Data Center revenue grew 125% to a record $6.7B, driven by the Mellanox acquisition (contributing ~10% of total company revenue) and the ramp of Ampere A100 GPUs. Hyperscale and cloud customers adopted A100 for AI training and inference, while vertical industries grew to represent over 50% of Data Center revenue. Management highlighted the A100 as NVIDIA's first 'universal' GPU, capable of training, inference, HPC, and data analytics.

Data Center revenue grew 58% to $10.6B. Hyperscale and cloud demand was described as 'outstanding,' with revenue more than doubling YoY. Inference-focused revenue tripled YoY. Growth was driven by Ampere architecture GPUs for AI workloads like natural language processing and deep learning recommendation systems. Networking grew but was gated by supply constraints.

Data Center revenue grew 41% to $15.0B despite a challenging year overall. Hyperscale demand was strong, but some cloud providers paused at year-end to recalibrate build plans. China revenue declined following the initial U.S. export controls in August 2022. The H100 Hopper GPU began ramping in Q3, with revenue already exceeding A100 by Q4. ChatGPT's November 2022 launch created an inflection point, and management noted that activity around AI infrastructure 'has just gone through the roof in the last 60, 90 days' by the FY23Q4 call.

Data Center revenue tripled to $47.5B, driven by the Hopper H100 platform for training and inference of large language models. Compute revenue grew 244% and networking revenue grew 133%. Large cloud providers represented over half of Q4 Data Center revenue. NVIDIA estimated approximately 40% of Data Center revenue was for AI inference. Enterprise AI adoption accelerated across healthcare, financial services, and automotive verticals, with the automotive data center vertical exceeding $1B.

Data Center revenue more than doubled to $115.2B. Hopper H100/H200 drove the majority of growth, with Blackwell contributing $11B in Q4 alone in its first quarter of shipments. Inference demand accelerated with the emergence of reasoning models like OpenAI O3 and DeepSeek R1. Cloud service providers represented about 50% of Data Center revenue in Q4. Consumer internet revenue grew 3x YoY. Enterprise revenue nearly doubled YoY. Sovereign AI emerged as a new demand driver.

Data Center revenue reached $193.7B, up 68% YoY, with Blackwell becoming the dominant platform. Networking revenue grew 142%, driven by NVLink compute fabric for GB200/GB300 systems. In Q3, Data Center compute was $43.0B (up 56% YoY) and networking was $8.2B (up 162% YoY). Hyperscalers continued to represent slightly over 50% of Data Center revenue. Management stated visibility to $500B of Blackwell and Rubin revenue through end of calendar 2026. Inference demand continued to accelerate with reasoning and agentic AI. Hopper continued to contribute approximately $2B per quarter even in its thirteenth quarter.

NVIDIA launched the Ampere architecture with the A100 GPU, its first 'universal' GPU for training, inference, and HPC. The A100 introduced TF32 for training and multi-instance GPU capabilities. Adoption began with hyperscalers and supercomputing centers. The GeForce RTX 30 Series brought Ampere to gaming.

Ampere was in full production across Data Center and Gaming. The A100 drove strong growth in cloud AI workloads. NVIDIA announced it achieved 20x performance gains over 3 years through full-stack innovation. The company began discussing the upcoming Hopper architecture. Supply constraints persisted across the product line.

The H100 Hopper GPU began ramping in Q3 FY23, with its transformer engine arriving 'just in time' for the generative AI wave. In Q4, H100 revenue already exceeded A100 revenue, reflecting customer urgency to adopt the new architecture. The H100 delivered up to 9x faster training and 30x faster inferencing on transformer models versus A100.

Hopper was the primary revenue driver, with the H100 and HGX systems powering the generative AI buildout. NVIDIA began shipping H200 with nearly double the inference performance of H100. The company announced Blackwell as the next-generation architecture and disclosed plans for an annual product cadence going forward. Non-GAAP gross margin reached 76.7% in Q4 FY24 as Hopper scaled.

Hopper H100/H200 drove most of the year's revenue, while Blackwell production began in Q4 with $11B of revenue in its first shipping quarter. Blackwell introduced the GB200 NVLink 72 rack-scale system, a fundamental architectural change connecting 72 GPUs and 36 CPUs in a single rack. Non-GAAP gross margins declined to 73.5% in Q4 as Blackwell ramped, reflecting higher manufacturing complexity. Management guided for low-70s margins during the ramp, improving to mid-70s later.

Blackwell became the dominant platform, contributing nearly 70% of Data Center compute revenue by Q1. GB300 (Blackwell Ultra) began sampling in Q1 and entered production by Q2, delivering 50% more FP4 inference compute than GB200. The transition was seamless due to shared architecture and physical footprint. Non-GAAP gross margins recovered from 61.0% in Q1 (depressed by the $4.5B H20 charge) to 75.2% in Q4 as cost structure and mix improved. Rubin, the next-generation platform with six new chips, entered fab and is on track for volume production in the second half of FY27.

In August 2022, the U.S. government imposed license requirements for exports of A100 and H100 GPUs to China and Russia. NVIDIA began developing alternative products (like the A800) for the China market. In October 2023, additional restrictions expanded the scope to include more products and more countries. Data Center revenue from China declined, and NVIDIA recorded a $702M inventory charge in Q3 FY23 related to lower China demand expectations. China sales as a percentage of Data Center revenue began declining.

China represented 14% of Data Center revenue for FY24, down from 19% in FY23. Following October 2023 export controls, China fell to a mid-single-digit percentage in Q4 FY24. NVIDIA began shipping alternative products not subject to license requirements but noted its competitive position in China was harmed. The company started shipping H20, a product designed specifically for the China market under the new restrictions.

Data Center revenue in China grew in absolute terms as H20 shipped, but remained well below pre-export-control levels as a percentage of total Data Center revenue. In January 2025, the outgoing U.S. administration published an 'AI Diffusion' rule imposing worldwide licensing requirements on a broad range of NVIDIA products. NVIDIA began planning for further restrictions.

In April 2025, the U.S. government required licenses for H20 exports to China, effectively ending NVIDIA's Hopper-based Data Center compute business in China. This triggered a $4.5B charge in Q1 FY26 for excess H20 inventory and purchase obligations. Prior to the restriction, H20 revenue was $4.6B in Q1 and had been running at $7-8B per quarter. NVIDIA could not ship $2.5B of H20 orders in Q1. For Q2, the company estimated $8B of lost H20 revenue. China Data Center revenue declined to low-single-digit percentages. Revenue from customers headquartered outside the U.S. fell to 31% of total revenue in FY26 from 41% in FY25. Management stated that Hopper cannot be further reduced for productive use under the new limits, and NVIDIA has no replacement product at present. The AI Diffusion rule was rescinded by President Trump, reopening non-China international markets.

Gaming revenue grew 41% to $7.8B, driven by the launch of GeForce RTX 30 Series (Ampere). Demand was described as 'incredible,' with channel inventories ending lower than they started. Only ~15% of the installed base had upgraded to RTX. Cryptocurrency mining contributed an estimated $100-300M to Q4 gaming revenue, though visibility was limited.

Gaming revenue grew 61% to $12.5B, led by continued Ampere RTX 30 Series demand fueled by gaming growth, hybrid work, and creator adoption. NVIDIA introduced Lite Hash Rate GPUs to limit Ethereum mining capability and launched dedicated Cryptocurrency Mining Processors (CMP), which generated $550M. Supply constraints persisted. Steam concurrent users grew 50% over two years.

Gaming revenue declined 27% to $9.07B as macroeconomic weakness and COVID disruptions in China created excess channel inventory. NVIDIA implemented pricing programs and deliberately undershipped GPUs to clear inventory. The Ada Lovelace architecture (RTX 40 Series) launched during the year. CMP revenue fell to nominal levels. The inventory correction contributed to $2.17B in total inventory provisions that compressed gross margins.

Gaming revenue recovered 15% to $10.4B as channel inventory normalized and the RTX 40 Series ramped. Higher sell-in to partners following the inventory correction drove growth. At CES, NVIDIA launched the GeForce RTX 40 SUPER Series. Management highlighted that over 100M RTX PCs were in the installed base and over 500 AI-enabled applications were available.

Gaming revenue grew 9% to $11.4B for the full year. Q4 revenue of $2.5B was down 11% YoY due to supply constraints affecting both Blackwell and Ada GPUs. The RTX 50 Series (Blackwell) was announced at CES with DLSS 4 and 2x performance improvements. Supply was expected to improve in Q1 FY26.

Gaming revenue grew 41% to $16.0B for FY26, driven by the Blackwell RTX 50 Series ramp. Record revenue of $4.3B was achieved in Q2 and Q3 as supply availability improved. The RTX 5060 brought Blackwell to mainstream gamers at $299. Management expects supply constraints for Gaming starting in Q1 FY27 and beyond. The Nintendo Switch 2 was announced using NVIDIA's custom RTX GPU with DLSS technology.

The Mellanox acquisition closed in April 2020 and contributed ~10% of NVIDIA's total revenue in FY21. Mellanox added InfiniBand, Ethernet, and ConnectX adapter products. Integration proceeded well, with shared SerDes technology and accelerated BlueField DPU roadmaps. Networking was primarily driven by HPC and cloud deployments.

Networking grew strongly but was supply-constrained. The Quantum InfiniBand and Spectrum switch businesses expanded. Revenue benefited from the first full year of Mellanox integration. Non-recurring Mellanox acquisition charges (including $161M inventory step-up) rolled off, benefiting gross margins.

InfiniBand led networking growth as the Quantum-2 400 Gb/s platform ramped, driven by AI and supercomputing demand. Generative AI model sizes growing at exponential rates drove the need for high-performance scale-out networking. Spectrum-4 Ethernet gained momentum. Data Center networking revenue grew 133% in FY24 (partly reflecting the FY23 ramp).

Networking exceeded a $13B annualized revenue run rate. Quantum InfiniBand revenue grew more than 5x YoY. NVIDIA launched Spectrum-X, a purpose-built Ethernet solution for AI, delivering 1.6x higher networking performance for AI versus traditional Ethernet. GPU networking attach rates were over 75%.

Networking revenue grew 51% for the full year, driven by Spectrum-X Ethernet for AI. Q4 networking declined 3% sequentially as the business transitioned from NVLink 8 (Hopper) to NVLink 72 (Blackwell). Spectrum-X was adopted by Microsoft Azure, OCI, and others for large AI factories. Cisco announced integrating Spectrum-X into its portfolio.

Data Center networking revenue grew 142% for FY26, reaching a record $11.0B in Q4 (up 263% YoY). NVLink compute fabric for GB200/GB300 systems was the primary growth driver. Spectrum-X Ethernet exceeded a $10B annualized revenue rate. InfiniBand revenue nearly doubled sequentially in Q2. NVIDIA introduced Spectrum XGS for connecting multiple data centers into gigascale AI super factories. NVLink Fusion was announced, enabling third-party CPUs and ASICs to connect to NVIDIA's platform via NVLink.

Non-GAAP gross margin was 62.3%, up 30 bps YoY. Improvement was driven by favorable product mix (higher Data Center contribution), partially offset by $424M in Mellanox acquisition-related charges including a $161M inventory step-up. Operating expenses grew 50% to $5.86B, driven primarily by the Mellanox acquisition.

Non-GAAP gross margin expanded to 64.9%, benefiting from lower Mellanox acquisition-related charges and a favorable high-end product mix in Gaming. Operating expenses grew 27% to $7.4B, well below the 61% revenue growth rate, driving significant operating leverage. Operating income grew 122% to $10.0B.

Non-GAAP gross margin contracted to 56.9% from 64.9%, primarily due to $2.17B in inventory provisions (7.5% impact on gross margin) for excess Gaming and Data Center (China-related) products. Operating expenses grew 50% to $11.1B, including the $1.35B Arm acquisition termination charge. Net income declined 55% to $4.37B.

Non-GAAP gross margin expanded to 72.7% from 56.9%, driven by the mix shift to Data Center revenue and lower inventory provisions (2.7% impact vs. 7.5% prior year). Operating expenses grew only 2% to $11.3B as the prior year's $1.4B Arm termination charge did not recur. R&D grew 18%. Operating income grew 681% to $33.1B, and net income grew 581% to $29.8B.

Non-GAAP gross margin expanded to 75.0%, benefiting from an even higher Data Center revenue mix. Operating expenses grew 45% to $16.4B, but revenue grew 114%, generating strong operating leverage. Operating income grew 147% to $81.5B. R&D grew 49% to $12.9B, driven by employee growth, compute infrastructure costs, and engineering development for new architectures.

Non-GAAP gross margin declined to 71.1% for FY26, pressured by the $4.5B H20 inventory charge in Q1 (which alone reduced Q1 non-GAAP margin to 61.0%) and the transition to more complex Blackwell rack-scale systems. Excluding the H20 charge, margins improved sequentially through the year, reaching 75.2% in Q4. R&D grew 43% to $18.5B. Total operating expenses grew 41% to $23.1B, slower than the 65% revenue growth. Net income grew 65% to $120.1B. Management targeted mid-70s gross margins going forward, noting input cost pressures but offsetting improvements in cost structure, cycle time, and mix.

NVIDIA paid $395M in dividends and did not repurchase stock. Cash was preserved for the pending $40B Arm acquisition announced in September 2020. NVIDIA issued $5.0B in new notes in March 2020. Capital expenditures guidance was $1.0-1.2B for FY22.

NVIDIA paid $399M in dividends and did not repurchase stock, again preserving cash for the Arm deal. The company issued $5.0B of notes in June 2021. Inventory purchase obligations grew to $6.9B from $2.57B a year earlier as NVIDIA secured supply chain capacity amid industry-wide shortages.

The Arm acquisition was terminated in February 2022 due to regulatory challenges, resulting in a $1.36B charge. With the deal off, NVIDIA returned $10.44B to shareholders ($10.04B buybacks, $398M dividends). Cash from operations declined to $5.64B from $9.11B due to lower net income and higher tax payments. The board authorized an additional $15B in buybacks.

Cash from operations increased to $28.1B from $5.6B. NVIDIA returned $9.9B to shareholders ($9.5B buybacks, $395M dividends). Capital expenditures guidance was $3.5-4.0B for FY25. The company completed a 10-for-1 stock split in June 2024.

NVIDIA returned $34.8B to shareholders ($34.0B buybacks, $834M dividends). The board approved a $50B additional buyback authorization. Capital expenditures increased to $3.4B. The company began making strategic equity investments in AI model companies to deepen ecosystem partnerships.

NVIDIA returned $41.4B to shareholders ($40.4B buybacks, $974M dividends). The board approved an additional $60B buyback authorization. Ecosystem investments totaled $17.5B in private companies and infrastructure funds, primarily to support AI startups and model makers. The company also provided $3.5B in land, power, and shell guarantees to early-stage companies. Capital expenditures increased to $6.1B. Management announced partnerships with OpenAI (up to 10 GW of AI data centers) and Anthropic (up to 1 GW of compute capacity with Grace Blackwell and Rubin systems). These investments were framed as expanding CUDA's reach rather than financial speculation, with management noting that NVIDIA's architecture is the only platform that runs every major AI model.