Sources: SEC EDGAR (10-K FY2026 · 8-K Q4/FY2026) · NVIDIA Newsroom · NVIDIA Blog · IEA, Electricity Mid-Year Update 2025 · April 2026
The Five Things That Matter
Nvidia is the dominant supplier of GPU-based computing infrastructure for artificial intelligence. Its Data Center segment generated $193.7 billion in FY2026 — more than TSMC’s entire annual revenue — growing at 68% year-on-year. Here are the five things that matter most.
Five key points:
1. The Data Center segment is the company. FY2026 Data Center revenue of $193.7 billion was 90% of total revenue, grew 68% year-on-year, and produced $62.3 billion in a single quarter. Q1 FY2027 guidance of $78.0 billion assumes zero revenue from China — that growth rate without one of the world’s largest markets.
2. CUDA is the moat, not the hardware. Over 20 years, the global AI developer community has written an enormous body of code in Nvidia’s CUDA software platform. None of it runs natively on competing hardware. Switching GPU vendors means rewriting production software — an organisational cost, not just a financial one.
3. Customers are committed to future generations, not just the current one. Meta’s partnership is explicitly “multigenerational.” OpenAI’s next-generation model training depends on Vera Rubin delivery, backed by up to $100B in Nvidia co-investment. CoreWeave is committed to 5 GW of AI factories by 2030 on Nvidia hardware.
4. Nvidia is building in the United States for the first time. Blackwell wafers at TSMC Arizona, supercomputer plants in Houston and Dallas, $500 billion in planned US AI infrastructure over four years. This reduces geopolitical exposure and builds domestic political capital for export control negotiations.
5. Three risks qualify the picture. The China market is excluded from guidance at roughly $32,000 per H200 unit. Grid capacity — not GPU supply — may become the binding data centre constraint. And Vera Rubin’s 10x inference cost reduction compresses hardware spend required per AI workload.
Key Metrics at a Glance
FY2026 Revenue
$215.9B
+65% year-on-year
Data Center Revenue
$193.7B
+68% year-on-year
Q4 FY26 Revenue
$68.1B
+73% year-on-year
R&D Spend FY26
$18.5B
vs. $8.7B in FY2024
Q1 FY27 Guidance
$78.0B
±2%, ex-China
Market Cap
~$4.0T
as of July 2025
Sources: NVIDIA Corporation Form 8-K Q4/FY2026 (February 25, 2026); Form 10-K FY2026 (February 25, 2026). SEC EDGAR CIK 0001045810.
I. Financial Architecture: A Platform Monetising Its Own Infrastructure Cycle
Nvidia's fiscal year does not align with the calendar year.
FY2026 covers January 27, 2025 through January 25, 2026 — almost entirely calendar year 2025. This matters when comparing figures against companies with December year-ends. All Nvidia SEC filings use this convention.
Nvidia closed FY2026 with $215.9 billion in total revenue, up 65% on the prior year. The Data Center segment alone — GPUs and the accelerated computing infrastructure around them — produced $193.7 billion, a 68% year-on-year increase. (NVIDIA Corporation 2026a) That single business line now generates more annual revenue than TSMC.
Q4 told the same story in sharper relief. Revenue of $68.1 billion for the quarter was 20% above Q3 and 73% above Q4 of the prior year; Data Center came in at $62.3 billion for those three months. (NVIDIA Corporation 2026a) Management guided Q1 FY2027 at $78.0 billion (±2%) and noted explicitly that the figure assumes zero Data Center compute revenue from China. Growth of that order with an entire major market excluded from the model.
Why gross margin is the more revealing number.
Gross margin is the percentage of each revenue dollar left after subtracting direct production costs. It signals pricing power and the ability to fund research and expansion. Nvidia’s FY2026 GAAP gross margin of 71.1% annually and 75.0% in Q4 is exceptionally high for a hardware manufacturer, where 40–50% is typical. It means Nvidia is not buying market share — it is selling at a premium customers accept.
Gross margins of 75.0% for Q4 and 71.1% for the full year (GAAP) confirm Nvidia is not buying growth through price. (NVIDIA Corporation 2026a) R&D spending: $18.5 billion in FY2026, against $12.9 billion in FY2025 and $8.7 billion in FY2024 — roughly 46% compounding annual growth over two years. (NVIDIA Corporation 2026b) Capital expenditure went from $1.1 billion in FY2024 to $6.0 billion in FY2026. Total R&D since inception has passed $76.7 billion. (NVIDIA Corporation 2026b)
Nvidia returned $41.1 billion to shareholders in FY2026 through buybacks and dividends, with $58.5 billion in repurchase authorisation remaining at year-end. (NVIDIA Corporation 2026a) Jensen Huang: “Enterprise adoption of agents is skyrocketing. Our customers are racing to invest in AI compute — the factories powering the AI industrial revolution and their future growth.” (NVIDIA Corporation 2026a)
II. What Nvidia Does: Platform, Hardware Roadmap, and Ecosystem Lock-in
Nvidia's business has two interlocking components: a GPU hardware roadmap that forces continuous customer reinvestment, and a software ecosystem — CUDA — that makes switching to a competitor's hardware prohibitively expensive. Together they form a compound lock-in that no competitor has replicated in full.
The Hardware Roadmap: Planned Obsolescence at Scale
Each GPU generation is so superior to its predecessor that customers have a strong economic case for upgrading whether or not an alternative exists. In the FY2026 Q4 filing, Nvidia disclosed the headline claim for the Vera Rubin platform: six new chips delivering “up to a 10x reduction in inference token cost, compared with the NVIDIA Blackwell platform.” (NVIDIA Corporation 2026a) Blackwell had already moved the bar substantially — “up to 50x better performance and 35x lower cost for agentic AI compared with the NVIDIA Hopper platform.” (NVIDIA Corporation 2026a)
What is inference, and why does its cost drive AI adoption?
Training is the initial process of teaching an AI model — performed once or periodically. Inference is what happens every time the model is used in production: generating a response, classifying an image, recommending content. Because inference runs continuously at massive scale, the per-unit cost — measured in “tokens” — is the primary economic constraint on AI deployment. A 10x inference cost reduction does not merely improve margins; it makes entire categories of AI application economically viable that were not before.
AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure are all confirmed as “among the first to deploy Vera Rubin-based instances” (NVIDIA Corporation 2026a) — the same four currently running Blackwell. They migrated together from Hopper to Blackwell. They are migrating together again. Nvidia's CUDA stack and NVLink interconnects are built into how these companies plan their infrastructure.
CUDA and NVLink: the software moat and hardware dependency.
CUDA (Compute Unified Device Architecture) is Nvidia’s proprietary software platform. Over two decades, an enormous body of AI research code, production software, and developer tooling has been written in CUDA — none of which runs natively on competing hardware. Switching means rewriting that software: an organisational cost, not just a financial one. This moat deepens with every additional line of CUDA code written globally.
NVLink is Nvidia’s proprietary high-speed inter-GPU interconnect. Workloads optimised for NVLink bandwidth do not transfer cleanly to competing architectures, adding a hardware dependency layer on top of CUDA.
The OpenAI partnership ties the hardware roadmap to a specific infrastructure commitment: “The first gigawatt of NVIDIA systems will be deployed in the second half of 2026 on the NVIDIA Vera Rubin platform.” (NVIDIA Corporation 2025a) Nvidia has committed up to $100 billion in OpenAI, disbursed per gigawatt deployed. OpenAI’s model development timeline is directly tied to Nvidia’s production schedule.
Interactive: The Three-Layer Lock-in
Interactive Explainer
Nvidia: The Three-Layer Lock-in
Click any layer to see how it traps customers — and why all three together make switching effectively impossible.
1
Hardware Layer
GPU Platform Cadence
+
How it works: Each GPU generation delivers performance gains so large that staying on the previous platform becomes economically difficult. Hopper gave way to Blackwell at 50x better performance per dollar. Blackwell Ultra will give way to Vera Rubin at 10x lower inference cost.
The trap: AI workloads grow faster than any single GPU generation can sustain efficiently. Customers who don’t upgrade fall behind competitors who do. This is not vendor pressure — it is an economic forcing function embedded in the technology itself.
Platform sequence: Hopper → Blackwell → Blackwell Ultra → Vera Rubin. AWS, Google Cloud, Azure, and Oracle confirmed as first deployers of each generation.
↓
2
Software Layer
CUDA Ecosystem & NVLink
+
How it works: CUDA is Nvidia’s proprietary software platform for GPU computing. Over 20+ years, the global AI developer community has written an enormous body of code in CUDA — research libraries, production models, optimised kernels. None of this runs natively on competing hardware architectures.
The trap: Switching from Nvidia GPUs requires rewriting or retooling years of CUDA-dependent software. The cost is not just financial — it is time, expertise, and the risk of performance degradation during migration. And since competitors’ software ecosystems are far thinner, the rewritten code will initially underperform CUDA equivalents.
This moat deepens with every additional line of CUDA code written anywhere in the world. Nvidia does not have to do anything to grow it.
↓
3
Contract Layer
Multi-Generational Partnerships
+
How it works: Nvidia’s largest customers are now contractually committed not just to the current GPU generation but to the next one as well. Meta’s partnership is explicitly “multigenerational.” OpenAI’s next-generation model training is conditioned on Vera Rubin delivery. CoreWeave has committed to 5 GW of AI factories by 2030 — all on Nvidia hardware.
The trap: Switching vendors now would require renegotiating or breaking partnerships that extend years into the future. For OpenAI, switching means delaying next-generation model training. For Meta, it means rebuilding infrastructure planning around an entirely new supplier relationship. The cost is not just hardware — it is time and competitive disadvantage during the transition.
Nvidia has invested up to $100B in OpenAI, disbursed per gigawatt deployed. This is no longer a supplier-customer relationship. It is a shared roadmap.
Lock-in Compound Effect
Each layer reinforces the others. A customer who could tolerate the hardware switching cost would still face CUDA migration. A customer who rewrote their CUDA code would still need to renegotiate multi-generational contracts.
The result
No competitor has yet replicated hardware performance leadership + CUDA ecosystem depth + multi-generational customer commitments simultaneously. Each is achievable individually. The combination is not.
Sources: NVIDIA Corporation Form 8-K (Q4 FY2026) and Form 10-K (FY2026), SEC EDGAR; NVIDIA Newsroom (OpenAI partnership, September 22, 2025). All figures fetch-verified.
III. Demand Environment: IEA Data on Data Centre Electricity and Capital
About the IEA source used here.
The International Energy Agency is an OECD-framework intergovernmental organisation with no commercial interest in its published data. The Electricity Mid-Year Update 2025 was prepared by the IEA’s Gas, Coal and Power Markets division and verified in full from the original PDF document.
Revenue figures tell you what Nvidia has already sold. Energy data tells you what the demand pipeline looks like going forward. The IEA's Electricity Mid-Year Update 2025 quantifies the infrastructure commitment the AI industry is making and how much power it will need to run.
The $320 Billion Hyperscaler Capital Commitment
The IEA states: “Investment in artificial intelligence and data centres continues to accelerate, with companies such as Meta, Amazon, Alphabet and Microsoft committing to spend USD 320 billion in 2025, up from USD 230 billion the previous year.” (IEA 2025, 13) A 39% year-on-year increase in AI infrastructure capital commitments by the four largest cloud operators is the most direct macroeconomic proxy for Nvidia Data Center demand.
What is a terawatt-hour (TWh)?
A terawatt-hour equals one trillion watt-hours — the entire United Kingdom consumes roughly 300 TWh per year. The IEA projects US data centre electricity consumption will increase by approximately 240 TWh by 2030 — equivalent to adding a large European country’s entire annual power demand to the US grid, driven almost entirely by AI computing infrastructure.
On consumption: US data centres used “around 180 TWh of electricity in 2024,” with demand “expected to steadily rise through 2030, with consumption projected to increase by approximately 240 TWh relative to 2024 levels.” (IEA 2025, 13–14) The hardware responsible for most of that consumption is Nvidia GPUs. Electricity growth is an independent physical proxy for Nvidia’s Data Center revenue — measurable without relying on the company’s own reported figures.
US Macro Electricity Demand Growth
US electricity demand grew 2.1% in 2024. The IEA projects 2.3% growth in 2025 and 2.2% in 2026 — rates it describes as “more than double the average growth rate over the past decade,” (IEA 2025, 13) attributing this to “the rapid expansion of data centres in both 2025 and 2026.” (IEA 2025, 11) Aggregate US demand is projected to rise from 4,112 TWh in 2024 to 4,296 TWh in 2026. (IEA 2025, 29) Utilities and grid operators are planning around these numbers.
“A major driver of this demand growth in the United States is the expansion of data centres, which consumed around 180 TWh of electricity in 2024, according to the IEA’s Energy and AI report. Investment in artificial intelligence and data centres continues to accelerate, with companies such as Meta, Amazon, Alphabet and Microsoft committing to spend USD 320 billion in 2025, up from USD 230 billion the previous year.”
— IEA, Electricity Mid-Year Update 2025, p. 13
For Nvidia, the sequence is circular. Hyperscaler capital commitments become GPU purchase orders. Those GPUs generate the electricity consumption the IEA is measuring. The AI workloads justify further infrastructure expansion, which feeds back into GPU procurement. What the IEA is tracking as an energy phenomenon is, at its source, a Nvidia revenue phenomenon.
IV. Strategic Partnerships and Manufacturing Repositioning
The FY2026 Q4 8-K reads, in part, as a list of industries Nvidia is entering. The common thread: each partnership plants Nvidia hardware inside workflows that are expensive to rebuild around a different vendor.
Hyperscalers and Foundation Model Providers
Nvidia formalised “a multiyear, multigenerational strategic partnership with Meta spanning on-premises, cloud and AI infrastructure, including the large-scale deployment of NVIDIA CPUs, networking and millions of NVIDIA Blackwell and Rubin GPUs.” (NVIDIA Corporation 2026a) “Multigenerational” means Meta’s procurement is synchronised with Nvidia’s hardware release schedule, not placed on the open market quarter by quarter. An “investment and deep technology partnership with Anthropic” is also disclosed, with Anthropic running Claude on Microsoft Azure on Nvidia hardware. (NVIDIA Corporation 2026a)
Physical AI, Life Sciences, and Government
Nvidia’s Cosmos and Isaac GR00T physical AI models have been adopted by Boston Dynamics, Caterpillar, Franka Robotics, LG Electronics, and NEURA Robotics. (NVIDIA Corporation 2026a) Siemens and Nvidia are building an “industrial AI operating system”; Dassault Systèmes and Nvidia an “industrial AI platform powering virtual twins.” (NVIDIA Corporation 2026a) A co-innovation AI lab with Eli Lilly targets drug discovery, alongside “a major expansion of NVIDIA BioNeMo.” (NVIDIA Corporation 2026a) Nvidia has joined the U.S. Department of Energy’s Genesis Mission as a private industry partner. (NVIDIA Corporation 2026a) CoreWeave is targeting “more than 5 gigawatts of AI factories by 2030” on Nvidia hardware — contracted partners alone represent multi-year GPU absorption at gigawatt scale. (NVIDIA Corporation 2026a)
US Manufacturing: $500 Billion Over Four Years
Until 2025, Nvidia designed chips in California and manufactured them entirely in Taiwan. Any disruption to Taiwan semiconductor production would simultaneously halt global AI chip supply with no short-term remedy.
Nvidia’s blog, published April 14, 2025: “Within the next four years, NVIDIA plans to produce up to half a trillion dollars of AI infrastructure in the United States through partnerships with TSMC, Foxconn, Wistron, Amkor and SPIL.” (NVIDIA Corporation 2025b) More than one million square feet of manufacturing space has been commissioned for Blackwell production in Arizona and AI supercomputers in Texas, with plants in Houston (Foxconn) and Dallas (Wistron) expected within 12–15 months of that announcement. (NVIDIA Corporation 2025b) Volume production at TSMC’s Phoenix facility began October 2025.
“The engines of the world’s AI infrastructure are being built in the United States for the first time. Adding American manufacturing helps us better meet the incredible and growing demand for AI chips and supercomputers, strengthens our supply chain and boosts our resiliency.”
Export controls have already taken a visible bite: the Q4 FY2026 guidance assumes zero China Data Center compute revenue. (NVIDIA Corporation 2026a) US manufacturing reduces exposure to further tightening and builds a domestic political constituency — a workforce in Arizona and Texas with a direct economic stake in Nvidia’s continued scale.
V. Risks and Assessment
China Exclusion
~$32,000/unit H200 market absent from all near-term guidance
Power Constraint
Grid capacity, not GPU supply, may become the AI infrastructure bottleneck
Unit Economics
10x Vera Rubin cost reduction compresses hardware spend per AI workload
China Market Exclusion
The Q1 FY2027 guidance of $78.0 billion explicitly excludes China Data Center compute revenue. (NVIDIA Corporation 2026a) At approximately $32,000 per H200 unit, China represents a large revenue opportunity currently inaccessible under export control constraints. Further tightening would widen the gap. The guidance is achievable without China — but the constrained potential is real.
Power Infrastructure as Binding Constraint
The IEA projects US data centre electricity consumption will increase by 240 TWh between 2024 and 2030. (IEA 2025, 13–14) Grid permitting timelines, interconnection queues, and transmission capacity — not GPU availability — may determine how fast that infrastructure can be built. A slower data centre construction rate reduces GPU deployment regardless of how much hardware is on order.
Unit Economics and Demand Compression
Vera Rubin’s 10x inference cost reduction will grow the market for AI applications. But it also means a given AI workload in 2027 will require substantially less hardware spend than the same workload in 2025. Performance improvements of this magnitude compress the capital expenditure needed for equivalent compute over time.
Assessment
Nvidia has built its market position across several layers simultaneously: hardware performance that forces upgrades, a software ecosystem that makes switching costly, customer commitments running across multiple product generations, and a domestic manufacturing base reducing geopolitical exposure. The combination — reinforced by $18.5 billion in annual R&D and $320 billion in committed hyperscaler capex — has no current competitor that replicates it in full. What the three risks address is the further question of how durable this position is over the two-to-four-year horizon that current market valuations implicitly price.
Terms Explained
GPU (Graphics Processing Unit). Originally for video game rendering; its parallel architecture proved ideal for AI. Nvidia’s Hopper, Blackwell, and Vera Rubin platforms are current standards for AI training and inference.
Inference token cost. The cost per “token” — roughly one word fragment — generated by a deployed AI model. The primary economic lever determining AI deployment at scale.
Hyperscaler. An operator of very large-scale data centre infrastructure: primarily AWS, Google Cloud, Microsoft Azure, and Oracle Cloud. Hyperscaler capital expenditure is the primary demand driver for Nvidia’s Data Center segment.
TSMC. Taiwan Semiconductor Manufacturing Company, the world’s largest contract chip manufacturer. Until 2025, virtually all Nvidia GPU wafers were produced exclusively at TSMC Taiwan facilities.
Source Credibility
Source
Grade
Note
NVDA Form 8-K, Q4/FY2026 (Feb 25, 2026)
Tier 1 — A+
SEC EDGAR mandatory disclosure; fetch-verified this session
NVDA Form 10-K, FY2026 (Feb 25, 2026)
Tier 1 — A+
SEC EDGAR annual statutory filing; fetch-verified this session
NVIDIA Newsroom (OpenAI PR, Sep 22, 2025)
Primary Disclosure — A
Official press release; fetch-verified this session
NVIDIA Blog (US Manufacturing, Apr 14, 2025)
Primary Disclosure — A
Official corporate channel; fetch-verified this session
IEA, Electricity Mid-Year Update 2025 (Jul 2025)
Tier 1 — A+
Intergovernmental; user-supplied PDF; full text verified
Bibliography
Chicago Author-Date, 17th edition. Alphabetical by institution.
International Energy Agency (IEA). 2025. Electricity Mid-Year Update 2025. Paris: IEA, July 2025. Gas, Coal and Power Markets Division. CC BY 4.0. www.iea.org.
NVIDIA Corporation. 2025a. “OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems.” NVIDIA Newsroom, September 22, 2025. https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems.
NVIDIA Corporation. 2025b. “NVIDIA to Manufacture American-Made AI Supercomputers in US for First Time.” NVIDIA Blog, April 14, 2025. https://blogs.nvidia.com/blog/nvidia-manufacture-american-made-ai-supercomputers-us/.
NVIDIA Corporation. 2026a. “NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026.” Form 8-K, Exhibit 99.1. Filed February 25, 2026. SEC EDGAR. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000019/q4fy26pr.htm.
NVIDIA Corporation. 2026b. Annual Report on Form 10-K for the Fiscal Year Ended January 25, 2026. Filed February 25, 2026. SEC EDGAR. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/nvda-20260125.htm.
U.S. Securities and Exchange Commission. n.d. EDGAR: Electronic Data Gathering, Analysis, and Retrieval System. Washington, D.C.: U.S. SEC. https://www.sec.gov/cgi-bin/browse-edgar.
Disclaimer. This briefing is an informational analysis based on publicly available SEC filings and does not constitute investment advice. All figures are sourced from the referenced disclosures.
Sources: SEC EDGAR (NVDA) · NVIDIA Newsroom · IEA · Published April 2026