NVIDIA

NVIDIA is no longer a chip company

NVIDIA is the defining infrastructure company of the artificial-intelligence build-out. Calling it a graphics-chip designer is now as incomplete as calling a cloud platform a server reseller. The company designs the processors, memory architecture, interconnect, networking, systems, software libraries and domain frameworks that turn electricity and data into useful computation. Its most important product is not an individual GPU. It is a programmable computing platform delivered at data-centre scale.

That distinction explains both the strength of the franchise and the seriousness of the debate. A faster accelerator can be copied or displaced. A platform that makes millions of developers productive, gives cloud providers a common service to sell, lets enterprises deploy the same code across clouds and generations, and improves every layer of the system together is much harder to unseat. But NVIDIA’s extraordinary scale has also turned its largest customers into motivated competitors. Amazon, Google and Microsoft are no longer building isolated chips; they are building accelerators, systems and software around workloads they understand better than any outside supplier.

The correct analytical question is therefore not whether a rival can produce a competitive chip. Several can. It is whether another architecture can deliver a lower usable cost per unit of intelligence after accounting for model development, utilisation, networking, power, deployment time, reliability and access to software. NVIDIA’s moat is the rate at which it improves that whole equation.

The business in one map

The portfolio looks sprawling only if each product is viewed separately. The same programmable architecture, developer tools and accelerated libraries can be reused across markets. A rendering technique developed for gaming can influence simulation; data-centre training can create models deployed on robots; networking developed for supercomputers becomes essential to AI factories. This reuse spreads engineering cost across a wider opportunity and gives NVIDIA more surfaces on which to establish its platform.

Accelerated computing: the migration beneath the AI cycle

Traditional CPUs are designed to execute a relatively small number of complex instruction streams with very low latency. They remain indispensable for operating systems, control logic and serial work. GPUs devote more of the silicon area to performing many similar operations in parallel. That architecture originally served computer graphics, where millions of pixels can be processed concurrently, but it also maps naturally to matrix operations used in machine learning, simulation and scientific computing.

The important qualification is that a GPU does not make arbitrary software faster. Code must expose parallel work and use algorithms and libraries suited to the architecture. Some workloads remain CPU-bound; others accelerate dramatically. This is why the software layer matters as much as transistor density. NVIDIA spent years turning theoretical parallel performance into something a developer can use through compilers, libraries, debuggers, frameworks and domain tools.

AI accelerated the migration because neural networks repeatedly perform dense linear algebra at vast scale. The useful result is not measured by peak operations on a specification sheet. It is measured by how quickly and cheaply a system can train a model or serve an answer while meeting accuracy, latency, reliability and power constraints. As models grew, the bottleneck moved from arithmetic to memory, then to communication between processors, then to power, cooling and orchestration. NVIDIA repeatedly expanded its design boundary to include the next constraint.

From Kepler to Rubin: a compounding architecture

NVIDIA’s architectural history is best read as a sequence of bottlenecks removed, not a catalogue of model numbers. Kepler made parallel compute more efficient and programmable. Pascal paired faster compute with high-bandwidth memory and the first NVLink. Volta introduced Tensor Cores specifically for deep-learning matrix operations. Ampere improved mixed precision and sparsity while allowing one GPU to be partitioned across workloads. Hopper added a Transformer Engine, lower-precision formats and faster interconnect for the model architecture that had become dominant.

Blackwell is the decisive conceptual shift. The company increasingly sells a rack whose accelerators, CPUs, switches, networking, memory, power and cooling have been engineered together. Rubin extends that logic across seven chips and multiple networking domains. This raises NVIDIA’s share of each deployment and makes comparisons based on GPU price less useful. It also raises execution risk: a rack-scale launch depends on many more components, manufacturing steps and thermal systems arriving together.

The product is the system

An isolated accelerator is valuable only if data reaches it, neighbouring accelerators communicate quickly, jobs remain fed with tokens and failures do not strand the cluster. AI infrastructure therefore has two different networking problems. Scale-up connects GPUs inside a tightly coupled computing domain so that a model can behave as if it is running on one very large machine. NVLink and NVLink switches address this problem. Scale-out connects many servers and racks across the data centre. InfiniBand and Spectrum-X Ethernet address that layer, with ConnectX adapters and BlueField processors managing data movement and infrastructure tasks.

The Mellanox acquisition was strategically transformative because it gave NVIDIA networking expertise, products and customer relationships before distributed AI became the industry’s defining workload. Networking is not an optional attachment to accelerator sales: communication time can determine cluster utilisation, and a small percentage improvement applied to an extremely expensive installation has large economic value. Control of networking also gives NVIDIA visibility into the system bottleneck and an additional surface for co-design.

Customers can still mix components. Ethernet is a broad ecosystem, hyperscalers design their own networks and open interconnect standards are improving. NVIDIA’s defence is to offer an integrated path whose performance, deployment time and support are easier to underwrite. NVLink Fusion is also strategically revealing: NVIDIA is allowing custom CPUs and accelerators to connect into parts of its rack architecture. That concedes that heterogeneous systems will exist while attempting to keep NVIDIA’s interconnect and system design at their centre.

CUDA is the moat—but “lock-in” is too simple

CUDA is the programming model and software environment that lets developers use NVIDIA’s parallel architecture. Its power comes from far more than a language interface. The stack includes optimised libraries for linear algebra, attention, data processing, genomics, simulation and many other domains; compilers and profilers; integration with leading AI frameworks; pretrained models and deployment tools; and a large body of code, documentation and human expertise accumulated over two decades.

The moat is partly switching cost, but productivity is the more important idea. A team chooses NVIDIA because a workload works today, is debuggable, has reference implementations, can be deployed through major clouds and is likely to run on the next architecture. The cost of changing platform includes rewriting and validating software, retraining engineers, reproducing performance and maintaining a second production path. Backward compatibility turns old development effort into an asset and makes each new library more valuable because it reaches a large installed base.

Scale strengthens the loop. More deployed hardware attracts software optimisation; better software raises real-world hardware performance; stronger application coverage drives more purchases; higher revenue funds the next architecture. NVIDIA also uses open models and frameworks to seed emerging markets such as robotics. “Free” software is often customer-acquisition and market-creation expenditure for the hardware platform.

But CUDA is not invulnerable. High-level frameworks can hide hardware differences. Compilers are becoming better at retargeting code. Large customers can afford dedicated teams to optimise for their own silicon. Open software ecosystems lower the cost of supporting AMD or custom accelerators, and inference engines increasingly schedule across heterogeneous hardware. The moat will narrow if developers experience hardware choice without a meaningful loss of performance, reliability or time to market. NVIDIA must therefore keep earning preference rather than relying on incompatibility.

How NVIDIA makes money

NVIDIA is fabless. It designs products and software but relies on manufacturing partners for wafers, advanced packaging, memory, assembly and testing. That model concentrates capital on architecture and ecosystem while allowing external suppliers to carry much of the fabrication base. The trade-off is dependence on a small number of highly specialised partners, particularly for leading-edge foundry capacity, high-bandwidth memory and advanced packaging.

The company monetises at several levels: chips and boards sold through module and system makers; integrated servers and racks; networking components; cloud-based access; and subscriptions such as NVIDIA AI Enterprise and virtual-GPU software. Hardware still provides the overwhelming economics. Software’s present value is mainly that it creates demand, raises utilisation, supports premium pricing and makes the installed base durable. Paid software can become meaningful, but the company does not yet provide enough standalone disclosure to treat it as a separate software franchise.

Channel accounting requires care. A direct customer may be an original-design manufacturer, system builder or cloud provider that integrates products for several end users. Reported customer concentration therefore shows bargaining and collection exposure but does not map neatly to final demand. Conversely, end-user concentration can be economically greater than the direct-customer table suggests because a handful of hyperscalers and model developers determine large deployment decisions.

These figures are dated 26 July 2026 and are included only to frame the business model. The durable issue is the combination of data-centre dominance, premium economics and unusually large forward supply commitments.

AI demand is several markets, not one

The phrase “AI compute” hides workloads with different economics. Pre-training learns general capabilities from vast datasets and rewards maximum cluster scale. Post-training adapts models through fine-tuning, reinforcement and synthetic data. Inference serves models to users and applications. Within inference, a simple classification request bears little resemblance to a reasoning agent that creates long contexts, calls tools, generates intermediate tokens and runs continuously. The hardware winner can differ by workload.

The common bear argument says that training belongs to GPUs but inference belongs to custom silicon. It is directionally sensible and analytically incomplete. Inference favours specialisation when a model and traffic pattern are stable, predictable and large enough to repay design cost. Yet frontier inference is becoming more computationally demanding as models reason, use tools and generate more tokens before producing an answer. Model architectures, context lengths and precision formats are also changing quickly. Programmability retains value when the target keeps moving.

NVIDIA does not need every inference operation to run on its hardware. It needs the pool of fast-changing and demanding workloads to grow faster than efficiency and substitution reduce compute per task. The key metric is not tokens alone but the cost, latency and complexity of producing economically useful outcomes.

Why hyperscalers both buy and compete

Hyperscalers are NVIDIA’s largest route to market and the strongest threat to its economics. They buy NVIDIA because customers demand it, software runs on it, frontier researchers move quickly on it and a common platform can be offered across regions and customer types. NVIDIA also acts as a neutral standard: a model developed on its architecture can often move among clouds, private systems and specialist AI providers with less adaptation.

The same customers build their own accelerators because infrastructure cost is strategically important, internal workloads are enormous and vertical control can improve efficiency. Google’s TPU, Amazon’s Trainium and Inferentia, and Microsoft’s Maia are increasingly accompanied by compilers, frameworks, rack designs and cloud services. This is no longer a contest between a complete NVIDIA platform and a bare internal chip.

The likely outcome is heterogeneous rather than winner-takes-all. NVIDIA remains favoured for frontier work, rapidly changing models, third-party demand and customers seeking portability. Custom silicon is strongest for stable, repetitive workloads inside a cloud whose owner controls the model, software and scheduling. The hyperscaler can use NVIDIA to attract demand while routing suitable internal or price-sensitive work to its own silicon. Over time this may reduce NVIDIA’s share without reducing the absolute market.

NVIDIA’s response is to widen the platform faster than customers can internalise it: annual architectures, rack-scale systems, networking, software and industry frameworks. NVLink Fusion is a more pragmatic response. By supporting third-party processors within NVIDIA-oriented systems, the company can retain interconnect and platform influence even when it does not supply every compute die. The risk is that openness makes the substitute easier; the opportunity is that NVIDIA becomes the architecture around heterogeneous AI factories.

The demand base beyond hyperscalers

Cloud concentration can obscure a broader market. Specialist AI clouds purchase systems to sell access to developers and model companies. Enterprises deploy private AI where latency, data control or predictable economics justify ownership. Governments fund sovereign infrastructure to keep strategic data, languages and capability within national control. Research institutions and regulated industries need supported systems that work in isolated environments. Industrial and edge customers eventually require models to run near machines.

NVIDIA is particularly well positioned where the customer lacks the engineering organisation to design a processor, networking fabric and software stack. A complete reference architecture compresses deployment time and makes performance more predictable. OEM, cloud and consulting partners extend distribution. The platform can therefore diffuse from a small number of frontier builders into a much larger population of users.

This demand is not equally mature. Hyperscaler capacity is deployed and monetised at scale; enterprise projects can remain experimental; sovereign purchases may depend on policy and funding; specialist clouds introduce credit and utilisation risk. Investors should distinguish installed, financed capacity with contracted users from announced ambitions. NVIDIA’s order book can be real while the end customer’s return remains uncertain.

Networking is a second franchise, not an accessory

In conventional servers, networking transports completed work. In distributed AI, the network participates in the computation. GPUs must repeatedly exchange model states and intermediate results. If communication stalls, valuable processors wait. This makes bandwidth, latency, congestion control, failure recovery and topology part of application performance.

NVIDIA occupies both major layers. NVLink creates a high-bandwidth scale-up domain within systems and racks. InfiniBand offers a tightly controlled scale-out fabric, while Spectrum-X adapts Ethernet for AI traffic and addresses customers committed to open networking. ConnectX adapters connect servers; BlueField offloads infrastructure, isolation and data movement. The portfolio lets NVIDIA tune the system from the accelerator through the fabric.

The strategic benefit is larger than incremental component revenue. Networking increases attach per AI factory, protects accelerator utilisation and creates an architectural barrier to chip-level substitution. The counterargument is equally important: cloud providers have deep networking expertise, Ethernet’s ecosystem is broad, and customers resist allowing a single supplier to control every layer. NVIDIA must prove that integration creates greater savings than multi-vendor choice.

Software economics: enablement before monetisation

NVIDIA’s software estate serves three economic functions. First, foundational software such as CUDA and CUDA-X makes hardware usable. Second, domain frameworks such as NeMo, Isaac and Omniverse shorten the path from infrastructure to an application. Third, commercial products such as NVIDIA AI Enterprise, NIM microservices and virtual-GPU software can generate recurring revenue and support.

The first function is already enormously valuable because it supports hardware share and pricing. The second expands the market by reducing specialist work. The third is the least proven in disclosed financial terms. NVIDIA may ultimately charge for enterprise support, deployment, optimisation and model services, but customers also expect a large software layer to be bundled with expensive systems. Open-source alternatives and cloud-managed services constrain standalone pricing.

Dynamo illustrates the strategy. Efficient inference requires requests to be routed, prefills and decoding to be scheduled, caches to be managed and GPUs to remain utilised. Better serving software can create more customer capacity from the same hardware, apparently reducing unit demand. In practice it can also lower the cost of AI enough to expand usage. NVIDIA benefits if software efficiency makes its platform the preferred place to run more inference, even when the software itself is inexpensive.

The sensible underwriting approach is conservative: value software first as moat and market creation, then recognise a separate recurring-revenue franchise only when revenue, retention and margins are visible. Describing all enabled hardware revenue as software economics double-counts the same advantage.

Physical AI: one architecture across train, simulate and run

Physical AI extends learned intelligence into robots, vehicles, factories and machines. The development problem is harder than generating text because a system must perceive a changing world, reason about physical consequences and act safely under latency and power constraints. Real-world data is expensive, dangerous edge cases are rare, and mistakes have physical costs.

NVIDIA’s proposition spans three computers. Data-centre systems train foundation and policy models. Omniverse and Cosmos generate, simulate and evaluate worlds so developers can expose systems to many conditions before deployment. DRIVE and Jetson computers run models in vehicles and robots with low latency. Isaac supplies robotics frameworks, while Cosmos world models help reason over and generate physical scenarios.

This is strategically attractive because each deployed machine can pull demand through the entire stack: training compute, simulation, software tools and edge processors. It also reuses assets built for data-centre AI and graphics. Open models and reference frameworks encourage adoption before the market is large enough to monetise directly.

The opportunity should not be confused with the current earnings base. Robotics markets are fragmented; qualification cycles are long; safety and integration dominate the bill of materials; and industrial customers often require customised systems. Automobile programmes can generate substantial design wins but take years to reach volume. Physical AI is credible strategic optionality because NVIDIA addresses the full development loop, not because every robot announcement is future revenue.

Gaming still matters

Gaming is now small beside data-centre AI, but it remains strategically important. GeForce established the installed base, developer relationships and rapid product cadence that made NVIDIA a platform company. RTX combined programmable shading, ray tracing and neural processing; DLSS uses AI to generate frames and improve image quality. The boundary between graphics and AI is disappearing as rendering becomes a hybrid of simulation and neural generation.

The gaming franchise also places capable local AI processors in millions of PCs. That can support creation, personal models, agents and applications that need low latency or privacy. It gives developers a consumer distribution surface that data-centre-only competitors lack. Yet gaming demand is cyclical, channel inventory can distort shipments and high-end pricing invites competition. It should be valued as a durable platform and innovation engine, not as a miniature version of data centre.

Supply chain, power and geopolitics

NVIDIA’s asset-light model does not make supply risk disappear; it converts fixed manufacturing cost into dependency and commitments. Leading GPUs require advanced wafers, high-bandwidth memory, complex packaging, substrates, networking components and liquid-cooling systems. A shortage at any layer can constrain the complete rack. Annual platform transitions amplify coordination risk because suppliers and customers must qualify many components on a compressed schedule.

The company’s enormous inventory and capacity commitments show management securing supply far ahead of shipment. In a rising market this enables growth and reassures customers. If demand changes, product transitions accelerate or a customer cancels, the same commitments can create excess inventory, prepayments and purchase obligations. The balance sheet can absorb more risk than smaller suppliers, but the absolute industrial exposure is now unprecedented for a fabless designer.

Power is becoming a demand constraint and a product feature. AI factories require generation, grid connections, cooling, land and construction. Better performance per watt raises the amount of useful output available from a fixed power envelope, which supports rapid system replacement. But if data-centre power cannot be brought online, customer budgets do not convert into installed systems. NVIDIA can improve the numerator of the equation; it cannot manufacture permits or transmission capacity.

Export controls have effectively removed NVIDIA from much of the advanced data-centre compute market in China. The immediate effect is lost sales; the longer-term effect is more important. Restrictions encourage local hardware, software and model ecosystems, reduce NVIDIA’s developer reach and may create competitors that later expand elsewhere. Designing reduced-capability products consumes resources and may still fail to obtain approval. China is therefore not merely an addressable-market haircut but a potential fracture in the global platform.

Competitive landscape

Traditional semiconductor analysis focuses on benchmark leadership, manufacturing node and price. Those remain necessary but insufficient. Each competitor must be judged on usable software, system availability, networking, developer support, cloud access and speed of deploying a real workload. NVIDIA can lose a benchmark and retain the platform; it can also win a benchmark and lose an account whose owner values vertical economics more highly.

The investment debates that matter

What could break the flywheel

How to judge NVIDIA from here

Quarterly revenue is a lagging summary of orders placed within a constrained system. The leading evidence sits beneath it. Watch whether new architectures reach production on time and whether customers deploy complete systems rather than merely receive components. Follow networking attach and system utilisation because they show whether NVIDIA is increasing its control of the AI factory. Examine supply commitments against actual deployment, power availability and end-user consumption.

For the moat, compare the time and engineering required to put the same production workload on NVIDIA, AMD and custom silicon. For inference, distinguish simple high-volume jobs from changing agentic workloads. For hyperscalers, separate continued NVIDIA purchasing from the share of internal workloads moving to their own chips. For software, demand recurring-revenue disclosure before capitalising it independently. For physical AI, privilege repeat production over partnerships and demonstrations.

The most important variable is the pace at which useful AI demand grows relative to three offsets: algorithmic efficiency, better hardware and substitution toward custom silicon. Efficiency is not automatically bearish. Lower cost can release far more usage, as it did across computing history. It becomes bearish only when demand elasticity is insufficient or when another platform captures the released volume.

Jensen Huang in his own words

Huang’s short formulations explain how management sees the market it is trying to create. They are useful strategic framing, not independent evidence; the operating tests in the preceding sections remain the basis for judging whether the vision becomes durable economics.

Bottom line

NVIDIA’s achievement is not that it designed the fastest AI chip. It transformed a parallel processor into an industry platform, then expanded that platform as the bottleneck moved from the chip to memory, networking, the rack, the data centre and software. CUDA created developer preference; rapid architecture cycles protected performance leadership; Mellanox made networking part of the product; and integrated systems turned NVIDIA from a component vendor into the architect of the AI factory.

The moat is strongest where workloads are difficult, changing and valuable enough that time to solution matters more than component price. It is weakest where a very large customer controls a stable workload and can optimise the entire stack around its own silicon. Both conditions can expand at the same time. NVIDIA can remain the standard for frontier and general-purpose accelerated computing while hyperscalers capture more repetitive inference.