NVIDIA

From PRIMO.ai
Jump to navigation Jump to search

Youtube search... ...Google search


NVIDIA is the company that turned the graphics processing unit (GPU) into the engine of modern artificial intelligence. Founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem, it began by making 3D graphics chips for PC games. In 2006 it released CUDA, a programming platform that let scientists and engineers run general-purpose math on GPUs. When deep learning took off around 2012, CUDA-powered GPUs became the default hardware for training neural networks, and NVIDIA's position in AI grew from there.

The problem NVIDIA solves is simple to state: modern AI requires enormous amounts of arithmetic, mostly matrix multiplication, performed on enormous amounts of data. A conventional CPU runs a small number of powerful cores optimized for sequential tasks. A GPU runs thousands of simpler cores in parallel, plus specialized Tensor Cores built for the exact matrix operations neural networks need. NVIDIA's advantage, however, has never been the chip alone. It comes from the full stack around it:

  • Silicon: GPUs (Hopper, Blackwell, Rubin), CPUs (Grace, Vera), networking chips (NVLink switches, ConnectX SuperNICs, BlueField DPUs, Spectrum-X Ethernet, InfiniBand), and, since its 2025 deal with Groq, low-latency inference processors (LPUs).
  • Systems: rack-scale computers such as the GB300 NVL72 and Vera Rubin NVL72, sold through cloud providers and server makers, plus DGX systems for enterprises and DGX Spark and RTX Spark for desktops and laptops.
  • Software: CUDA and hundreds of CUDA-X libraries, RAPIDS for data science, NGC containers, NIM inference microservices, the Dynamo inference server, and open models such as Nemotron, Cosmos, Isaac GR00T, and Alpamayo.
  • Blueprints for entire data centers: the DSX reference designs and Omniverse digital twins that describe how to build what Huang calls "AI factories."

The paradigm shift NVIDIA has pushed since about 2023 is that a data center is no longer a building full of general-purpose servers. It is a single, purpose-built machine that turns electricity into "tokens," the units of text, code, images, and actions that AI models produce. In this framing, the unit of competition is no longer a chip. It is the cost and speed of producing useful tokens per watt and per dollar across an entire facility.

By October 2026 that strategy had made NVIDIA one of the most valuable companies in history. It became the first company to pass $4 trillion in market value in July 2025 and the first to pass $5 trillion on October 29, 2025. Fiscal 2026 revenue (the year ended January 2026) reached $215.9 billion, and quarterly revenue reached $96.2 billion in the quarter ended July 2026. The company has also become a major investor and dealmaker across the AI industry, which brings new questions about competition, concentration, and the sustainability of AI spending.


NVIDIA's real product is no longer a chip but an entire AI factory, co-designed from transistors to software. Its lead depends on keeping that whole stack a generation ahead, every year, while the industry tests whether the tokens it produces are worth what the buildout costs.



Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube
Load video
YouTube

Company Snapshot (October 2026)

Item Detail
Founded April 1993, by Jensen Huang, Chris Malachowsky, and Curtis Priem
Headquarters Santa Clara, California; a second major campus, "Constellation," is under construction in Taipei's Beitou-Shilin Technology Park
CEO Jensen Huang (founder; recipient of the 2026 IEEE Medal of Honor)
Fiscal year Ends in late January; fiscal 2027 runs February 2026 to January 2027
Fiscal 2026 revenue $215.9 billion, up 65 percent; Data Center $193.7 billion
Latest reported quarter Q2 fiscal 2027 (ended July 26, 2026): $96.2 billion revenue, up 106 percent year over year
Guidance About $108 billion for Q3 fiscal 2027, excluding any data center compute revenue from China
Reporting platforms (since Q1 fiscal 2027) Data Center (Hyperscale; AI Clouds, Industrial and Enterprise, or "ACIE") and Edge Computing
Current data center platform Vera Rubin (in full production; ramping through 2026); Blackwell Ultra (GB300) still shipping in volume
Pending acquisition Hugging Face, $12.93 billion, expected to close in the first half of 2027 subject to regulatory approval
Next major events GTC San Jose, March 15 to 18, 2027

Mechanisms

NVIDIA's business works as a flywheel with four interlocking parts.

1. An annual, full-stack hardware cadence. At Computex 2024 NVIDIA moved from a two-year to a one-year rhythm for data center architectures. Each generation is designed as a system, not a part:

  • Hopper (2022 to 2024): H100 and H200, the chips that powered the first wave of large language models.
  • Blackwell (2025): B200 and the GB200 NVL72 rack, which joins 72 GPUs into one NVLink domain that software can treat as a single giant GPU.
  • Blackwell Ultra (late 2025 to 2026): GB300, with more memory and stronger low-precision inference.
  • Vera Rubin (2026): a platform of seven chips (Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch with co-packaged optics, and the Groq 3 LPU), five rack-scale systems, and a pod-scale supercomputer aimed at agentic AI.
  • Rubin Ultra (2027) in the new vertical "Kyber" rack, and Feynman (2028) with a new Rosa CPU, LP40 LPU, BlueField-5, and optical scale-up.

2. Software lock-in through usefulness. CUDA turned 20 in 2026. Two decades of libraries (cuDNN, cuBLAS, NCCL, TensorRT, cuDSS, and many others), tooling, and developer habit mean that most AI frameworks and models run best, or first, on NVIDIA hardware. NVIDIA now extends that layer upward with open models (Nemotron for language, Cosmos for world models, GR00T for humanoid robots, Alpamayo for self-driving), NIM microservices, the Dynamo inference server, and agent tooling such as the Agent Toolkit, NemoClaw, and the OpenShell secure runtime. Every new model that ships optimized for NVIDIA reinforces the installed base.

3. An ecosystem that manufactures and sells the systems. NVIDIA designs; TSMC fabricates and packages (CoWoS advanced packaging); SK Hynix, Samsung, and Micron supply high-bandwidth memory (HBM); Taiwanese and US manufacturers (Foxconn, Quanta, Wistron, and others) build racks; and cloud providers, "neoclouds" such as CoreWeave and Nebius, server makers, and governments deploy them. Huang said at Computex 2026 that each Vera Rubin system contains nearly 2 million parts from about 150 partners in Taiwan, and that NVIDIA expects to spend roughly $150 billion a year there. NVIDIA increasingly contributes rack designs to the Open Compute Project and publishes DSX reference designs so partners can build complete AI factories faster.

4. Demand creation and capital deployment. NVIDIA invests heavily across its customer and supplier base: equity stakes in OpenAI ($30 billion), Anthropic (up to $10 billion), Intel ($5 billion), and Nokia ($1 billion); dozens of startup investments; capacity reservations with suppliers; and acquisitions such as the pending Hugging Face deal. These investments expand the market for NVIDIA systems, but they also raise the "circularity" question discussed below.

How inference changed the mechanism. Until about 2024, demand was driven mostly by training ever-larger models. Since then, demand has increasingly come from inference: serving models to users and, more recently, running AI agents that reason in long chains, call tools, read files, and work for minutes or hours. NVIDIA responded in three ways:

  • Disaggregated serving. Inference has two phases. Prefill processes the prompt and is compute-heavy. Decode generates output tokens one at a time and is limited mainly by memory bandwidth. NVIDIA's Dynamo software routes these phases to different GPUs and manages the key-value (KV) cache that stores context.
  • A dedicated low-latency engine. Through its December 2025 licensing and hiring deal with Groq, NVIDIA added the LPU, a processor that keeps model weights in on-chip SRAM for very fast decode. The Groq 3 LPX rack (256 LPUs) is designed to sit beside Vera Rubin racks, with GPUs handling prefill and attention and LPUs handling latency-critical decode.
  • CPUs and storage for agents. Agents spend much of their time on non-GPU work: running code, querying databases, retrieving documents. The Vera CPU, launched as a standalone product at Computex 2026, and the BlueField-4 STX storage architecture target that work.

Where uncertainty remains. The mechanism depends on conditions NVIDIA does not control: TSMC capacity and Taiwan's geopolitical stability, HBM supply, electrical power for gigawatt-scale sites, customers' ability to fund multi-hundred-billion-dollar capital budgets, export rules, and whether AI services generate enough revenue to justify the hardware. Performance claims for unreleased products (Rubin Ultra, Feynman) are company targets, not measured results.

Why This Matters

For the AI industry. Almost every frontier lab, cloud provider, and AI startup depends substantially on NVIDIA hardware and software. NVIDIA's roadmap shapes what models can be trained and how cheaply they can be served. When NVIDIA says Vera Rubin reduces inference cost several-fold, that is effectively a forecast for the price of AI services.

For the economy and markets. NVIDIA's revenue roughly doubled year over year in mid-2026, and its market value has exceeded $5 trillion. Its results are watched as a gauge of AI capital spending by Microsoft, Google, Amazon, Meta, Oracle, OpenAI, and others. Data center revenue in the quarter ended July 2026 was about 93 percent of total revenue.

For science and engineering. NVIDIA systems power leading supercomputers, including the US Department of Energy's planned Solstice system at Argonne National Laboratory (100,000 Blackwell GPUs, announced October 2025) and Europe's first exascale system, JUPITER, where NVIDIA reported a record simulation of fault-tolerant quantum computing in October 2026. GPU-accelerated libraries such as cuDSS (generally available October 2026) move engineering simulation onto GPUs, and NVIDIA committed $1 billion over five years to US science in October 2026.

For nations ("sovereign AI"). Governments increasingly treat AI compute as strategic infrastructure. NVIDIA reported more than $30 billion of sovereign AI revenue in fiscal 2026, more than triple the prior year, with large programs in South Korea, the United Kingdom, Germany, France, Saudi Arabia, Japan, India, and elsewhere.

For computing beyond data centers. NVIDIA is pushing into PCs (RTX Spark), robotics (Jetson Thor and Isaac GR00T), autonomous vehicles (DRIVE Hyperion and Alpamayo), telecom (AI-RAN with Nokia), and even orbital data centers (Vera Rubin Space-1). Physical AI contributed more than $6 billion of revenue in fiscal 2026, by NVIDIA's account.

Competition and limits. NVIDIA's position is strong but contested:

  • AMD signed a deal with OpenAI in October 2025 for up to 6 GW of Instinct GPUs.
  • Custom chips are growing: Google's TPUs (including a 2025 agreement for Anthropic to use up to one million TPUs), Amazon's Trainium, and Broadcom-designed accelerators (including a 10 GW OpenAI program announced in 2025).
  • Huawei and other Chinese suppliers are filling demand that US export rules and Chinese policy keep from NVIDIA.
  • Inference specialists and new architectures challenge the GPU's efficiency at serving models, which partly explains the Groq deal.

If NVIDIA's thesis is right (that AI agents and physical AI create sustained, compounding demand for tokens), then AI factories become as fundamental as power plants, and the companies that build the most efficient ones capture enormous value. If it is wrong (if AI revenue lags spending, or cheaper alternatives catch up), the industry could face overbuilt capacity, and NVIDIA's investments in its own customers would amplify the downturn.

What Remains Uncertain

  • Durability of demand. NVIDIA guided to roughly 70 percent revenue growth for fiscal 2028 and described it as supply-constrained. Rising interest rates (US 10-year yields above 5 percent in October 2026) and questions about AI return on investment have produced sharp stock swings even as results beat expectations.
  • Circular financing. Analysts have questioned arrangements in which NVIDIA invests in companies that then spend heavily on NVIDIA systems. Huang said in March 2026 that NVIDIA's OpenAI and Anthropic stakes would likely be its last in those companies ahead of their expected public listings.
  • Execution. Vera Rubin is the most complex system NVIDIA has built. Its ramp, and the 2027 to 2028 roadmap, must hold to justify current expectations.
  • Geopolitics. Manufacturing is concentrated in Taiwan; China's market is largely closed by a mix of US and Chinese policy; and antitrust regulators in China, the US, and Europe scrutinize NVIDIA's conduct and deals.
  • Power and grids. Gigawatt-scale AI factories need new generation and transmission, which move slower than chip roadmaps.

Recent Developments

This section covers the most important developments from roughly late 2024 through October 21, 2026, grouped by theme. Company performance claims are identified as such.

Recent Developments

Record growth and a new reporting structure

  • Fiscal 2026 (ended January 25, 2026): revenue $215.9 billion (up 65 percent). Data Center revenue was $193.7 billion, networking exceeded $31 billion for the year, and Q4 alone brought $68.1 billion. NVIDIA returned $41.1 billion to shareholders and said nearly 9 GW of Blackwell-based infrastructure had been deployed within a year of launch.
  • Q1 fiscal 2027 (ended April 26, 2026): revenue $81.6 billion (up 85 percent); Data Center $75.2 billion; networking a record $14.8 billion (up 199 percent). NVIDIA added $80 billion to its buyback authorization, raised its quarterly dividend from $0.01 to $0.25 per share, reorganized reporting into Data Center (Hyperscale and ACIE) and Edge Computing, and began including stock-based compensation in non-GAAP results.
  • Q2 fiscal 2027 (ended July 26, 2026): revenue $96.2 billion (up 106 percent); Data Center $89.0 billion (up 117 percent), of which Hyperscale was about $48.7 billion and ACIE about $40.3 billion; Edge Computing $7.2 billion; gross margin 75.0 percent. Huang said "compute is revenue" and that Vera Rubin was in full production. The Q3 outlook of about $108 billion assumes no data center compute revenue from China.
  • Market value: first company to $4 trillion (July 2025) and $5 trillion (October 29, 2025). After the August 2026 report, the stock added more than $400 billion in a day, taking market value near $5.5 trillion; by mid-October 2026 it traded roughly 10 percent below its 52-week high amid a broader sell-off in AI chip stocks tied to rising bond yields.

From Blackwell to Vera Rubin

  • GTC 2025 (March 2025) introduced Blackwell Ultra, the Dynamo inference server, the Vera Rubin roadmap, Isaac GR00T N1 for humanoid robots, and Spectrum-X and Quantum-X silicon photonics switches.
  • Computex 2025 introduced NVLink Fusion, which lets other companies' CPUs and custom accelerators plug into NVIDIA's NVLink fabric.
  • GTC Washington, D.C. (October 2025) brought announcements with Uber, Nokia, Palantir, and the Department of Energy, plus the NVQLink interconnect for linking quantum processors to GPU supercomputers. Huang said Blackwell and Rubin bookings through 2026 totaled about $500 billion, and the first Blackwell wafer was produced at TSMC's Arizona fab.
  • CES 2026 (January 5, 2026) declared Vera Rubin in "full production," described a six-chip platform (Vera CPU with 88 custom Arm cores, Rubin GPU with 336 billion transistors, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6), and claimed about 5 times Blackwell's inference performance and up to 10 times lower cost per token. Some analysts read "full production" as a signal to investors that Rubin was on schedule rather than evidence of volume shipments.
  • GTC 2026 (March 16 to 19, San Jose) formally launched Vera Rubin as seven chips, five rack-scale systems, and one supercomputer; added the Groq 3 LPX rack; previewed the 600 kW Kyber rack for Rubin Ultra (2027) and the Feynman generation (2028); announced the DSX AI Factory reference design and Omniverse DSX Blueprint; introduced Vera Rubin Space-1 for orbital data centers; launched DLSS 5; and unveiled NemoClaw for enterprise agents. Huang said he expected about $1 trillion in combined Blackwell and Vera Rubin purchase orders through 2027. Microsoft Azure brought up the first Vera Rubin rack.
  • GTC Taipei and Computex (June 1, 2026) reiterated full production, launched the Vera CPU as a standalone data center processor, introduced RTX Spark for Windows PCs with Microsoft and MediaTek, and announced the Constellation campus in Taipei.
  • Fall 2026: NVIDIA contributed Vera Rubin rack and compute-tray designs to the Open Compute Project (October 12), Google Cloud announced A5X instances on Vera Rubin with general availability expected in November 2026 (October 21), and management said Vera Rubin racks were running at hyperscalers and neoclouds.

The inference turn: Groq and disaggregated serving

On December 24, 2025, Groq announced a non-exclusive licensing agreement with NVIDIA for its inference technology, and Groq founder Jonathan Ross (a co-inventor of Google's TPU), president Sunny Madra, and other staff joined NVIDIA. CNBC reported the value at about $20 billion, which would make it NVIDIA's largest transaction; NVIDIA said it was not acquiring the company, and GroqCloud continued independently. Analysts noted the structure likely reduced antitrust exposure compared with a full merger. By GTC 2026 the technology had become the Groq 3 LPX rack (shipping in the third quarter of 2026), and the Feynman generation will include an LP40 LPU co-designed with the Groq team. What this establishes: NVIDIA concluded that GPUs alone were not the optimal answer for latency-critical decode. What it does not establish: how much of future inference will run on LPUs versus GPUs, which will depend on measured cost per token in production.

Agentic AI software

NVIDIA has positioned Vera Rubin as built "to run agents," not just models. Its software response includes:

  • Nemotron open language and reasoning models, expanded in 2026 with models for European languages, biology, and quantum computing.
  • NemoClaw (GTC 2026), an enterprise platform for autonomous agents.
  • OpenShell, an open-source, policy-controlled runtime with kernel-level sandboxing between an agent and the systems it can affect, and the OpenShell Reference Architecture for Safe AI Agents (GTC Berlin, October 21, 2026).
  • Agent Toolkit, expanded in October 2026 to ground enterprise agents in company data, and confidential computing features so frontier model builders can run on shared infrastructure securely.

These are early-stage products; their adoption relative to agent frameworks from model labs and cloud providers remains to be seen.

NVIDIA returns to the PC: RTX Spark and Intel

  • RTX Spark (announced June 1, 2026; laptops on sale October 16, 2026 from $2,399) is an Arm-based system-on-chip developed with MediaTek and derived from the GB10 superchip in DGX Spark. It combines a 20-core CPU (10 Cortex-X925 and 10 Cortex-A725 cores) with a Blackwell RTX GPU of 6,144 CUDA cores, up to 128 GB of unified memory, and up to 1 petaflop of FP4 AI performance, built on a TSMC 3 nm process. Launch systems come from ASUS, Dell, HP, Lenovo, Microsoft (Surface Laptop Ultra), and MSI. NVIDIA and Microsoft pitch it as a platform for running personal AI agents locally, with new Windows security primitives and OpenShell controlling what agents can access. Independent reviews will determine how it compares with Apple, Qualcomm, AMD, and Intel laptops.
  • DGX Spark (shipped October 2025) remains NVIDIA's Linux desktop for AI developers, with DGX Station for larger local workloads.
  • Intel partnership (September 18, 2025): NVIDIA bought $5 billion of Intel stock at $23.28 per share (since cleared by regulators). Intel is building custom x86 CPUs that connect to NVIDIA GPUs over NVLink for data centers, and x86 system-on-chips with integrated RTX GPU chiplets for PCs. As of mid-2026, Intel's CEO confirmed the work was ongoing; consumer products had not yet launched.

Physical AI: robotics and autonomous vehicles

  • Robotics: Jetson Thor (August 2025) brought Blackwell-class compute to humanoid and industrial robots. Isaac GR00T foundation models, Cosmos world models, and Omniverse simulation form a "train in simulation, deploy in the real world" pipeline. On October 21, 2026, NVIDIA expanded Isaac with a reference humanoid (the Agile ONE SN from Agile Robots, built on GR00T and Jetson Thor with planned Halos safety integration), new datasets, and runtime updates.
  • Autonomous driving: At CES 2026 NVIDIA released Alpamayo, open vision-language-action models that reason through rare driving situations and explain their decisions (Alpamayo 1 has about 10 billion parameters and is intended for development, not direct deployment). Mercedes-Benz's CLA is the first production car to ship with NVIDIA's full DRIVE AV stack.
  • Robotaxis: Uber and NVIDIA announced in October 2025 a plan to scale toward 100,000 Level 4 vehicles starting in 2027 on the DRIVE AGX Hyperion 10 platform, with partners including Stellantis, Lucid, and Mercedes-Benz, plus a joint data factory built on Cosmos.
  • October 2026: Volvo Cars (October 15) and Mercedes-Benz (October 16) expanded their partnerships, and at GTC Berlin Bolt with Lucid, Mercedes-Benz, Volkswagen Commercial Vehicles, and Einride showed autonomous-mobility use cases on Hyperion.

Europe and sovereign AI

  • 2025: At GTC Paris and London Tech Week (June 2025), NVIDIA announced European AI infrastructure partnerships, including an industrial AI cloud in Germany with Deutsche Telekom (10,000 GPUs) and large commitments to the United Kingdom (120,000 Blackwell GPUs) and Germany (100,000), according to NVIDIA's regional statements.
  • South Korea (October 31, 2025): more than 260,000 Blackwell GPUs for the government (about 50,000), Samsung, SK Group, and Hyundai Motor Group (about 50,000 each), and Naver Cloud (about 60,000), plus AI factories, sovereign model projects, and 6G work.
  • GTC Berlin (October 20 to 22, 2026): Huang met German Chancellor Friedrich Merz on October 19 to discuss industrial AI. In the October 21 keynote he described a five-layer AI economy (energy, computing, AI factories, models, applications) and urged Europe to produce AI, not just consume it. Announcements included DSX POD (containerized, factory-tested Vera Rubin AI factories built by partners such as Bull, Dell, HPE, Lenovo, Schneider Electric, and Vertiv, available in early 2027), DSX designs with closed-loop liquid cooling, European cloud partner expansions, Vera CPU adoption by Nebius and Nscale, industrial digital twins, a protein-design model (RFD4-Proteina) with healthcare partners, and telecom "distributed intelligence" networks with Nokia. Coverage reported that NVIDIA's European commitments total close to $10 billion.

Strategic investments and the circularity debate

  • OpenAI: In September 2025 the companies signed a letter of intent for at least 10 GW of NVIDIA systems, with NVIDIA to invest up to $100 billion progressively. The arrangement never became a binding contract. In February 2026 NVIDIA instead took a $30 billion equity stake in OpenAI's roughly $110 billion funding round (alongside Amazon and SoftBank), not tied to deployment milestones.
  • Anthropic: In November 2025 Microsoft, NVIDIA, and Anthropic announced that Anthropic would commit up to 1 GW of compute on Grace Blackwell and Vera Rubin systems and purchase $30 billion of Azure capacity, with NVIDIA investing up to $10 billion and Microsoft up to $5 billion. It was NVIDIA's first deep technology partnership with Anthropic.
  • Huang's stance: At a Morgan Stanley conference on March 4, 2026, Huang said the $100 billion OpenAI opportunity was "probably not in the cards" and that the OpenAI and Anthropic investments would likely be NVIDIA's last in those companies, citing their expected public listings.
  • Other stakes: Intel ($5 billion), Nokia ($1 billion for about 2.9 percent, October 2025), and many startups through NVentures. Critics argue that investing in customers can inflate demand signals; NVIDIA says its investments aim to expand and deepen its ecosystem. The debate is unresolved and turns on whether these customers generate durable revenue.

Acquisitions

NVIDIA has historically preferred licensing, acqui-hires, and smaller purchases, partly because its $40 billion bid for Arm collapsed under regulatory opposition in 2022. Recent activity:

  • Run:ai (GPU orchestration; completed late 2024 after an EU review), plus a series of 2025 software deals including Lepton AI, Gretel (synthetic data), CentML, and SchedMD (developer of the Slurm workload manager), and an acqui-hire of Enfabrica's networking team.
  • Groq (December 2025): licensing and talent deal, reportedly about $20 billion (see above).
  • Kumo AI (mid-2026): a predictive-AI software company, reportedly for about $400 million.
  • Hugging Face (announced September 3, 2026): a definitive agreement to acquire the leading open model and dataset hub for $12.93 billion, of which about $11.9 billion goes to Hugging Face investors and about $1 billion to an employee retention program. Expected to close in the first half of 2027 subject to regulatory approval. Huang pledged that Hugging Face will remain an open, multi-cloud, multi-accelerator platform supporting every model builder, and NVIDIA said developers will not need its chips to use it. NVIDIA was already the largest contributor of models to the platform. Regulators are expected to examine whether ownership could favor NVIDIA hardware or software; NVIDIA's SEC filing also flagged the risk that governments restrict open-source models.

China and export controls

  • 2025: In April the US required licenses for the China-specific H20, and NVIDIA took a charge of about $4.5 billion. In August the administration allowed H20 sales in exchange for a 15 percent revenue share, but Chinese authorities discouraged purchases and later told companies to stop ordering the RTX Pro 6000D. In September, China's State Administration for Market Regulation (SAMR) announced a preliminary finding that NVIDIA had violated anti-monopoly law by not fully honoring conditions attached to its 2020 Mellanox acquisition; penalties were not specified.
  • December 2025 to 2026: The US allowed H200 exports to approved Chinese customers with a 25 percent cut to the US government, under a licensing framework that capped volumes. Chinese customs initially blocked shipments in January 2026, then regulators reportedly granted in-principle approval for major buyers such as Alibaba, ByteDance, and Tencent. Actual volumes stayed small: NVIDIA said Hopper shipments to China were below 1 percent of Data Center revenue in Q2 fiscal 2027, and its guidance continues to exclude China. On October 20, 2026, Reuters reported that ByteDance, Alibaba, and Tencent had ordered about $35 billion of NVIDIA chips; how much will actually be delivered remains uncertain given both governments' approvals.

Science, quantum, and government

  • Supercomputers: In October 2025 NVIDIA and Oracle announced Solstice (100,000 Blackwell GPUs) and Equinox (10,000) at Argonne, part of seven new US government systems.
  • Quantum: NVQLink (2025) connects quantum processors to GPU supercomputers for real-time error correction, and CUDA-Q supports hybrid programming. In October 2026 NVIDIA ran a record simulation of "magic state cultivation," a leading approach to fault-tolerant quantum computing, on JUPITER.
  • US science: On October 8, 2026, NVIDIA committed $1 billion over five years to US research capacity in quantum computing, healthcare, and energy, framed as support for the administration's Genesis Mission.

Supply chain and manufacturing

In April 2025 NVIDIA said it planned to produce up to $500 billion of AI infrastructure in the US over four years, with partners including TSMC, Foxconn, and Wistron. On October 21, 2026 it announced further capacity agreements, including a $1.1 billion capacity-reservation advance for KYEC to build test facilities in the US and a plan by SPIL to expand co-packaged-optics engine capacity 15 times by mid-2027. In Taiwan, NVIDIA expects spending of roughly $150 billion a year and plans to house about 4,000 staff at its Constellation campus by 2030.

Products and Roadmap

Generation or product Timing Key facts (company specifications or claims unless noted)
Hopper (H100, H200) 2022 to 2024 Workhorse of the first generative AI wave; H200 later approved for export to China under license
Blackwell (B200, GB200 NVL72) 2025 About 208 billion transistors; 72 GPUs per NVLink domain; FP4 inference
Blackwell Ultra (GB300) Late 2025 to 2026 More HBM and faster FP4 inference; main revenue driver through mid-2026
Vera Rubin NVL72 Full production 2026; cloud availability from late 2026 Rubin GPU (about 336 billion transistors, about 50 PF NVFP4, HBM4 at about 22 TB/s); Vera CPU (88 custom Arm cores); NVLink 6 at about 3.6 TB/s per GPU; about 3.6 exaflops NVFP4 and 260 TB/s per rack; fully liquid-cooled. NVIDIA earlier labeled this rack "NVL144" by counting GPU dies.
Groq 3 LPX Third quarter of 2026 256 LPUs per rack for low-latency decode beside Vera Rubin
Vera CPU (standalone) 2026 Data center CPU for agent workloads; adopted by partners including Nebius and Nscale
Rubin Ultra in Kyber rack Second half of 2027 Vertical compute trays; about 600 kW per rack; about 15 exaflops FP4 (NVL576 counting dies)
Feynman 2028 New GPU (reportedly on TSMC A16), Rosa CPU, LP40 LPU, BlueField-5, CX10, copper and optical Kyber scale-up to 1,152 GPUs
DSX and DSX POD 2026 to 2027 Reference designs for complete AI factories; containerized Vera Rubin factories from partners in early 2027
DGX Spark and DGX Station From October 2025 Desktop AI supercomputers for developers (Linux)
RTX Spark Laptops from October 16, 2026 Arm SoC with MediaTek; 6,144 CUDA cores; up to 128 GB unified memory; Windows PCs
GeForce RTX 50 series and DLSS 2025 onward Blackwell consumer GPUs; DLSS 4.5 (2026) and DLSS 5 (announced at GTC 2026)
Jetson Thor and Isaac GR00T From August 2025 Robot computer and humanoid foundation models
DRIVE AGX Hyperion 10 and Alpamayo 2025 to 2027 Level 4 vehicle platform and open reasoning models for autonomous driving
Vera Rubin Space-1 Announced 2026 Radiation-tolerant compute module for orbital data centers; still in development

Partnerships, Investments, and Agreements

Date Partner What was agreed Status
Sept 2025 OpenAI Letter of intent: at least 10 GW of NVIDIA systems and up to $100 billion of investment Never finalized; replaced by a $30 billion equity investment (Feb 2026)
Sept 2025 Intel $5 billion equity stake; custom x86 CPUs on NVLink and x86 SoCs with RTX chiplets Investment closed; products in development
Oct 2025 Nokia $1 billion for about 2.9 percent; AI-RAN products for 5G-Advanced and 6G; T-Mobile US trials from 2026 Active; expanded with telecom operators at GTC Berlin 2026
Oct 2025 Uber Scale toward 100,000 Level 4 robotaxis from 2027 on DRIVE Hyperion; joint data factory Active
Oct 2025 South Korea (government, Samsung, SK, Hyundai, Naver) More than 260,000 Blackwell GPUs and joint AI factories Deploying
Oct 2025 US Department of Energy and Oracle Solstice and Equinox systems at Argonne In progress
Nov 2025 Anthropic and Microsoft Up to 1 GW on Grace Blackwell and Vera Rubin; NVIDIA invests up to $10 billion Active
Dec 2025 Groq Non-exclusive license and hiring of leadership; reported value about $20 billion Completed; Groq 3 LPX product
Jan 2026 Mercedes-Benz CLA ships with full DRIVE AV stack; partnership expanded Oct 2026 Active
June 2026 Microsoft and MediaTek RTX Spark Windows PCs for local AI agents Shipping since Oct 16, 2026
Sept 2026 Hugging Face Acquisition for $12.93 billion Pending regulatory approval; expected close first half of 2027
Oct 2026 Google Cloud Vera Rubin A5X instances General availability expected Nov 2026
Oct 2026 KYEC, SPIL, and other suppliers Capacity reservations and new US and Asian facilities Announced

Financial Results

Period Revenue Data Center Notes
Q3 FY2026 (ended Oct 2025) $57.0 billion $51.2 billion Blackwell ramp
Q4 FY2026 (ended Jan 2026) $68.1 billion $62.3 billion Networking about $11 billion in the quarter
Full FY2026 $215.9 billion $193.7 billion Up 65 percent; $41.1 billion returned to shareholders
Q1 FY2027 (ended Apr 2026) $81.6 billion $75.2 billion Networking $14.8 billion; dividend raised to $0.25; $80 billion added to buyback
Q2 FY2027 (ended July 2026) $96.2 billion $89.0 billion Gross margin 75.0 percent; Edge Computing $7.2 billion
Q3 FY2027 outlook About $108 billion Not broken out Excludes data center compute revenue from China; results expected late November 2026


The Physics

NVIDIA's products are shaped by a handful of physical and mathematical limits. Understanding them explains most of its design choices.

1. Parallelism and the end of easy scaling. Since Dennard scaling ended in the mid-2000s, smaller transistors no longer deliver proportionally lower power per transistor, so clock speeds stopped rising. Performance now comes from doing more operations in parallel and moving less data. A Blackwell GPU contains about 208 billion transistors across two dies; a Rubin GPU contains about 336 billion. Huang has claimed Rubin delivers roughly 5 times Blackwell's inference performance with only about 1.6 times the transistors, which, if borne out, would come from co-design (precision formats, memory, interconnect) rather than raw transistor count.

2. The roofline: compute versus memory bandwidth. Any workload is limited either by arithmetic throughput or by how fast data reaches the arithmetic units:

attainable FLOP/s = min(peak FLOP/s, memory bandwidth x arithmetic intensity)

Arithmetic intensity is operations performed per byte moved. A Rubin GPU's quoted figures are about 50 petaflops of NVFP4 (4-bit) inference and roughly 22 terabytes per second of HBM4 bandwidth, a ratio of about 2,300 operations per byte. Generating one token for one user with a dense model performs only about 2 operations per parameter, and at 4 bits each parameter is half a byte, so the intensity is only about 4 operations per byte. Single-stream decode therefore runs far below peak compute: it is memory-bound. This single fact explains batching (serving many users at once so each weight read is reused), the KV-cache memory problem, the value of more HBM per GPU, and NVIDIA's interest in Groq's SRAM-based LPUs.

A useful back-of-envelope bound for batch-of-one decode on a dense model:

max tokens per second (one sequence) ~ memory bandwidth / bytes of weights read per token

A 70-billion-parameter model stored at 4 bits occupies about 35 GB. At 22 TB/s, one GPU could stream those weights about 630 times per second, so roughly 630 tokens per second is an upper bound for a single sequence before overheads. Real systems differ (mixture-of-experts models read only some weights per token, KV-cache reads add traffic, and batching changes everything), but the bound shows why bandwidth, not FLOPs, often decides user-facing speed.

3. Precision as a physics lever. Halving the bits per number roughly doubles math throughput per unit of silicon and halves the memory and bandwidth needed per parameter. NVIDIA moved from FP16 to FP8 (Hopper) to FP4 (Blackwell and Rubin). Its NVFP4 format stores small blocks of 4-bit values (16 values per block) with a shared FP8 scale factor plus a per-tensor scale, which preserves accuracy far better than plain 4-bit numbers. The open question for every new format is how much model quality is lost, and that is measured model by model.

4. Interconnect and the size of the "one big GPU." Large models are split across many GPUs, which must constantly exchange data (all-reduce for training, all-to-all for mixture-of-experts routing). Communication time grows with message size divided by link bandwidth. NVLink 6 provides about 3.6 TB/s per GPU, and a Vera Rubin NVL72 rack totals about 260 TB/s of all-to-all bandwidth. Copper links are cheap and power-efficient but reach only a couple of meters, which is why NVIDIA packs 72 GPUs (and in Kyber, 144 GPU packages) into one rack and is moving to co-packaged optics (lasers integrated beside the switch chip) for longer links. Feynman's roadmap includes both copper and optical NVLink scale-up, reaching 1,152 GPUs across eight racks.

5. Power, heat, and the gigawatt factory. Every operation dissipates energy as heat. Rack power has climbed from roughly 10 to 20 kW in conventional servers to about 120 to 130 kW for a GB200 NVL72, and NVIDIA's 2027 Kyber rack for Rubin Ultra is designed for about 600 kW. Air cannot carry that heat, so Vera Rubin is fully liquid-cooled. The required water flow follows from basic thermodynamics:

flow (kg/s) = power (W) / (specific heat (J/kg K) x temperature rise (K))

For a 600 kW rack with a 10-degree temperature rise, that is about 600,000 / (4,186 x 10), or roughly 14 liters of water per second. NVIDIA designs Vera Rubin to accept coolant around 45 degrees Celsius, warm enough that facilities can reject heat with dry coolers instead of water-hungry chillers; its DSX reference designs claim near-zero facility cooling-water consumption. At the facility level, deals are now described in gigawatts: Anthropic's 2025 agreement covers up to 1 GW of NVIDIA systems, and the 2025 OpenAI letter of intent referenced 10 GW.

6. Cost per token as the governing metric. Combining these limits, the figure of merit becomes tokens per second per watt and tokens per dollar of capital. NVIDIA claims Vera Rubin lowers cost per token by up to about 10 times versus Blackwell and needs about one quarter as many GPUs to train large mixture-of-experts models, and at GTC Berlin Huang cited an analyst chart showing roughly 30 times better tokens per dollar than Blackwell. These are vendor and vendor-cited claims; independent benchmarks (for example MLPerf results and third-party inference studies) are the way to check them as systems reach customers.

Legacy Products and Earlier Resources

Several resources linked at the top of this page reflect earlier generations of NVIDIA technology and remain useful for learning:

  • NVIDIA Jetson Nano introduced many developers to edge AI. Its successors are the Jetson Orin family (including the low-cost Orin Nano Super developer kit) and, for robotics, Jetson Thor.
  • The A100 (Ampere, 2020) powered early large language model training, including the systems behind the original ChatGPT. It has been succeeded by Hopper, Blackwell, and Rubin.
  • Transfer Learning Toolkit became the TAO Toolkit; much of its role in adapting pre-trained models is now also served by NIM microservices and open model families.
  • GTC session archives from 2018 remain online, while current sessions live on NVIDIA On-Demand.
  • NVIDIA Omniverse™ has grown from a design-collaboration tool into the simulation layer for robotics, industrial digital twins, and the Omniverse DSX Blueprint for designing AI factories.

Further Resources

References

Financial results

Products and platforms

GTC Berlin and Europe

Partnerships, investments, and acquisitions

Policy, science, and Taiwan

Additional Viewing

Load video
YouTube

NVIDIA GTC Keynote 2026 (March 16, 2026, about 2 hours 20 minutes). Jensen Huang's flagship keynote launching Vera Rubin as a seven-chip, rack-scale platform for agentic AI, integrating Groq's LPU for low-latency inference, previewing Rubin Ultra in the 600 kW Kyber rack and the 2028 Feynman generation, and introducing NemoClaw, DSX AI factory designs, and Vera Rubin Space-1. It is the single best source for NVIDIA's current architecture and roadmap.

Load video
YouTube

NVIDIA GTC Taipei 2026 Keynote (June 1, 2026, about 2 hours). Huang declares that "useful AI has arrived," confirms Vera Rubin full production, launches the standalone Vera CPU for agent workloads, and introduces RTX Spark with Microsoft and MediaTek. Useful for understanding NVIDIA's push into CPUs and PCs and its deep dependence on Taiwan's supply chain.

Load video
YouTube

NVIDIA Live at CES 2026 with Jensen Huang (January 5, 2026, about 90 minutes). Huang presents the six-chip Vera Rubin platform and its claimed gains over Blackwell, introduces the Alpamayo reasoning models for autonomous driving, and frames 2026 as the year physical AI goes mainstream. Watch it alongside analysts' caution about how "full production" should be interpreted.

The GTC Berlin 2026 keynote (October 21, 2026) is available from NVIDIA's GTC Berlin keynote page.

slide-1-YZ-r7-768x462.jpg

Retrieved from "https://primo.ai/w/index.php?title=NVIDIA&oldid=48768"