§ 01 Business Overview
Nvidia makes the hardware that AI runs on. Not some of it. Most of it. Nvidia holds approximately 80% of the AI accelerator market, with data center revenue reaching $193.7 billion in fiscal year 2026. Every major language model, every recommendation engine, every autonomous vehicle training run, every drug discovery simulation happening at scale right now is running on Nvidia silicon. The company did not plan to be in this position. It spent three decades building the best graphics chips in the world for video games. Then ChatGPT launched in November 2022 and the entire trajectory of the business changed overnight.
Nvidia's fiscal year ends in January. For fiscal year 2026, the company generated $215.9 billion in total revenue, up 65% from the prior year. Data center revenue was $193.7 billion, representing more than 90% of the total business. Gaming, which was once the core franchise, contributed roughly $11.4 billion. Automotive, professional visualization, and other segments make up the remainder.
The product driving everything right now is the Blackwell GPU architecture. Blackwell chips are not incremental improvements on prior generations. They are a fundamental redesign of how AI compute is packaged and delivered, moving from single-GPU configurations to massive interconnected clusters spanning thousands of processors. A single GB200 NVL72 rack contains 72 Blackwell GPUs linked by NVLink, delivering 30 times the AI inference performance of the previous H100 generation. Every Blackwell unit Nvidia can produce is already sold. The company has stated the entire 2026 Blackwell production calendar is committed to customers.
Q1 FY2027 earnings report on May 20, 2026 - eight days from now - with consensus revenue at $78.8 billion and EPS of $1.78. Goldman Sachs is running $80.05 billion in revenue and $1.86 in EPS, both above consensus, and says the bar for stock outperformance is high because expectations are already elevated. Jensen Huang stated at GTC 2026 that Nvidia has demand visibility of over $1 trillion in purchase orders for Blackwell and Vera Rubin systems through 2027. That is not a market size estimate. It is a claimed order book.
§ 02 Competitive Moat · Strong
Nvidia's moat is the most debated in technology right now because it operates on two completely different levels and the bear case only attacks one of them.
The hardware moat. Nvidia designs the best AI accelerator chips in the world. Its gross margins prove it. Nvidia achieves 79 to 88% gross margins on its H100 and B200 chips. AMD's comparable margins are 64 to 68%. Intel's Gaudi 3 operates at 58.4%. A 20-point gross margin premium over the nearest competitor in a hardware business is not a slight edge. It is structural pricing power, earned by building chips with higher performance per watt, per rack unit, and per dollar of total cost of ownership than anyone else. The Blackwell architecture's NVLink interconnect, which allows thousands of GPUs to operate as a single system, is an integration achievement that took years to develop and cannot be replicated by copying the chip design alone.
The software moat: CUDA. This is the more important and more durable layer. CUDA is the programming framework that developers use to write code for Nvidia GPUs. It has been available since 2006 - 20 years of compounding. Over those two decades, millions of researchers, engineers, and data scientists built their tools, models, and workflows on CUDA. PyTorch, TensorFlow, JAX, and virtually every major AI framework runs natively on CUDA. Libraries like cuDNN, cuBLAS, and NCCL are baked into the infrastructure of modern AI research. Switching away from CUDA does not mean buying a different chip. It means rewriting every tool, framework, and production pipeline your organization runs, then retraining every engineer who built on the old stack. That is not a purchasing decision. It is a multi-year infrastructure migration. As more developers build on CUDA, more tools and libraries emerge, attracting additional developers - a classic network effect that compounds Nvidia's competitive position with each passing year.
The custom silicon threat is real but it attacks inference, not training. Google's TPUs, Amazon's Trainium, Microsoft's Maia, and Meta's MTIA are all optimized for inference workloads - running already-trained models to generate predictions at scale. For inference, the CUDA moat is thinner because the customer controls the stack and can optimize for a narrower task. For training, where the workload is complex, iterative, and deeply integrated with research tooling, CUDA's advantage is close to absolute. Market share will decline from 87% in 2024 to approximately 75% in 2026 as AMD and custom silicon scale, but absolute revenue continues growing from $100 billion to over $150 billion because the total market is expanding faster than share declines. Losing share in a market growing at 40% per year is not a crisis. It is arithmetic.
The newest layer of the moat is sovereign AI. Sovereign AI revenue exceeded $30 billion in fiscal year 2026, more than tripling year over year, as governments in Canada, France, the Netherlands, Singapore, and the UK build national AI infrastructure on Nvidia platforms. Sovereign buyers are not price-sensitive in the way hyperscalers are. They are buying strategic capability with national security implications. Once a government builds its AI infrastructure on Nvidia hardware and CUDA software, the switching cost is political as well as technical.
§ 03 Financial Snapshot
The three-year revenue trajectory is unlike anything in the history of large-cap public companies. From $26.97 billion to $215.93 billion in three fiscal years - an 8x increase in revenue at a company that was already a large-cap semiconductor business. Operating income went from $5.58 billion to $130.39 billion across the same window. For context, $130 billion in annual operating income puts Nvidia in the same conversation as Apple and Saudi Aramco on absolute profit generation.
| Fiscal Year | Revenue | Operating Income | Op. Margin | Net Income | Net Margin |
|---|---|---|---|---|---|
| FY2023 (Jan 2023) | $26.97B | $5.58B | 20.7% | $4.37B | 16.2% |
| FY2024 (Jan 2024) | $60.92B | $32.97B | 54.1% | $29.76B | 48.9% |
| FY2025 (Jan 2025) | $130.50B | $81.45B | 62.4% | $72.88B | 55.8% |
| FY2026 (Jan 2026) | $215.93B | $130.39B | 60.4% | ~$118B+ | ~55%+ |
The direction of margin tells an important story. Operating margin went from 20.7% in FY2023 to 62.4% in FY2025, because data center GPUs carry fundamentally different economics than gaming chips. A gaming GPU sells to a consumer for $500 to $1,500. An H100 sells to a hyperscaler for $30,000 to $40,000. An NVL72 rack sells for over $3 million. The product mix shift toward data center is not just a volume story. It is a margin expansion story driven by the pricing power that comes from selling infrastructure that customers have no substitute for.
Gross margin peaked at approximately 78% in Q1 FY2026, then compressed to the 73 to 74% range as the Blackwell architecture launched. New GPU generations always carry higher initial production costs that normalize as yields improve and manufacturing scales. Management guided Q4 FY2026 gross margins to 74.8% GAAP and 75% non-GAAP, suggesting stabilization as Blackwell production scales. The direction of travel on margins heading into FY2027 is stable to slightly recovering, not compressing.
Key metrics heading into May 20 earnings:
| Metric | Figure |
|---|---|
| Q1 FY2027 revenue guidance | $78B +/- 2% |
| Consensus revenue estimate | $78.8B |
| Goldman Sachs estimate | $80.05B |
| Consensus EPS | $1.78 |
| Goldman EPS estimate | $1.86 |
| Implied YoY growth | ~77% |
| FY2027 full-year consensus revenue | $367.7B (+70% YoY) |
| FY2027 consensus EPS | $8.25 (+73% YoY) |
| Forward P/E (FY2027) | ~15 to 18x |
| PEG ratio | 0.63 |
| 3-year historical average P/E | 19.4x |
| Analyst consensus | 48 Buy, 9 Strong Buy, 1 Sell |
| 12-month consensus price target | $269.17 (pre-split equivalent) |
The forward P/E of approximately 15.7x represents a 19% discount to Nvidia's 3-year historical average multiple of 19.4x. Customer concentration is a notable variable: two customers represented 36% of FY2026 revenue. On China, Q1 FY2027 guidance explicitly excludes all China data center compute revenue. Jensen Huang estimated the Chinese market at approximately $50 billion in potential annual revenue - a stream that is effectively zero under current export controls with no clear return timeline. Wells Fargo estimates $25 billion in potential China revenue upside if export policies reverse, representing approximately 9% of current annual revenue run rate.
§ 04 Risk Rating
Export controls are a permanent variable, not a one-time charge. China represented an estimated 20 to 25% of Nvidia's data center revenue before restrictions tightened. That revenue is currently zero in guidance. The $4.5 billion H20 inventory charge in Q1 FY2026 was the accounting recognition of that loss, but the ongoing revenue absence is the larger issue. Any reversal in export policy could deliver meaningful upside, but continued restrictions represent a ceiling on near-term revenue that consensus models cannot fully account for because the policy trajectory is unpredictable. Jensen Huang has argued publicly that restricting older chip sales to China accelerates Huawei's domestic silicon development, which is the opposite of what U.S. policy intends. Whether that argument prevails in Washington is not a financial question. It is a geopolitical one.
Customer concentration at extreme scale. Two customers accounted for 36% of FY2026 revenue. At $215.9 billion in total revenue, 36% is approximately $77 billion flowing from two purchasing relationships. If either of those customers accelerates its custom silicon buildout, delays a major order, or shifts inference workloads off Nvidia hardware, the revenue impact is visible immediately in the quarterly results. The concentration risk is partially mitigated by the CUDA switching cost argument, but partially is not fully.
The inference transition. AI workload economics are shifting. Training large models requires the maximum compute density that only Nvidia can currently deliver at scale. Inference - running trained models to serve users - is more cost-sensitive and more amenable to custom silicon. As AI applications mature from research and development into production deployment, the mix of training versus inference workloads changes. If the inference share of total AI compute grows faster than expected, the hardware moat compresses at the margin because TPUs, Trainium, and AMD MI-series chips are viable inference alternatives. Vera Rubin promises a 10x reduction in inference token cost versus Blackwell, which is Nvidia's answer to this risk, but Rubin does not ramp until the second half of 2026.
The $1 trillion demand statement is not a contract. Jensen Huang's GTC 2026 claim of $1 trillion in demand visibility through 2027 covers Blackwell, Blackwell Ultra, and Rubin systems. That visibility is based on customer conversations, purchase commitments, and order backlog that Nvidia has not fully disclosed at the line-item level. Hyperscaler capex plans can and do change. If Microsoft, Google, Meta, or Amazon moderates AI infrastructure spending in response to slower-than-expected AI monetization, the order book converts to revenue more slowly than the $1 trillion figure implies.
The risk rating is 6 rather than higher because the business is executing at a level that has essentially no peer in semiconductor history, the CUDA moat is genuinely durable for training workloads, the sovereign AI diversification is reducing hyperscaler concentration, and Vera Rubin arriving in the second half of 2026 is a second-half catalyst that current consensus may be undermodeling.
§ 05 Bull vs. Bear
Bull case: The compute industrial revolution framing Jensen Huang uses on every earnings call is not marketing language. It describes something real. Every major technology company in the world is simultaneously rebuilding its infrastructure around AI accelerated compute. Hyperscalers are collectively spending over $630 billion in capital expenditure in 2026. Sovereign governments are building national AI infrastructure. Enterprises are deploying AI in workflows that previously ran on CPUs. OpenAI has committed to at least 10 gigawatts of Nvidia systems. Meta signed up for millions of Blackwell and Rubin GPUs. The demand is not concentrated in one customer or one use case. It is dispersed across every economic sector simultaneously.
At 15 to 18x forward earnings on 70% projected revenue growth, Nvidia is not priced like the dominant infrastructure provider of the AI era. It is priced like a cyclical semiconductor company that might miss next quarter. The PEG ratio of 0.63 means the market is paying less per unit of growth than the growth rate itself would suggest. A company already above $200 billion in annual revenue is expected to grow 71% in fiscal year 2027, and the earnings growth of 73% in the same period is acceleration rather than the deceleration bears had been anticipating.
The China upside is not in the numbers. $25 billion in potential annual China revenue at Nvidia's current gross margins represents significant earnings upside that zero current consensus models include. Any partial policy reversal is asymmetric upside.
Bear case: Excellence is now assumed, not rewarded. Nvidia has beaten earnings expectations for 14 consecutive quarters. At $200 billion-plus in revenue scale, growth naturally becomes harder to accelerate, and investors are beginning to price in the possibility that monetization lags the infrastructure buildout being funded by hyperscalers. The May 20 earnings report requires not just a beat but a beat above an already elevated bar. Goldman expects $80 billion against a $78.3 billion consensus, and even Goldman says the bar for stock outperformance is high.
The custom silicon programs at Google, Amazon, Microsoft, and Meta are not toy projects. They are multi-billion dollar engineering efforts backed by the largest technology companies in the world. For inference specifically, the economics of custom silicon improve with every generation as those companies accumulate production experience. The process is slow, but it is directional. Five years from now, the inference workload that runs on Nvidia today may run predominantly on custom silicon, and Nvidia's revenue base narrows to training, which is a smaller and less predictable market than inference at scale.
The training-to-inference transition timing is the most important unknown in the Nvidia thesis. If inference adoption is slower than expected, training demand extends and Nvidia wins. If inference scales faster than expected and custom silicon captures it, Nvidia's addressable market shrinks before Vera Rubin's economics can compete.
Buy on Weakness. Entry interest: $98 to $108. At current prices around $111 to $115, Nvidia is fairly but not cheaply valued given the earnings event in eight days. Paying above $110 heading into a print where the bar is already elevated and Goldman is warning that a beat alone may not move the stock is not a compelling risk-reward setup. The better entry is on weakness - either a post-earnings pullback if guidance disappoints relative to an elevated bar, or a broader market pullback that takes Nvidia down despite business results remaining intact.
At $98 to $108, the forward P/E compresses to approximately 13 to 15x on FY2027 consensus estimates that themselves exclude China revenue upside and may be undermodeling Vera Rubin's second-half contribution. At that range, the CUDA moat, the $1 trillion order backlog, and the sovereign AI diversification are priced conservatively enough to offer real margin of safety.
The thesis weakens materially if the Vera Rubin ramp slips or hyperscaler capex commentary turns cautious on the May 20 call. Watch for those two signals specifically.
§ 06 What to Watch
Gross margin guidance for Q2 FY2027 and the full year. Goldman expects management to reiterate mid-70% gross margin guidance for calendar 2026. Any softening of that language would matter more than a revenue beat because it signals either Vera Rubin transition costs are running higher than expected or competitive pricing pressure is emerging. Gross margin trajectory is the earliest visible signal of whether the pricing moat is holding.
Vera Rubin ramp timeline specificity. Nvidia has been vague on the precise quarter when Vera Rubin revenue recognition begins in earnest. Any sharpening of that timeline on the May 20 call - with specific shipment volumes or customer commitments - is a positive catalyst. Any softening of Rubin language is a negative one.
Hyperscaler commentary on AI capex through the rest of 2026. Alphabet, Meta, Microsoft, and Amazon all report quarterly and provide capex guidance. If any of them signals moderation in AI infrastructure spending in the back half of 2026, Nvidia's Q3 and Q4 order book is at risk. This is the most important external variable and it will not show up in Nvidia's own earnings call. Watch the hyperscaler calls that come before and after May 20.
China policy developments. Jensen Huang estimated the Chinese market at approximately $50 billion in annual opportunity. Any licensing agreement, policy carveout, or export control modification that reopens even a portion of that market is immediate upside to a number that consensus has set to zero.
Sovereign AI order book updates. The $30 billion in sovereign AI revenue in FY2026 tripled year over year. Management commentary on fiscal 2027 sovereign pipeline - particularly new country commitments or expanded agreements with existing sovereign customers - will signal whether this diversification continues to reduce hyperscaler concentration risk.
This analysis introduced the concept of platform lock-in versus product lock-in, and why they are fundamentally different kinds of moats.
Product lock-in is what most people think of when they hear the word moat. A customer buys your product, finds it better than alternatives, and keeps buying it because switching costs are real but not enormous. Adobe's file format lock-in works this way. ServiceNow's workflow dependency works this way. These are real moats but they can be overcome given enough competitive pressure and a sufficiently better alternative.
Platform lock-in is structurally different. A platform becomes the substrate on which other people build things. Once enough people build on the platform, the platform acquires value that has nothing to do with whether the underlying product is the absolute best. CUDA has been accumulating platform lock-in for 20 years. The reason it is hard to displace is not that CUDA the language is irreplaceable. It is that millions of developers built careers, tools, libraries, and companies on CUDA, and all of that accumulated work only runs on Nvidia hardware. Switching away from CUDA means abandoning not just a product but an entire ecosystem of compounding human capital investment.
This distinction matters for understanding why the custom silicon threat is real for inference but not for training. Inference is running a finished model against new inputs. The workload is narrow, well-defined, and optimizable for a specific chip architecture. Platform dependency matters less because you are not iterating on the model or integrating with a research ecosystem. Training is building the model from scratch, iterating on it, debugging it, and integrating it with every tool in a research pipeline that was built on CUDA. Replacing CUDA in a training workflow means replacing the entire research stack, not just the chip. That is platform lock-in in practice, and no competitor has come close to breaking it.