Observed action and outcome
NVIDIA continued CUDA beyond the 2007 public beta, adding programming features, reusable libraries, developer distribution, and integrated accelerated-computing systems. CUDA 6 addressed memory and library friction, cuDNN supplied reusable neural-network primitives under a permissive license, and DGX-1 combined accelerators, software, development tools, storage, and networking. The record shows the products were announced; it does not show whether management used the exact internal gates proposed in Part A. claim.nvidia.cutoff.cuda-public-toolkit claim.nvidia.outcome.software-system-expansion
Reported consolidated results were uneven before becoming much larger. Revenue was USD 3.326 billion in FY2010, USD 4.130 billion in FY2014, USD 9.714 billion in FY2018, USD 16.675 billion in FY2021, and USD 60.922 billion in FY2024. Income from operations was negative USD 98.945 million in FY2010 and USD 32.972 billion in FY2024. These are company-wide endpoints, not CUDA revenue or a causal return. claim.nvidia.outcome.reported-consolidated-scale claim.nvidia.outcome.attribution-bounded table.nvidia.outcome.consolidated-scale
Reported consolidated R&D expense was USD 908.851 million in FY2010, USD 1.336 billion in FY2014, USD 3.924 billion in FY2021, and USD 8.675 billion in FY2024. The filings do not allocate those amounts to CUDA, so this series measures the company investment base rather than CUDA investment. claim.nvidia.outcome.rd-scale-not-cuda-investment table.nvidia.outcome.rd-scale
What the technical record establishes
General-purpose programming on GPUs predated CUDA. Brook for GPUs and competing programmable architectures establish prior work, making this a commercialization, distribution, and platform-layering case—not a claim that NVIDIA invented GPGPU. claim.nvidia.cutoff.gpgpu-prior-art claim.nvidia.cutoff.competitive-boundary evidence.nvidia.cutoff.gpgpu-prior-art
The 2012 AlexNet paper reported a winning 15.3-percent top-five test error versus 26.2 percent for the second-best entry and said training took five to six days on two GTX 580 GPUs. This independently establishes the reported research result and GPU configuration; it does not prove that CUDA alone caused the result. claim.nvidia.outcome.alexnet-gpu-milestone evidence.nvidia.outcome.alexnet-result evidence.nvidia.outcome.alexnet-gpu-training
CUDA 6, cuDNN, and DGX-1 illustrate a layered mechanism: expose hardware through a programming surface, reduce recurring developer work with reusable libraries, and package hardware plus software into deployable systems. Their issuer performance claims and peak specifications are not treated as independently verified customer economics. claim.nvidia.outcome.software-system-expansion evidence.nvidia.outcome.cuda6-product evidence.nvidia.outcome.cudnn-library evidence.nvidia.outcome.dgx-launch
Adoption signals are company-defined, not proof of lock-in
NVIDIA reported more than two million CUDA downloads in 2013, more than 2.2 million developers using CUDA and other software tools in FY2021, and support for more than 3,500 applications in FY2024. Downloads, developers, and applications have different definitions and are not a comparable cohort. They support a distribution thesis but do not independently measure active retention, workload portability, switching cost, or profit. claim.nvidia.outcome.ecosystem-adoption-claims assumption.nvidia.outcome.ecosystem-not-lockin
The FY2010 filing also described AMBER acceleration and broader Tesla OEM availability. Those are issuer claims about use and distribution; the same filing reported an operating loss, underscoring why product evidence and economic evidence must remain separate. claim.nvidia.outcome.early-adoption-claim table.nvidia.outcome.consolidated-scale
Data Center is an outcome signal, not a CUDA revenue series
NVIDIA reported Data Center revenue of USD 1.93 billion in FY2018 and USD 47.5 billion in FY2024. The FY2018 disclosure included Tesla, GRID, and DGX; the FY2024 disclosure described processors, systems, networking, software, services, DGX Cloud, and generative-AI workloads. The anchors remain separately defined, and no CUDA CAGR or CUDA-attributable revenue is calculated. claim.nvidia.outcome.data-center-signal-not-cuda-series table.nvidia.outcome.data-center-anchors assumption.nvidia.outcome.metric-boundaries
FY2014 provides an important non-monotonic checkpoint: consolidated revenue declined 4 percent while GPU business, gaming, Tesla, and Quadro moved differently. The outcome was never one clean CUDA series. evidence.nvidia.outcome.fy2014-mixed-results claim.nvidia.outcome.reported-consolidated-scale
Required causal controls
Hardware and supply mattered. NVIDIA identified TSMC capacity constraints that prevented fulfillment and hurt revenue and gross margin; later results cited Pascal, Volta, and Ampere ramps; and FY2024 supply still depended on foundries, memory suppliers, and CoWoS packaging. Reuters described HGX as a complex system whose missing components could delay shipment. Software ecosystem, hardware performance, and physical supply are therefore separate mechanisms. claim.nvidia.outcome.hardware-and-supply-co-causes evidence.nvidia.outcome.fy2014-tsmc-constraint evidence.nvidia.outcome.fy2024-supply
Gaming and cryptocurrency affected FY2018 growth, while NVIDIA later said cryptocurrency volatility and pandemic-era behavior changed demand across Gaming, Data Center, mobile workstations, and professional visualization. These external demand effects cannot be credited to the original platform decision. claim.nvidia.outcome.crypto-pandemic-co-causes evidence.nvidia.outcome.fy2021-crypto-pandemic
NVIDIA paid USD 7.13 billion for Mellanox and reported that Mellanox contributed 10 percent of FY2021 revenue. The acquisition added interconnect products and a broader computing, networking, and storage stack, so its contribution remains an acquisition and networking control. claim.nvidia.outcome.mellanox-networking-control table.nvidia.outcome.mellanox-consideration table.nvidia.outcome.mellanox-revenue-share
FY2024 results also coincided with a generative-AI investment wave and later full-stack execution. Reuters recorded management's structural-demand thesis and an analyst's warning that customers might be buying GPUs before establishing how to monetize them. This is evidence of both demand and uncertainty, not a forecast that demand must persist. claim.nvidia.outcome.generative-ai-demand-control evidence.nvidia.outcome.reuters-demand evidence.nvidia.outcome.reuters-skepticism-supply
Competition and regulation remained material. The FY2024 filing listed GPU, custom-chip, and cloud-provider competitors and reported that China declined from 19 percent to 14 percent of Data Center revenue amid export restrictions. claim.nvidia.outcome.competition-export-controls table.nvidia.outcome.china-data-center-share
Causal assessment and counterfactual
The primary bounded hypothesis is that the programming model, libraries, developer distribution, and integrated systems contributed to NVIDIA's later platform position. The sequence and product record support contribution, but not a causal percentage. hypothesis.nvidia.platform-layering-contributed claim.nvidia.outcome.attribution-bounded
The rival hypothesis assigns more explanatory weight to GPU and semiconductor advances, foundry and system supply, gaming and cryptocurrency, pandemic behavior, Mellanox and networking, generative-AI demand, competition, export controls, and later execution. The observational record cannot eliminate it. hypothesis.nvidia.hardware-demand-and-execution-dominated assumption.nvidia.outcome.partial-attribution
The feasible graphics-first counterfactual would have kept a bounded CUDA tool available while delaying broader investment until workload, contribution, portability, and cost gates cleared. No controlled comparator identifies its forgone adoption, savings, revenue, profit, cash flow, or valuation effect. counterfactual.nvidia.graphics-first assumption.nvidia.outcome.counterfactual-unquantified
Transferable learning and abstention
For a company exposing differentiated hardware through a developer platform while standalone economics remain unknown, release compilers, libraries, tools, and systems in measurable stages. Expand only after frozen active-developer, production-workload, contribution, investment, portability, reliability, and supply gates clear. The rule remains a candidate, not a universal law. rule.stage-deep-technology-platform-under-unknown-economics
The case contains no stable CUDA revenue, CUDA gross profit, CUDA investment, counterfactual cash-flow, capitalization, or market-price packet. It therefore produces no CUDA-only return, shareholder-return attribution, optimal investment budget, or target price. claim.nvidia.outcome.return-valuation-abstention claim.nvidia.cutoff.valuation-inputs-missing