Nvidia (NVDA)

1. Business Overview

Few companies in the history of capitalism have pivoted so decisively from one era to another as Nvidia. Founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem, Nvidia spent its first two decades as a semiconductor company best known for making graphics processing units for gamers. Today, it is the backbone of the global artificial intelligence infrastructure build-out and, briefly in late 2025, the world’s first company to reach a $5 trillion market capitalisation.

The company’s core insight — that the massively parallel processing architecture underlying a GPU is precisely the computational structure required to train and run neural networks — turned out to be one of the most consequential technological bets of the modern era. The GPU was designed to handle millions of simultaneous mathematical operations for rendering graphics; AI turns out to need exactly the same capability, at enormously larger scale.

Nvidia generates revenue across several product families, though the concentration is now striking. In fiscal year 2025, the company recorded $130.5 billion in total revenue, representing a 114% increase year-over-year. The vast majority of this — roughly 87 cents of every dollar — comes from the Data Center segment, which encompasses the H100, H200, and now Blackwell-generation GPUs sold to hyperscalers (Microsoft Azure, Google Cloud, AWS, Oracle), sovereign AI programmes, and enterprise customers building private AI infrastructure. The remaining revenue is split principally between Gaming (GeForce RTX cards for consumers) and smaller but growing segments in Automotive and Professional Visualisation.

What makes the business model exceptional is the stack. Nvidia does not just sell silicon; it sells a complete system. Its NVLink and NVSwitch interconnect technology enables thousands of GPUs to function as a unified compute cluster. Its software platform, CUDA, has accumulated over two decades of optimised libraries, frameworks, and developer tooling. Its DGX and HGX server systems are sold as complete solutions. The customer buying a Blackwell GPU cluster is, in practice, buying into a full computing ecosystem, not merely a chip.

2. Industry Context

The market in which Nvidia now primarily competes — AI accelerator chips for data centres — did not meaningfully exist five years ago. Today, it is the most important segment in the global semiconductor industry, and it is expanding at a pace that defies conventional forecasting. Every major technology company on the planet is simultaneously building or expanding AI infrastructure, and that infrastructure runs almost exclusively on accelerated compute.

The competitive landscape divides into three distinct tiers. Nvidia occupies the first tier alone, with an estimated 80% or more of the data-centre AI accelerator market by revenue. When cloud providers’ bespoke silicon is excluded, Nvidia’s share of the GPU-based data centre compute market may be as high as 98%.

AMD’s Instinct MI300X represents the strongest external challenge to Nvidia’s dominance, offering 192GB of HBM3 memory compared to 80GB in the H100, at competitive pricing. However, AMD faces the classic chicken-and-egg problem: developers will not adopt ROCm until software support improves, but software support will not improve without developer adoption. AMD has made genuine strides — its data centre segment reached $4.3 billion in Q3 2025, up 22% year-over-year — but that single quarterly figure compares against Nvidia’s $30.8 billion data centre revenue in the same period.

The third tier consists of custom silicon developed by the hyperscalers themselves — Google’s TPUs, Amazon’s Trainium and Inferentia, Microsoft’s Maia — and a handful of well-funded startups such as Groq and Cerebras. These represent a more structural long-term threat: customers who have historically been Nvidia’s best clients developing the means to reduce their dependence on it.

The industry’s defining characteristic at present is that demand substantially exceeds supply. Nvidia is supply-constrained, not demand-constrained, and has been for most of the past two years. This is both a statement of how extraordinary the AI investment cycle has been, and a reminder that the company’s current financial results are, at least in part, a function of a constrained supply environment that will eventually normalise.

3. Economic Moat

Nvidia’s competitive advantage is among the most multi-layered in the technology sector, and understanding each component separately is important because their durability differs.

CUDA and software lock-in is the deepest and most defensible part of the moat. CUDA — Compute Unified Device Architecture — is the programming platform that allows developers to write code that runs on Nvidia GPUs. Launched in 2007, it has been refined continuously for nearly two decades and has accumulated an enormous ecosystem: the company dominates with full-stack AI solutions and widespread adoption, with the competition now extending into semiconductor manufacturing, chiplet design, 3D stacking, and multi-die packaging — but CUDA’s dominance as a software framework remains the defining factor. The millions of engineers who have learned to write CUDA code, the billions of lines of production AI code that assume Nvidia hardware, the pre-optimised libraries for every major AI framework — all of this creates switching costs that are genuinely structural rather than merely commercial. Migrating a large AI training workload from Nvidia to AMD’s ROCm platform is not simply a matter of swapping hardware; it requires re-engineering, re-testing, and accepting performance uncertainty.

Architecture and product cadence constitutes a second layer. Nvidia has maintained a roughly two-year GPU architecture release cycle (Volta, Ampere, Hopper, Blackwell) that has consistently outpaced competitors in raw compute performance. Each generation has been sufficiently ahead of alternatives to justify the price premium, and the forward roadmap (Rubin architecture is already announced) ensures that customers who commit to Nvidia are committing to a roadmap they can plan around.

NVLink and system-level integration creates a third layer that is often underappreciated. At scale, training large AI models requires not just fast individual GPUs, but fast interconnects that allow clusters of GPUs to share memory and communicate with low latency. Nvidia’s NVLink and NVSwitch technology makes its multi-GPU clusters significantly more efficient than comparable AMD or Intel configurations. This systems-level advantage is harder to replicate than chip performance alone.

Intellectual property and manufacturing partnerships are the final structural layer. Nvidia has deep and exclusive partnerships with TSMC for cutting-edge fabrication and with HBM memory suppliers. Its portfolio of AI-related patents is extensive. These do not prevent competition, but they raise the cost and timeline for anyone attempting to replicate the full stack.

The honest assessment, however, is that the moat has an Achilles heel: it is fundamentally a software and ecosystem advantage built on top of a hardware product. If a competitor were to achieve genuinely superior hardware performance AND build sufficient software parity, the switching cost argument weakens. The hyperscalers’ internal chip programmes are the most credible long-term threat to this structure, precisely because they bypass the need to match Nvidia in the open market.

4. Financial Quality

Nvidia’s fiscal year 2025 produced $130.5 billion in revenue, $81.5 billion in operating income, and $72.9 billion in net income — a growth rate of 114%, 147%, and 145% respectively versus the prior year. These are not merely impressive numbers; they are historically unusual for a company of this size.

The gross margin profile deserves particular attention. GAAP gross margins for the full year were 75.0%, up from 72.7% in fiscal year 2024. For a hardware company, these margins are extraordinary — comparable to premium software businesses, and reflective of both pricing power and the extent to which Nvidia has moved up the value stack from commodity chip supplier to full-system provider. Crucially, margins were expanding even as revenue grew at triple digits, suggesting that the business was not buying growth at the expense of quality.

However, there is a notable recent development. Nvidia’s outlook for the first quarter of fiscal year 2026 guided for GAAP gross margins of approximately 70.6%, down from 73-75% in recent quarters. The Q1 FY26 results confirmed this, showing non-GAAP gross margins of 61.0%, down 17.9 percentage points year-over-year. This compression is primarily attributable to the ramp-up costs associated with the Blackwell architecture, H20 export restriction charges (discussed under Risks), and the always-difficult yield improvement curves in early production ramp. Whether this represents a temporary trough or the beginning of a more structural margin decline is arguably the most important near-term financial question for the company.

Cash generation is exceptional by any measure. Nvidia generates more free cash flow than most sovereign wealth funds deploy annually, and has been systematically returning it to shareholders: the quarterly dividend was increased by 150% in mid-2024, and buyback programmes have been sustained at substantial scale. The balance sheet is fortress-quality, with net cash well in excess of any near-term capital requirements.

Return on invested capital (ROIC) has reached levels rarely seen outside software companies, reflecting the asset-light nature of the fabless semiconductor model: Nvidia designs its chips but outsources fabrication to TSMC, keeping capital expenditure modest relative to revenues. This fabless structure amplifies financial returns in periods of high demand, though it also introduces supply-side risk during periods of constrained TSMC capacity.

5. Management and Capital Allocation

Jensen Huang is one of the most consequential technology executives of his generation. As a co-founder who has led the company continuously since its inception in 1993, he brings an unusual combination of engineering depth, strategic vision, and long-term conviction. The pivot toward AI compute — specifically, the decision to invest in CUDA as a general-purpose computing platform in 2006, years before AI became commercially relevant — is now widely regarded as one of the best strategic bets in technology history. It was made without certainty of reward, against considerable internal and external scepticism, and over a decade before it produced visible financial returns.

His compensation structure includes substantial equity components tied to long-term performance, aligning his incentives meaningfully with those of long-term shareholders. He retains a meaningful ownership stake in the company, though it has been reduced from his earlier holdings through decades of diversification.

Capital allocation has been notably shareholder-friendly. In periods of cash generation, Nvidia has prioritised buybacks and dividend increases over empire-building acquisitions. The one notable exception — the attempted $40 billion acquisition of ARM Holdings in 2020 — was ultimately blocked by regulators in the UK and EU. Whether that outcome was good or bad for shareholders remains debatable, but the episode suggests management has ambitions beyond organic growth.

One governance concern worth noting is the degree to which the company’s strategy and culture are concentrated in Huang personally. Succession is not an imminent risk, but it is a legitimate long-term consideration for any investor with a multi-decade time horizon.

6. Risks and Red Flags

Geopolitical and export control risk is perhaps the most immediate threat. The US government has progressively tightened restrictions on the sale of advanced AI chips to China, and the situation remains in flux. In April 2025, the US imposed restrictions on the export of Nvidia’s H20 chips to China — until then the most advanced AI chip legally exportable to the country — with the restriction placed indefinitely. As a result of these export restrictions, Nvidia expected to incur a charge of approximately $5.5 billion in Q1 FY26. China had historically represented a significant revenue opportunity, and its progressive closure is a material headwind. The risk cuts both ways: restrictions hurt Nvidia’s revenue directly, but they also, ironically, accelerate Chinese domestic chip development, which could eventually threaten Nvidia’s global market position if Chinese alternatives reach competitive performance levels.

The Jevons Paradox and model efficiency risk received dramatic illustration in January 2025. When DeepSeek released its R1 reasoning model, Nvidia’s stock fell approximately 17%, erasing close to $600 billion in market capitalisation in a single day. The fear was straightforward: if AI models could be trained at a fraction of the previously assumed cost, demand for Nvidia’s chips might be lower than investors had priced in. The counterargument — known as the Jevons Paradox, the 19th-century observation that efficiency improvements in a resource tend to increase rather than decrease overall consumption — suggests that cheaper AI inference will expand demand so dramatically that total chip requirements increase. Both arguments have merit, and the tension between them will define Nvidia’s revenue trajectory over the next several years.

Customer concentration and hyperscaler bargaining power is a structural concern. A substantial portion of Nvidia’s revenue flows from a handful of hyperscalers — Microsoft, Google, Amazon, and Meta — each of whom is simultaneously Nvidia’s best customer and its most credible potential competitor, via internal chip development. As their own custom silicon matures, they have increasing leverage in pricing negotiations and increasing incentive to reduce Nvidia dependence.

Gross margin compression from the Blackwell ramp and changing product mix is near-term financial risk. The shift from mid-70% to low-60% gross margins, even if partially transitional, changes the unit economics of the business and raises questions about the medium-term financial profile.

Valuation risk is arguably the central risk for any new investor. At current prices, Nvidia is priced to perfection in an industry characterised by rapid technological change, geopolitical uncertainty, and potential demand cyclicality. Any shortfall in the AI infrastructure build-out — whether from macro slowdown, efficiency breakthroughs, or simply the lumpy nature of large capital programmes — could produce severe multiple compression even if the underlying business remains strong.

7. DAFO (SWOT) Analysis

Strengths are centred on the combination of a two-decade software moat, a sustained hardware leadership cycle, and extraordinary financial productivity. The CUDA ecosystem and NVLink system integration represent genuinely difficult-to-replicate advantages. The financial quality — 75% gross margins on $130 billion in revenue — is a testament to pricing power that is the hallmark of a business with real monopoly characteristics in its core market.

Weaknesses are real but largely manageable in the near term. The dangerous concentration — both in product (Data Center is effectively the whole company) and customer (the top handful of hyperscalers) — means that any deterioration in the AI infrastructure investment cycle would flow through Nvidia’s results with unusual directness. The TSMC dependency, while common to all fabless designers, has particular severity given the geopolitical risk around Taiwan.

Opportunities are compelling and genuinely long-duration. The next phases of AI — physical AI, agentic systems, sovereign national programmes, enterprise adoption — are all compute-intensive and all likely to require accelerated silicon. Nvidia’s automotive platform (DRIVE) and its Isaac robotics infrastructure represent early-stage optionality on large markets that are still taking shape.

Threats deserve careful weight rather than dismissal. The hyperscalers’ internal chip programmes are not hobby projects — Google’s TPU programme is a decade old and a genuine alternative for inference workloads. AMD’s MI300 series, while not yet software-competitive, is hardware-competitive and attracting serious enterprise interest. The export control dynamic around China is a genuine structural revenue headwind, not a temporary disruption.

8. Investment Thesis

The bull case rests on three propositions, each of which has substantial supporting evidence. First, the AI infrastructure build-out is still in early innings — the hyperscalers are spending at rates previously reserved for transformative industrial investments, and that spending has a multi-year horizon. Second, Nvidia’s CUDA moat is deeper and more durable than it appears from the outside: the cost of migrating production AI workloads is enormous, and the software ecosystem compounds with every passing year. Third, new phases of AI — particularly inference scaling and agentic systems — may require significantly more compute per dollar of economic output than training-era AI, potentially validating the Jevons Paradox argument for sustained demand. A business generating $72 billion in annual net income, with a credible roadmap for growth, in a sector where demand is structurally expanding, is difficult to dismiss as overvalued purely on the basis of near-term multiples.

The bear case is also serious, and should not be dismissed as mere valuation anxiety. The combination of margin compression (gross margins down from 78% to 61% in a single year is a significant shift), China revenue impairment from export controls, and the real possibility that hyperscaler custom silicon gradually erodes Nvidia’s market share in inference — its fastest-growing workload — constitutes a meaningful risk to the growth narrative. Beyond those specifics, there is a simpler concern: Nvidia is priced for a future in which it dominates an industry that does not yet fully exist, at margins that are historically unprecedented for hardware companies. That future is plausible, but it is not guaranteed, and the stock leaves little cushion for disappointment.

What type of investor does this suit? Nvidia is best suited for a long-term investor with a high tolerance for near-term volatility who believes that AI represents a genuine multi-decade structural transformation of the global economy — and that Nvidia’s software ecosystem gives it the durable competitive advantages required to remain the primary beneficiary of that transformation. It is not a value investment and never will be. It is not a low-risk defensive holding. It is a high-conviction, high-multiple, high-volatility bet on a specific structural thesis about the future of computing. An investor who is wrong about the durability of the CUDA moat, or about the pace of AI infrastructure investment, or about the regulatory trajectory, will find that the current valuation provides very little downside protection.

The most honest conclusion is this: Nvidia is a genuinely exceptional business operating at the centre of the most consequential technology transition in a generation. The economic moat is real, the financial quality is extraordinary, and the management has earned the right to be taken seriously on its long-term vision. But exceptional businesses can be poor investments if purchased at the wrong price, and Nvidia’s valuation already prices in a great deal of the optimistic scenario. The appropriate stance depends less on whether Nvidia is a great company — it clearly is — and more on whether the current market price adequately compensates for the geopolitical risk, competitive evolution, and cyclical uncertainty that are genuine features of the landscape it navigates.

This analysis is prepared for informational and educational purposes only. It does not constitute investment advice, and no part of it should be interpreted as a recommendation to buy, hold, or sell any security. All financial data sourced from Nvidia SEC filings and public disclosures.