CompanyGraph
NVIDIA Corporation
NVDA · United States
Designs GPUs whose parallel matrix computation architecture converts electrical power into AI and HPC workloads, then binds those workloads to that architecture through accumulated software depth.
CUDA's parallel thread model requires AI frameworks and inference engines to be written against its specific instruction set, so every production system built on TensorRT or CUDA-optimized libraries accumulates switching debt measured in engineering quarters — debt that replicates the architecture's hold across millions of developers at no incremental fabrication cost. That software depth, however, can only reach the market as physical silicon if TSMC allocates sufficient 4nm and 5nm wafer starts, a capacity shared with Apple and other high-volume customers, meaning chip volume is capped by a single foundry's allocation regardless of capital deployed. U.S. export controls on A100 and H100 sales to China then remove over 20% of the addressable data center market from that already constrained supply, compressing the pool of deployable chips further. The entire system therefore depends on developers continuously writing new code against CUDA rather than absorbing the one-time rewrite cost to port away, because if that porting decision is made at scale, the switching debt that justifies TSMC's indispensable role and the software ecosystem's hold dissolves together.
How does this company make money?
The company sells chips per unit to distributors and OEMs (manufacturers who build the chips into their own products): H100 data center GPUs are priced at $25,000–$40,000 per unit, GeForce gaming cards at $300–$1,600 per unit. Additional mechanics include licensing arrangements tied to CUDA development tools and AI framework optimizations.
What makes this company hard to replace?
CUDA-optimized code bases require months of rewriting to migrate to AMD ROCm or Intel OneAPI platforms. TensorRT — an inference engine (software that runs trained AI models efficiently) — is embedded in production AI systems in ways that make switching costs measurable in engineering quarters. GeForce Experience software and RTX ray-tracing pipelines lock game developers into RTX hardware certification cycles.
What limits this company?
TSMC's advanced-node wafer starts at 4nm and 5nm are finite and shared with Apple and other high-volume customers, so chip supply cannot exceed TSMC's allocation to this company regardless of capital deployed. Because all advanced GPU fabrication runs through those nodes at a single foundry, any allocation shortfall or geopolitical disruption to Taiwan directly caps the volume of CUDA-capable chips that can reach market.
What does this company depend on?
The mechanism depends on TSMC foundry capacity at 4nm and 5nm process nodes, Samsung and SK Hynix HBM3 high-bandwidth memory (the fast memory stacked directly on the chip that feeds the GPU's parallel cores), the CUDA software platform developed internally, substrate materials sourced through Taiwan Semiconductor Manufacturing, and chip packaging services from Advanced Semiconductor Engineering.
Who depends on this company?
Microsoft Azure and AWS data centers running GPT and Claude model training would face months-long delays scaling AI services if GPU supply were disrupted. Autodesk and Adobe creative software users would lose real-time 3D rendering and video editing acceleration. Tesla and Waymo autonomous vehicle development would slow due to lack of simulation compute for neural network training.
How does this company scale?
The CUDA software architecture replicates freely across unlimited GPU installations once developed, enabling software-hardware lock-in across millions of developers at no incremental cost. Physical chip production cannot scale beyond TSMC's advanced-node wafer capacity, creating hard supply limits during demand surges that no amount of capital can immediately overcome.
What external forces can significantly affect this company?
U.S. Commerce Department export controls restricting A100 and H100 sales to China eliminate over 20% of potential data center GPU addressable market. Taiwan geopolitical tensions threaten TSMC foundry access that produces all advanced GPU chips. European Union AI Act compliance requirements may restrict certain GPU compute capabilities for AI model training applications.
Where is this company structurally vulnerable?
If major cloud providers or AI companies standardize training and inference pipelines on a non-CUDA architecture — converting the rewrite cost from a switching barrier into a one-time sunk cost they absorb once — the accumulated library depth loses its lock-in force, because the ecosystem advantage exists only as long as developers are writing new code against CUDA rather than porting away from it.