NVIDIA
High-performance GPU computing platform and software ecosystem for AI training and inference.
CUDA software ecosystem + leading architecture iteration + deep integration with TSMC advanced packaging, forming a triple moat across hardware, software, and manufacturing.
Company Overview
NVIDIA was founded in 1993 and is headquartered in Santa Clara, California, USA. The company initially focused on the design and sales of graphics processing units (GPUs). With the introduction of CUDA (Compute Unified Device Architecture) in 2006, it expanded GPUs from graphics rendering to general-purpose computing. After the deep learning revolution broke out in 2012, NVIDIA GPUs became the core hardware for AI training and inference thanks to their powerful parallel computing capabilities, quickly establishing market dominance. From 2023 to 2026, global demand for AI computing power grew explosively, and NVIDIA maintained an absolute lead in the data center GPU market. The latest Blackwell B300 series is expected to ship in volume in the second half of 2026, further consolidating its technological edge. The company's business now spans multiple areas, including data center GPUs, automotive, professional visualization (ProViz), and gaming, with data center revenue accounting for more than 88% of total revenue.
The data center business is NVIDIA's absolute core engine, covering GPU accelerator cards for AI training and inference (such as H100/H200, B200/B300), Grace Hopper CPUs, networking equipment (Mellanox InfiniBand/Spectrum-X Ethernet), and DGX series complete server systems. The customer base spans all major global cloud service providers (AWS, Azure, GCP, Meta, Oracle) and large enterprises. The gross margin for this business remains above 75%, and with the Blackwell architecture upgrade, the unit price per card has risen from approximately $30,000 for the H100 to $50,000-$60,000 for the B200, continuously driving revenue growth. In Q2 of fiscal 2026 (as of July 2026), data center revenue exceeded $28 billion, up 150% year-over-year.
The GeForce RTX series GPUs target consumer-grade gaming, creative production, and light AI applications. The RTX 40/50 series incorporates Tensor Cores and AI acceleration units, supporting AI-enhanced technologies such as DLSS and frame generation. This business is subject to PC cyclical fluctuations but remains one of the company's cash flow stabilizers.
The Quadro/RTX series addresses professional scenarios such as workstations, rendering, and virtual simulation. The automotive business (autonomous driving chips Orin/Thor) contributes approximately 2% of revenue, with intelligent cockpit and autonomous driving computing platforms serving as long-term growth points.
| Product Line | Revenue Share | Core Customers | Gross Margin |
|---|---|---|---|
| Data Center GPU/Systems | ~88% | Cloud service providers, Enterprises | ~75% |
| Gaming GPU | ~6% | Consumers, DIY enthusiasts | ~60% |
| Professional Visualization | ~3% | Design, Film/TV, Engineering | ~65% |
| Automotive | ~2% | Automakers (NIO, BYD, etc.) | ~55% |
Technical Moat
Moat 1: Irreplaceability of the CUDA Ecosystem
Since its launch in 2006, CUDA has accumulated over 20 years of compatibility, with 4.5 million developers worldwide building AI frameworks (PyTorch, TensorFlow, JAX, etc.) on CUDA. All mainstream AI model training and inference deeply depend on CUDA libraries (such as cuDNN, TensorRT, NCCL). Although AMD ROCm continues to catch up, the gap in compatibility, stability, and performance remains at least 3–5 years. NVIDIA has also been continuously adding AI-specific instructions through CUDA 12.x versions (such as FP8 Transformer Engine and sparsity support), making it difficult for competitors to replicate.
Moat 2: Leading GPU Architecture Iteration Speed
NVIDIA maintains a cadence of launching a new architecture every two years: Ampere (2020) → Hopper (2022) → Blackwell (2024) → Blackwell Ultra/B300 (2026). Each generation delivers significant improvements in compute unit count, interconnect bandwidth (NVLink, NVSwitch), and memory capacity (HBM3e/4). The Blackwell B300 adopts TSMC's CoWoS-L advanced packaging, integrating up to 12 HBM4 stacks with memory bandwidth reaching 8 TB/s. This rapid iteration cycle leaves competitors perpetually chasing the previous generation.
Moat 3: Deep Integration with Supply Chain and Advanced Packaging
NVIDIA has established a deeply exclusive partnership with TSMC, occupying over 60% of TSMC's CoWoS capacity. By 2026, CoWoS capacity has expanded to 50,000 wafers per month (12-inch equivalent), with NVIDIA having locked in the majority of capacity through 2027 in advance. Meanwhile, the company's in-house NVLink interconnect technology and Mellanox networking solutions build a full-stack competitive defense spanning from compute to networking.
| Dimension | Data |
|---|---|
| Global AI training GPU market share | ~85% |
| Global AI inference GPU market share | ~70% |
| Industry ranking | No. 1 |
| Main competitors | AMD (MI300X), Intel (Gaudi 2/3), Huawei (Ascend 910B/920) |
| Key downstream customers | AWS / Microsoft / Google / Meta / Oracle / Elon Musk's xAI, etc. |
| Substitution threat | Custom AI chips (TPU, Trainium, LPU) are penetrating specific scenarios, but their versatility falls short of GPUs |
NVIDIA holds an absolute advantage in the HPC/AI chip market, but competitors are narrowing the gap: AMD's MI300X has secured partial orders from Microsoft and Meta; Huawei's Ascend 910B has gained share in restricted domestic scenarios. Nevertheless, NVIDIA maintains a gross margin of over 80%, backed by its software ecosystem and full-stack solutions (DGX + InfiniBand).
Financial Performance & Growth
| Metric | Data |
|---|---|
| Revenue (FY2025) | USD 130.497 billion |
| Gross Margin | 73.5% (FY2025) |
| Net Margin | 55.1% (FY2025) |
| FY2026 Expected Revenue | ~USD 180 billion (analyst consensus) |
| Core Growth Thesis | Data center GPU per-card ASP uplift + inference demand surge + rising enterprise customer penetration |
Growth Drivers:
- Blackwell B200/B300 Volume Ramp: In 2026, B300 is expected to contribute over 60% of data center revenue, with per-card ASP 2x higher than the H-series.
- Inference Demand Surge: AI application deployment (GPT-5, Claude 4 on-device inference, etc.) is driving inference GPU demand. NVIDIA has launched dedicated inference solutions such as the L40S and Grace Hopper.
- Global AI Arms Race: Cloud vendors' CapEx will exceed USD 250 billion in 2026, with GPU procurement accounting for approximately 60%.
- Automotive & Robotics: Thor SoC is entering mass production and shipment, offering significant long-term growth potential.
Key Risks:
- Technology Substitution Risk: Specialized AI chips (e.g., Cerebras wafer-scale chips, Groq LPU) may deliver superior cost-performance in specific scenarios, eroding the moat of GPU versatility.
- Geopolitical and Export Control Risk: The US government continues to tighten AI chip export restrictions to China, costing NVIDIA roughly 20% of its addressable market and accelerating the rise of competitors such as Huawei and AMD.
- Supply Chain Concentration Risk: Any shortfall in TSMC's CoWoS capacity expansion will constrain B300 supply; Samsung's HBM4 supply stability remains uncertain.
- Demand Cyclicality Risk: If AI capital expenditure growth decelerates, cloud providers may reduce orders, leading to inventory buildup and pricing pressure.
- Antitrust and Compliance Risk: The US Department of Justice and the EU are investigating NVIDIA's bundling and exclusive procurement practices, which could cause short-term share price volatility.
Core Investment Thesis / Industry Value Summary:
NVIDIA is the most critical "shovel seller" in the AI infrastructure layer. Leveraging its CUDA ecosystem, generational architectural leadership, and supply chain lock-in, it currently captures over 80% of profits in the global AI compute market. Despite competitive catch-up and geopolitical frictions, its data center business is expected to sustain over 30% CAGR over the next 3-5 years, establishing NVIDIA as the highest-certainty value benchmark in the AI industry chain. Investors should closely monitor the Blackwell B300 volume ramp-up pace and shifts in export control policy.