NVIDIA

The product became an environment in which other organizations think

NVIDIA was incorporated in 1993 and introduced CUDA in 2006. By fiscal 2026, the company described its offering not as a chip alone but as a programmable architecture spanning GPUs, networking, systems, hundreds of libraries and software kits, models, training data, and services. It reported more than 7.5 million developers using CUDA or its other software tools.¹

That stack coordinates work across product generations. Code, libraries, documentation, trained operators, server designs, and cloud configurations can accumulate instead of being rebuilt around each processor. NVIDIA itself says a larger developer population and installed base increase the platform's value to customers.¹

Founder continuity does not amount to majority voting control. NVIDIA's 2026 proxy reported that Jen-Hsun Huang had served as president, chief executive, and director since 1993 and beneficially owned 3.58% of outstanding shares as of the proxy's record date. The board classified each of the other nine director nominees as independent under Nasdaq rules.² Formal checks therefore coexist with authority attached to Huang's offices, tenure, and disclosed ownership stake.

The coordination has a second face. A technical choice can become an institutional commitment when software, skills, workflows, and data-center architecture depend on the same vendor's interfaces. Capability and dependency then travel together. The governing question is not whether the platform is useful, but how much practical authority a customer retains after building on it.

CUDA joined generations of hardware to accumulated developer knowledge

CUDA gave developers a supported programming environment that NVIDIA could carry across successive GPUs. In its fiscal 2026 filing, the company says CUDA runs on all NVIDIA GPUs and sits beneath domain-specific libraries, frameworks, software kits, and APIs. That is participant evidence for the intended architecture and current product boundary, not an independent adoption study.¹

The 2012 AlexNet paper records one consequential use. Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton trained a large convolutional neural network for five to six days on two NVIDIA GTX 580 GPUs, splitting it because a single card had only 3 GB of memory. Their model won the ILSVRC-2012 competition with a 15.3% top-five test error rate, compared with 26.2% for the second-best entry; the authors also made their CUDA-based implementation public.³

AlexNet demonstrates a bounded capability, not that NVIDIA alone caused the subsequent expansion of deep learning. The result also depended on the ImageNet dataset, model design, training methods, researchers, and a competition that made performance comparable. What NVIDIA supplied was a reusable hardware and software path that outside researchers could exploit and that the company could later optimize around.

Independent regulatory evidence supports the dependency side of the loop. In a 2024 inquiry into generative AI, France's Autorité de la concurrence described CUDA as proprietary, exclusive to NVIDIA chips, and the sector's most widely used environment. Stakeholders told the authority that hardware, software, ecosystem maturity, and community size create inertia; AMD's ROCm appeared marginal in that inquiry.⁴ The finding is specific to the authority's 2024 generative-AI record. It should not be generalized to every accelerated workload, geography, or future alternative.

Full-stack co-design moved coordination above the chip

NVIDIA calls its Blackwell data-center offering “extreme co-design”: chips, networking, systems, software, and algorithms are architected together. The filing also places GPUs, CPUs, interconnects, networking, rack-scale systems, software, libraries, models, datasets, and services inside the platform.⁵ This is a company account of design intent; it does not independently establish the performance or necessity of every layer.

Co-design moves the unit of adaptation above the chip. NVIDIA can respond to a bottleneck by changing several layers together. Customers may also become dependent on several layers at once: a programming model, optimized libraries, networking, a cloud configuration, and staff practice can function as one institutional commitment even when purchased as separate products.

The physical system is distributed. NVIDIA uses a fabless strategy and assigns wafer fabrication, memory, assembly, testing, and packaging to suppliers. Its fiscal 2026 filing identifies TSMC and Samsung as foundries; SK Hynix, Micron, and Samsung as memory suppliers; and Hon Hai, Wistron, and Fabrinet among assembly, testing, and packaging partners. The company said the supply chain was mainly concentrated in Asia and reported $95.2 billion of outstanding inventory purchase and long-term supply and capacity obligations at year end.⁶ Micron separately announced in February 2024 that it had begun volume production of HBM3E memory intended for NVIDIA's H200.⁷ That supplier-controlled record corroborates one production relationship, but does not disclose realized volume, margin, bargaining power, or working conditions.

Cloud access adds another decision point. The French inquiry found that building on-site computing capacity was costly and difficult amid GPU shortages, making cloud providers an important route to computing power.⁴ NVIDIA can coordinate the architecture without owning the factories or every route by which users reach it. Its power therefore coexists with dependence on suppliers, cloud providers, energy systems, and customer capital.

Public authority can redraw the market

In October 2022, the U.S. Bureau of Industry and Security amended the Export Administration Regulations to control specified advanced-computing chips, supercomputer end uses, semiconductor-manufacturing items, and related transactions involving the People's Republic of China. The rule stated national-security and foreign-policy purposes, including restricting military modernization and capabilities used for human-rights abuses.⁸ The legal record establishes the rule and its stated rationale, not whether it achieved those purposes.

The boundary continued to move. Effective January 15, 2026, BIS changed its licensing policy so applications to export NVIDIA H200, AMD MI325X, and similar processors from the United States to end users in China or Macau could receive case-by-case review if conditions concerning domestic supply, foundry capacity, recipient controls, shipment volume, and third-party testing were met.⁸ Access was neither a simple sale nor a permanent ban: public authority could change which product, customer, inspection process, and geography were permissible.

NVIDIA records the commercial side. Its fiscal 2026 filing says an April 2025 H20 licensing requirement led to a $4.5 billion charge for excess inventory and purchase obligations; later licenses produced about $60 million in H20 revenue. The company said that, at fiscal year end, it was effectively foreclosed from China's data-center compute market.⁹ That is reliable evidence of the company's reported financial treatment and market assessment, not the experience of Chinese customers, researchers, workers, or competitors.

The episode makes platform authority plural. NVIDIA decides what it will design and support; suppliers determine what capacity can be delivered; cloud providers mediate rented access; and states can condition a geographic path. A user may experience their combined decisions as price, waiting time, an unavailable instance, or a denied shipment without receiving a decision from any single institution.

The gains are also distributed. NVIDIA reported fiscal 2026 revenue of $215.938 billion and net income of $120.067 billion, while returning $40.4 billion through share repurchases and $974 million through dividends.¹⁰ The same filing reported about 42,000 direct employees in 38 countries, including 31,000 in research and development, and a 3.7% turnover rate.¹¹ Its proxy calculated median-employee total compensation of $282,050 and a 129-to-1 chief-executive pay ratio, while explaining that it reused the fiscal 2024 median employee and that cross-company ratios are not directly comparable.²

Independent reporting supplies a narrower, less favorable view. Current and former employees interviewed by Bloomberg described grueling, high-stress work that left little time to use the wealth created by equity compensation.¹² The public report does not disclose a sampling frame or establish prevalence. Taken together, high reported compensation and low turnover cannot stand in for evidence about workload, voice, contingent work, or conditions in supplier factories.

The material account is system-wide and cannot be reduced to a processor's efficiency. Berkeley Lab estimated that U.S. data centers used 176 TWh of electricity in 2023, or 4.4% of national consumption, and projected a range of 325–580 TWh in 2028. It estimated 66 billion liters of direct site water consumption and nearly 800 billion liters of indirect water consumption through electricity in 2023.¹³ The model ties recent growth partly to accelerated AI servers, but it does not allocate the totals to NVIDIA or distinguish the net effect of more efficient computation from growth in demand.

NVIDIA's fiscal 2026 sustainability report supplies a different boundary. It reports 10.701 million metric tons of Scope 3 greenhouse-gas emissions, including 9.302 million from purchased goods and services, compared with 10,390 metric tons for market-based Scope 1 and 2 combined. It also reports 451,297 cubic meters of water withdrawal while warning that a method change makes the fiscal 2026 water figures incomparable with prior years.¹⁴ These are corporate inventory figures, not a complete life-cycle account of fabrication, customer data-center use, displacement, or rebound.

Technical capability, financial value, constrained access, and material demand are outcomes of the same coalition. They should be evaluated together without pretending that the available sources can assign every benefit or burden to one company.

Portability is tested when authority moves

The French competition authority said NVIDIA appeared dominant in the IT components needed to train foundation models and recorded concerns about supply restrictions, contract terms, discrimination, CUDA dependence, and links to specialist cloud providers.⁴ The authority was identifying potential risks, not finding that NVIDIA had committed an infringement. Its opinion also acknowledged that the generative-AI sector was moving quickly and that relevant markets could not yet be defined precisely.

A European Commission merger review supplies useful counterevidence against treating every adjacent software layer as closed. In December 2024 the Commission found that NVIDIA likely held a dominant position in the global market for discrete data-center GPUs, yet cleared its acquisition of Run:ai unconditionally. Competitors confirmed that widely used compatibility tools would keep other GPU-orchestration software available, and the Commission found credible alternatives to Run:ai.¹⁵ That Phase I transaction review concerns orchestration software in the European Economic Area; it neither dissolves the French authority's CUDA concerns nor proves portability for an entire workload.

A list of competing processors therefore does not establish portability. A meaningful alternative must be reachable through software, skills, manufacturing capacity, cloud availability, performance, and cost appropriate to the work. A team retains more authority when it can inspect total costs, move a workload, preserve trained knowledge, appeal an allocation decision, and continue operating after a supplier, cloud provider, or state changes direction.

Capability and authority require different organizational lenses

Innovation, entrepreneurship, and renewal and strategy, competition, and adaptation illuminate the repeated recombination of processors, networking, systems, and software. Knowledge, expertise, and professional autonomy asks what developers can carry elsewhere; learning, quality, and reliability asks how libraries, tests, and field experience improve across generations; and work design, productivity, and automation asks which human tasks the stack makes easier, harder, or newly dependent.

Structure, hierarchy, and scale separates the central architecture from the distributed industrial network. Coordination, communication, and common understanding examines the shared interfaces, while delegation, decentralization, and responsibility follows decisions across NVIDIA, suppliers, clouds, customers, and states. Decision making, judgment, and bounded rationality, measurement, accounting, and control, and executive attention, information, and organizational sensing keep forecasts, benchmarks, road maps, purchase commitments, and exceptions from collapsing into a single performance number.

Purpose, mission, and institutional legitimacy tests broad claims about what accelerated computing is for. Authority, legitimacy, and acceptance asks why affected people should accept platform decisions; cooperation, incentives, and organizational equilibrium tracks how gains and dependencies are distributed; and governance, stewardship, and accountability asks who can inspect, contest, and revise those arrangements. Culture, informal organization, trust, and voice and organizational ignorance identify what financial, adoption, and turnover aggregates do not reveal about experience, remedy, or hidden cost.

The comparisons attached to NVIDIA are editorial, not historical genealogies. Netscape asks what changes when a platform's shared code is placed in an open institution. Andrew S. Grove offers a contrast between managing an integrated semiconductor manufacturer and coordinating a fabless network. The organizational-intelligence lens asks how accumulated developer knowledge becomes infrastructure, while benefit for all life keeps owners, users, workers, communities, and ecosystems inside the same performance account.

The unresolved boundary lies between a product and a shared substrate: which interfaces should be portable, what continuity can customers demand, and which publics should have standing when changes in chips, licenses, cloud terms, or supply commitments determine who can continue the work?

The unresolved evidence runs through the whole stack

The present record establishes NVIDIA's reported architecture, suppliers, governance, workforce, financial results, and export-control exposure; the AlexNet paper's bounded technical result; two regulators' transaction- and sector-specific 2024 competition assessments; a limited independent account of employee experience; official U.S. export rules; a modeled U.S. data-center resource account; and NVIDIA's corporate emissions inventory. It does not yet establish:

• representative migration costs for workloads moving from CUDA to competing environments across fields and organization sizes;
• first-person experience from independent developers, small laboratories, cloud customers, and people whose access changed under export controls;
• workload, voice, contingent-labor, and occupational-health evidence from NVIDIA employees and supplier-factory workers;
• facility-level effects on electricity prices, water, land, housing, employment, and public revenue in manufacturing and data-center communities; or
• an NVIDIA-attributable product life cycle that joins fabrication, transport, use, displacement, rebound, repair, reuse, and end of life.

Those gaps keep the community impact research-needed and limit any net claim about workers or ecosystems. They also connect the trained skill examined in knowledge, expertise, and professional autonomy to the scale problem in structure, hierarchy, and scale: how can a centrally designed stack coordinate a distributed industrial network without making contest or exit merely theoretical?

Paths into deeper study

• Trace one current workload through source code, libraries, compiler, GPU, networking, fabrication, cloud access, and electricity, then measure the cost and lost functionality of moving it to an alternative.
• Pair procurement and license records with interviews from universities, startups, cloud customers, and researchers in restricted markets.
• Add worker-controlled evidence from direct, contingent, construction, logistics, and supplier-factory labor rather than inferring experience from company aggregates.
• Build facility-level environmental and community cases before assigning NVIDIA a share of data-center or semiconductor-system effects.