NVDA — NVIDIA Corporation

NVIDIA controls roughly 85% of the AI accelerator market through a moat that is structural rather than circumstantial — the CUDA platform represents nineteen years of accumulated software optimization, tooling, and developer expertise that a competitor with better silicon still cannot displace overnight. At a forward earnings multiple of 21 times, the stock appears superficially reasonable, but that multiple is applied to projected earnings that assume today's AI infrastructure supercycle sustains without interruption. Good business, meaningfully overpriced.

---

The scale of money flowing into AI infrastructure is without precedent in the history of technology investment. Hyperscalers — Microsoft, Google, Amazon, and Meta — disclosed combined capital expenditure plans exceeding $380 billion for 2025, the majority directed toward data centers and the accelerated compute clusters that power large language models, inference at scale, and the next generation of AI applications. This number is not speculative; it is a sum of publicly committed budgets from four of the five largest companies by market capitalization. The bet being made is that AI will restructure every major industry and that the companies who own the infrastructure layer own a toll road into that future. NVIDIA sits at the center of this bet, collecting the toll.

The consensus position on NVIDIA has coalesced into something almost theological: CUDA is unassailable, Jensen Huang is a visionary, and the AI buildout makes NVIDIA's dominance self-reinforcing. There is truth in this view. There is also danger in it. When a narrative becomes consensus, the narrative is in the price. The analytical question is not whether NVIDIA is exceptional — it plainly is — but whether exceptional is worth any price, and how much of the future the current price has already consumed.

The AI semiconductor market presents a structure that would have seemed fantastical five years ago. In 2024, the global AI accelerator market generated roughly $28 billion in revenue. Independent projections for 2032 range from $256 billion to $363 billion, implying compound annual growth rates of 29 to 37 percent. These projections carry the uncertainty that always accompanies extrapolation at extreme growth rates, but their direction is not in question: the world is building AI infrastructure at a rate that has no historical parallel, and that infrastructure requires specialized silicon that general-purpose processors cannot efficiently replace.

The structural reason GPUs dominate AI training is mathematical. Neural network training and inference are embarrassingly parallel computations — the same operation applied to millions of data points simultaneously. The GPU was designed for exactly this pattern: thousands of small processing cores executing identical instructions in lockstep across large data arrays. Intel's general-purpose CPU, optimized for sequential processing and branching logic, is structurally the wrong tool for the job. Custom silicon like Google's TPU addresses specific model architectures efficiently but lacks the generality to run every framework on every workload. The GPU's combination of programmability and parallel throughput explains why the world's AI models run on it — and why that is unlikely to change even as alternatives improve at the margins.

NVIDIA's business today is not recognizable as the graphics chip company of ten years ago. In fiscal year 2025, data center revenue reached $115.2 billion, representing 88% of total company revenue of $130.5 billion. The remaining 12% is split between gaming ($11.4 billion), professional visualization, and automotive — businesses that remain meaningful but are not the investment thesis. The data center segment sells three distinct things: the GPU hardware itself (Hopper generation in FY2025, Blackwell in FY2026), the networking fabric that connects those GPUs into coherent clusters at rack scale (NVLink and NVSwitch), and increasingly, software subscriptions under the NVIDIA AI Enterprise umbrella — a per-GPU, per-year licensing model that transforms one-time hardware revenue into recurring software revenue. The Blackwell GB300 NVL72 — a rack-scale system integrating 72 Blackwell GPUs with NVLink interconnect — is sold as a complete AI factory, not as individual chips. This shift from component to system changes the competitive dynamic: a hyperscaler cannot simply swap in AMD Instinct accelerators without replacing the entire rack architecture.

The CUDA moat is the single most important fact about NVIDIA as an investment. CUDA is a parallel computing platform and programming model, first released in 2006, that allows developers to write code targeting NVIDIA GPUs using C, C++, and Python. Around CUDA, NVIDIA has built a library ecosystem — cuDNN for deep neural networks, cuBLAS for linear algebra, NCCL for multi-GPU communication, the Nsight toolchain for profiling and debugging — that represents nineteen years of accumulated optimization work. PyTorch and TensorFlow, the two dominant AI frameworks, run on CUDA as their primary execution path. The global population of developers who have built skills around CUDA numbers in the millions.

The economic consequence of this installed base is switching costs that exceed any realistic performance gap. A hyperscaler considering AMD's Instinct MI450 — expected to arrive in 2026 on a 2nm process — faces a calculus that is not simply FLOPS-per-dollar. It must weigh the cost of porting its inference stack, retraining engineering teams, losing CUDA-specific library performance, and accepting the operational risk of a platform with far smaller community support and fewer documented solutions. NVIDIA's own research and the operational experience of hyperscalers who have tested alternatives consistently shows that raw hardware performance advantages below roughly 30 to 40 percent do not trigger migration; the switching cost exceeds the performance gain. AMD has not breached this threshold at scale. The data confirms the claim:

Company	AI GPU Market Share (2025)	GAAP Gross Margin	Software Ecosystem
NVIDIA	~85%	~74%	CUDA (19 yrs, millions of developers)
AMD	~10%	~50%	ROCm (limited enterprise adoption)
Intel	~2%	~43%	oneAPI (early stage)
Google TPU	Internal use only	N/A	Internal only

The 24 percentage point gross margin gap between NVIDIA and AMD is not a temporary product-cycle advantage. It reflects the pricing power that comes from selling into a captured ecosystem. When NVIDIA raises its prices — as it effectively did by moving from A100 to H100 to Blackwell at successively higher ASPs — customers pay, because the alternative is a multi-year re-platforming project. AMD's gross margin of 50% is not bad in absolute terms; it is simply the margin of a company competing on price and performance against a platform that has already captured its customers' workflows. Intel's 43% reflects an even weaker competitive position. The margin gap is the moat, expressed in numbers.

The honest bear case on the moat is not AMD. It is Google, Amazon, and Meta. These four hyperscalers — NVIDIA's four largest customers, representing 40 to 50 percent of the company's revenue — are building custom silicon specifically designed to reduce their dependence on NVIDIA. Google's TPU v6 (Trillium) is now deployed at scale for internal model training. Amazon's Trainium 3 targets cost-per-FLOP efficiency in training workloads. Meta's MTIA v3 is the most ambitious internal AI compute program among the four. The addressable question is not whether these custom chips work — they do — but how large a fraction of total AI compute spend they can realistically replace. If Google and Meta divert 15 to 20 percent of their AI compute spend to internal silicon, the headwind to NVIDIA is $15 to $25 billion in annual revenue — meaningful but not existential, since the overall market is expanding fast enough to partially absorb the displacement. The risk is structural, not cyclical: as these customers become more capable of building their own silicon, the fraction of their spend flowing to NVIDIA may decline even as total spend grows. It has not happened yet. It is worth watching for in the gross margin trajectory.

NVIDIA's financial results for fiscal year 2025 defy easy comparison to prior periods. Revenue of $130.5 billion grew 114 percent year over year. Free cash flow reached $60.9 billion, up 125 percent. GAAP gross margins for the full year came in at approximately 74.6 percent — a level that reflects both the hardware pricing power described above and the increasing mix of Blackwell systems, which command higher ASPs than their Hopper predecessors. GAAP earnings per share for FY2025 were $2.94 on a post-split basis, up 147 percent year over year. The reported numbers are not distorted by significant one-time items in the annual figures; the GAAP-to-non-GAAP reconciliation shows non-GAAP EPS of $2.99, a $0.05 per share difference that reflects primarily stock-based compensation adjustments rather than material accounting differences.

In fiscal 2026, the pattern continued through the first three quarters: quarterly revenues of $44.1 billion, $46.7 billion, and $57.0 billion respectively. The $57.0 billion Q3 FY2026 result set a company record, with GAAP gross margins of 73.4 percent. One exception requires note: Q1 FY2026 GAAP gross margin was 60.5 percent, depressed by a $5.5 billion inventory charge related to the H20 GPU product, which was subject to expanded U.S. export controls restricting sales to China. This charge is genuinely non-recurring — it reflects a specific geopolitical event, not a deterioration in underlying economics. On a normalized basis, Q1 FY2026 gross margins were consistent with the broader trend. The H20 situation is nonetheless instructive about the China exposure: analysts estimate the total regulatory impact of U.S.-China trade restrictions at $15 to $16 billion in annual revenue, representing a market NVIDIA has effectively lost access to.

Jensen Huang co-founded NVIDIA in 1993 and has run it for 32 years. The company he leads today generates more revenue in a single quarter than the entire business did in fiscal year 2022. That transformation was not accidental — Huang placed a decisive bet on general-purpose GPU computing with CUDA's launch in 2006, at a time when the market for GPU compute outside gaming was essentially zero. The bet compounded. Under his direction, NVIDIA has maintained a one-year hardware cadence that keeps the performance lead consistent, and has invested in software infrastructure that competitors have consistently underinvested in. Capital allocation has been shareholder-friendly: the company authorized a $50 billion share repurchase program in August 2024, executed over $45 billion in repurchases over the prior ten quarters, and returned $15.4 billion to shareholders in a single quarter in fiscal 2026. The diluted share count is declining despite meaningful equity compensation. Huang personally holds approximately 854 million shares — a position so large that his 64 reported sell transactions over five years, even at elevated prices, represent a small fraction of a sustained, substantial stake. He has not sold conviction; he has harvested liquidity from a position that has compounded beyond any practical need to hold the full amount.

The growth trajectory is best understood by watching data center revenue, GAAP gross margin, and free cash flow together — these three numbers make it impossible to tell a story about NVIDIA that isn't grounded in evidence.

Fiscal Year	Data Center Revenue	YoY Growth	GAAP Gross Margin	Free Cash Flow
FY2021	$6.7B	—	63.3%	$4.7B
FY2022	$11.0B	+64%	64.9%	$8.1B
FY2023	$15.0B	+36%	56.9%	$3.8B
FY2024	$47.5B	+217%	72.7%	$27.0B
FY2025	$115.2B	+142%	74.6%	$60.9B

Two inflection points dominate this table. The first is FY2023 to FY2024: data center revenue tripled in a single year as the first wave of large language model training drove hyperscalers to purchase every H100 NVIDIA could produce. The second is FY2024 to FY2025: revenue more than doubled again as the Blackwell architecture ramped into production and the addressable market expanded from model training into inference-at-scale. The gross margin trajectory is almost as striking as the revenue line: margins fell in FY2023 as NVIDIA worked through the post-COVID inventory normalization in its legacy gaming business, then snapped back sharply as data center mix overwhelmed the lower-margin segments. The FY2025 gross margin of 74.6 percent is not only the highest in the company's history — it is higher than almost every other semiconductor company in the world, achieved at $130 billion in revenue. The pricing power is real.

The structural driver of this trajectory is the AI infrastructure investment cycle. Hyperscalers collectively spent over $380 billion on capital expenditures in 2025, with a substantial majority directed toward data center capacity and the compute clusters inside those data centers. Industry forecasts project data center AI capital spending growing at approximately 40 percent annually through 2030 — a figure that would make it the largest single capital expenditure category in the global economy within this decade. NVIDIA, with 85 percent market share in AI GPU hardware, captures a disproportionate fraction of this spend.

The penetration argument for NVIDIA is unusual relative to most businesses analyzed this way. The addressable population is not consumers or enterprises in the traditional sense — it is the computational capacity being installed to run AI models. By one estimate, total AI infrastructure spending from now through 2030 will exceed $3 trillion. NVIDIA's FY2025 data center revenue of $115 billion represents roughly 4 percent of that cumulative spend. Even if NVIDIA's market share erodes from 85 to 60 percent over this period due to hyperscaler custom silicon development, and even if overall AI infrastructure spending grows at half the projected rate, NVIDIA's annual revenue opportunity through the end of the decade is substantially larger than today's $130 billion total. This is not a company approaching market saturation. The question is not whether the opportunity is large. The question is what to pay for access to it.

At approximately $140 per share, NVIDIA trades at a trailing price-to-earnings multiple of roughly 35 times based on FY2025 GAAP earnings of $2.94 per share, and at a forward multiple of approximately 21 times based on FY2027 earnings projections. The 21 times forward figure is the one that generates optimism. It is below the semiconductor sector median of 28 times forward earnings, and for a company with NVIDIA's competitive position, a below-sector multiple feels like an anomaly. It is not an anomaly — it reflects that the "forward" earnings in the denominator embed continued rapid growth. Forward earnings of approximately $6.70 per share by FY2027 require data center revenue to sustain close to the current quarterly run rate for two years, with minimal disruption from export restrictions, hyperscaler custom silicon, or cyclical digestion.

The framework for evaluating purchase price requires normalized pre-tax earnings — what the business earns at a normal point in its cycle. This is where NVIDIA presents a genuine analytical challenge. The business has grown 114 percent in a single year and is in the third year of an AI infrastructure supercycle. FY2025 is not a trough. It may be mid-cycle, or it may be approaching peak. The normalized pre-tax earnings base for FY2025 is approximately $83.4 billion, or $3.42 per share at the current share count — derived from $71.7 billion in GAAP net income divided by a 14 percent effective tax rate. At $140 per share, that is a multiple of 41 times normalized pre-tax earnings — far above the 15 times threshold at which growth stops being required to justify ownership. The buy price implied by this baseline is $51 per share, roughly 64 percent below the current price.

The objection to this calculation is fair: using FY2025 as the normalized base ignores that the business has continued growing materially since then. The Q3 FY2026 quarterly result of $57 billion in revenue annualizes to approximately $228 billion — implying a run-rate pre-tax earnings base closer to $6.40 per share. At $140, that gives a multiple of 22 times normalized pre-tax — still above 15 times, but close enough that the conversation changes. The buy price at that earnings level would be approximately $96 per share. The honest answer is that the normalized earnings base sits somewhere between $3.42 and $6.40 per share, depending on where the AI infrastructure cycle plateaus. The market is pricing the midpoint optimistically.

The most intelligent bear on NVIDIA makes a specific argument: the current AI capex cycle is driven by hyperscaler competition for AI positioning — a race to deploy capacity before competitors — rather than by demonstrated economic returns on that capacity. When return-on-investment discipline reasserts itself, which it eventually must, AI capex growth decelerates. NVIDIA's revenue, currently growing at 100 percent annually, mean-reverts to the growth rate of the underlying compute demand — perhaps 20 to 30 percent annually rather than 100 percent. In that scenario, today's $140 price, priced for sustained hypergrowth, reflects multiple compression risk that more than offsets the underlying business compounding. The counter is that even a 20 to 30 percent AI compute growth rate sustains NVIDIA's revenue trajectory for a decade, and the CUDA moat means margin and market share erosion is gradual rather than sudden. Both arguments are internally consistent. The bear does not need a catastrophe; it only needs deceleration.

For the stock to become compelling at this level of business quality, either the share price needs to decline toward the $90 to $100 range — where normalized pre-tax earnings at the current run rate approach 15 times — or the quarterly revenue run rate needs to sustain materially above $57 billion for long enough to establish that figure as the new normalized floor. A second scenario: export restrictions easing, particularly if U.S.-China trade policy shifts, could restore $15 billion or more in annual revenue and move the normalized earnings base upward without requiring multiple expansion. A third: hyperscaler custom silicon development stalls or delivers disappointing performance, removing the structural headwind to NVIDIA's market share at its four largest customers. None of these scenarios are implausible. None of them are certain.

NVIDIA has built the most defensible platform in the history of semiconductors, and Jensen Huang has demonstrated for three decades that he knows how to extend it. The CUDA moat is real, the gross margin premium is validated, and the AI infrastructure opportunity is genuinely large. None of that is in dispute. What is in dispute is the price. At $140, the stock prices in a future that is probable but not certain, at a premium that leaves no room for the future to be merely good rather than perfect. The company deserves to be owned — at the right price. The right price is not today's.