NVIDIA -- How the Business Works
NVIDIA is a fabless semiconductor company that designs GPU architectures
and licenses them for manufacture by TSMC. The company generates revenue by selling
accelerated computing hardware, networking equipment, and software/services to hyperscale
cloud providers, enterprises, gamers, and automakers. The CUDA software ecosystem -- with
5.9M+ developers and 4,400+ accelerated applications -- creates deep switching costs that
lock customers into the NVIDIA platform across hardware generations.
CY2025 Revenue
$215.9B
+65% YoY
Gross Margin
75.0%
Q4 CY2025 GAAP
Data Center Share
~92%
AI accelerator market
CUDA Developers
5.9M+
Ecosystem lock-in
Revenue by segment -- Q4 CY2025
Total quarterly revenue: $68.1B
Data Center
$62.3B (91.5%)
Gaming
$3.7B (5.5%)
Professional Visualization
$1.3B (1.9%)
Automotive
$0.6B (0.9%)
OEM / IP
$0.2B (0.2%)
Segment data from NVIDIA Q4 FY2026 (CY2025) earnings. Source: Daloopa.
| Segment | Revenue | % of Total | Market Share | Key Customers |
|---|---|---|---|---|
| Data Center | $62.3B | 91.5% | ~92-97% | Microsoft, Google, Amazon, Meta, Oracle |
| Gaming | $3.7B | 5.5% | ~80% | Consumers, OEMs (Dell, HP, Lenovo) |
| Professional Viz | $1.3B | 1.9% | ~90% | Enterprise workstation users |
| Automotive | $0.6B | 0.9% | ~20-30% | Mercedes-Benz, BYD, robotaxi firms |
| OEM / IP | $0.2B | 0.2% | — | Licensing, legacy |
Business model flow
Step 1
Design GPU Architecture
NVIDIA designs chip architectures (Hopper, Blackwell, Rubin) and the full
networking stack (NVLink, NVSwitch, ConnectX). R&D spend: $15.4B in CY2025.
▼
Step 2
Fabless Manufacturing (TSMC)
NVIDIA owns no fabs. TSMC manufactures all chips on leading-edge nodes (4nm, 3nm).
This asset-light model enables 75% gross margins -- among the highest in semis.
▼
91.5% of Revenue
Hyperscalers & Enterprise
GPU clusters (H100, B200, GB300), DGX systems, networking.
$30K-$40K per GPU. Sold to MSFT, GOOG, AMZN, META, ORCL.
5.5% of Revenue
Gamers & Creators
GeForce RTX GPUs sold through retail and OEMs.
Annual upgrade cycle driven by new architecture launches.
3.0% of Revenue
Auto, ProViz & Other
DRIVE Orin/Thor for autonomous vehicles, RTX workstation GPUs,
OEM licensing. Emerging growth vectors.
▼
The Moat
CUDA Ecosystem → Platform Lock-In → Annual Upgrade Cycle
5.9M+ developers trained on CUDA. 4,400+ accelerated applications (TensorRT, Triton,
NeMo, NIM). Switching to AMD ROCm requires rewriting code, retraining teams, and
accepting inferior tooling. This creates an annual upgrade cycle where existing customers
buy the next-gen GPU rather than switch vendors.
Platform approach: hardware + software + services
Hardware
GPUs: H100, B200, GB300 NVL72
Systems: DGX SuperPOD, HGX
Networking: NVLink, NVSwitch, ConnectX-8, Spectrum-X
Roadmap: Annual cadence -- Blackwell (2024) → Rubin (2026) → Feynman (2028)
Systems: DGX SuperPOD, HGX
Networking: NVLink, NVSwitch, ConnectX-8, Spectrum-X
Roadmap: Annual cadence -- Blackwell (2024) → Rubin (2026) → Feynman (2028)
Software
CUDA: 5.9M developers, industry standard
Libraries: TensorRT, cuDNN, NCCL, Triton
Frameworks: NeMo, NIM microservices
Enterprise: NVIDIA AI Enterprise (recurring SaaS)
Libraries: TensorRT, cuDNN, NCCL, Triton
Frameworks: NeMo, NIM microservices
Enterprise: NVIDIA AI Enterprise (recurring SaaS)
Services & Ecosystem
DGX Cloud: GPU-as-a-service via CSPs
Omniverse: Digital twin simulation
DRIVE: Autonomous vehicle platform
Sovereign AI: National AI infrastructure partnerships
Omniverse: Digital twin simulation
DRIVE: Autonomous vehicle platform
Sovereign AI: National AI infrastructure partnerships
Competitive moat assessment
| Moat Factor | Strength | Detail |
|---|---|---|
| Switching Costs | Very High | CUDA rewrite cost is prohibitive. Years of developer training, optimized codebases, and toolchain dependencies. |
| Network Effects | High | More developers on CUDA → more libraries → more developers. 4,400+ accelerated apps create self-reinforcing ecosystem. |
| Pricing Power | Absolute | $30K-$40K per GPU at 75% gross margins. Customers pay willingly for performance/watt leadership and lowest TCO. |
| Scale Advantage | Massive | $15.4B R&D spend in CY2025 -- more than AMD's entire revenue. Outspends all competitors combined. |
| Can Customer Replace? | No (12-month) | AMD ROCm improving but years behind in ecosystem maturity. Custom ASICs (Google TPU, Amazon Trainium) serve narrow use cases only. |
Risks to the business model
Customer Concentration
Top 4 hyperscalers (MSFT, GOOG, AMZN, META) likely represent 50%+ of Data Center revenue.
Each is developing custom silicon alternatives (TPU, Trainium, MTIA).
Export Controls
U.S. government restrictions on China exports cost NVIDIA ~$4.5B in H20 write-downs
(Q1 CY2025). Further controls could restrict additional markets.
TSMC Single-Source
100% reliance on TSMC for leading-edge manufacturing. Any disruption (geopolitical,
capacity, yield) directly impacts NVIDIA's ability to deliver.
Data sourced from Daloopa, NVIDIA 10-K/10-Q filings, and earnings transcripts.