. NVIDIA also holds a reported $30 billion equity stake in OpenAI as part of OpenAI's March 2026 funding round.
Bear case. Revenue is heavily concentrated in a handful of hyperscaler customers. All four major hyperscalers are developing custom AI silicon (Google TPUs, Amazon Trainium/Inferentia, Microsoft Maia, Meta MTIA) with the stated goal of reducing NVIDIA dependency. AMD ROCm is maturing and AMD hardware is priced 25–40% below comparable Blackwell inference configurations. Export controls on advanced chips to China represent a material revenue headwind. Cerebras wafer-scale architecture has demonstrated substantially faster inference on large memory-bound workloads, a specialized competitive pressure in the inference tier.
Watch items. Hyperscaler custom silicon adoption curves; ROCm ecosystem maturation; U.S. export control policy; Blackwell-to-Rubin architecture cadence; whether inference-specialized alternatives such as Cerebras capture meaningful revenue share at frontier model labs.
Relationship to Portfolio
NVIDIA is the infrastructure dependency that makes both Cerebras and Anthropic relevant as investment theses. Cerebras is explicitly positioned as an inference-tier alternative to NVIDIA GPU clusters, targeting the memory-bandwidth bottleneck that limits NVIDIA throughput on large models. Anthropic is one of NVIDIA's largest indirect customers; Anthropic's growth directly drives demand for compute infrastructure NVIDIA provides.