NVIDIA NASDAQ NVDA Public

NVIDIA Deep Dive: Building the Full-Stack Physical-AI Infrastructure with PhysicsNeMo, Warp, Cosmos World Foundation Models and Omniverse

NVIDIA (NVDA) builds full-stack physical-AI infrastructure with PhysicsNeMo, Warp, Cosmos world foundation models and Omniverse digital twins; Isaac Lab 3.0 reached general availability on 20 October 2026, Cosmos 3 weights opened on 31 May, and the Hugging Face acquisition awaits regulatory approval.

Executive Summary

NVIDIA's position in physical AI cannot be captured in one sentence, because the company appears in four separate layers at once: compute for training and inference, GPU numerical libraries, a differentiable simulation language, and the world models and robot policies that run on top of simulation. Treating those four layers as one thing produces the wrong conclusion, namely that NVIDIA wants to become a CAE vendor. Separating them reveals what the company actually sells: the infrastructure that turns physics simulation into an AI workload. The revenue, however, still comes from compute.

This report makes three arguments.

First, physical AI is a genuine second growth curve, but it is not yet the revenue base. Fiscal 2026, which ended on 25 January 2026, produced revenue of 215.9 billion dollars, up 65 percent. The second quarter of fiscal 2027, ended 26 July 2026, produced 96.22 billion dollars, up 106 percent year on year, of which 89.0 billion came from Data Center, or 92.5 percent, at a gross margin of 75.0 percent. Guidance for the third quarter is 108.0 billion dollars, and it explicitly assumes no Data Center compute revenue from China. In the same filings, Omniverse, Isaac, PhysicsNeMo and DSX have no separately reported revenue. Any statement that physical AI already contributes a specific dollar figure therefore lacks a primary source.

Second, the moat sits in numerical libraries and differentiable simulation, not in rendering. Most observers equate NVIDIA's simulation capability with the visual output of Omniverse, which mistakes the display layer for the capability layer. The checkable assets are the sparse solvers and quantum chemistry libraries inside CUDA-X, namely cuDSS for direct sparse systems, cuISS for iterative sparse systems and cuEST for electronic structure; Warp, a Python-fronted simulation language that compiles to the GPU and supports differentiation; and Newton, an open-source physics engine written in Warp. Those three determine whether physics equations can run on GPUs in a differentiable, batched form. Omniverse only determines how good it looks once it runs.

Third, open source is an ecosystem instrument, not a commercial concession. Newton sits under Apache-2.0 within the Linux Foundation, and weights for Cosmos and Isaac GR00T are published openly, which looks inconsistent with a company running a 75 percent gross margin. The context resolves the apparent contradiction. Open licensing lowers the barrier to training robots on NVIDIA's simulation stack and raises the compute demand that follows. Open source here generates demand.

Evidence grading must come first. In this report, financial statements and 10-K disclosures are primary evidence. Performance multiples in company press releases, including 475x, 252x, 10x, 50x and 20x, are vendor claims without a published test matrix. Independent third-party reproduction is essentially absent across the physical AI line. Wherever only a vendor claim exists, this report says so rather than presenting the multiple as fact.

This revision adds three delivery-status changes, each recorded in DISCREPANCIES. Isaac Lab 3.0 has reached general availability: the v3.0.0 tag appeared on GitHub on 20 October 2026 and NVIDIA announced general availability at GTC Berlin on 21 October. Cosmos 3 is no longer a forthcoming model, since the Nano and Super checkpoints have been downloadable from Hugging Face and the Cosmos 3 repository since 31 May 2026, with the Edge and DROID policy variants added later. NVIDIA also signed a definitive agreement to acquire Hugging Face on 2 September 2026 at a headline consideration of about 12.93 billion dollars, with completion expected in the first half of 2027 subject to regulatory approval. The previous revision of this report described Isaac Lab 3.0 as being in early access and Cosmos 3 as announced with delivery unconfirmed, and both statements are superseded by what the official channels now show.

Company Overview

The legal entity is NVIDIA Corporation, incorporated in California in April 1993 and reincorporated in Delaware in April 1998, headquartered at 2788 San Tomas Expressway, Santa Clara, California, and traded on the Nasdaq Global Select Market as NVDA. The fiscal year ends on the last Sunday in January. Fiscal 2026 closed on 25 January 2026, and fiscal 2027 is a 53-week year. That detail matters, because coverage frequently conflates fiscal 2026 with calendar 2026, and the two are almost a year apart.

At the end of fiscal 2026 the company employed approximately 42,000 people across 38 countries, of whom roughly 31,000, or 74 percent, worked in research and development. More than 80 percent of employees hold technical roles, more than half hold an advanced degree, turnover was 3.7 percent, more than 40 percent of new hires came from employee referrals, and cumulative research and development investment since inception exceeds 76.7 billion dollars. Revenue of 215.9 billion dollars against 42,000 employees implies about 5.1 million dollars of revenue per employee, a figure without peer among large technology companies.

Reporting is organised into two segments, Compute and Networking and Graphics, which produced 88.3 billion and 7.9 billion dollars respectively in the second quarter of fiscal 2027. The platform view splits differently: Data Center at 89.0 billion dollars, comprising 48.7 billion from Hyperscale and 40.3 billion from AI Clouds, Industrial and Enterprise, plus Edge Computing at 7.2 billion. That split deserves separate attention, because Hyperscale accounts for 55 percent of Data Center revenue and is the starting point for understanding customer concentration.

Read backwards through a physical AI lens, a handful of dates explain the current position. CUDA arrived in 2006 and turned a graphics processor into a general parallel processor, which is the precondition for everything that follows. Real-time ray tracing and RTX in 2018 brought physically correct rendering into consumer hardware. The Mellanox acquisition in 2019 pulled networking into system-level design. Omniverse in 2021 was the first attempt to sell collaborative real-time digital twins as a product. At GTC 2025, NVIDIA joined Google DeepMind and Disney Research to launch the Newton physics engine project and handed it to the Linux Foundation, the first time core simulation capability was released as a foundation-governed open project. GTC on 16 March 2026 was the densest physical AI announcement to date: Cosmos 3, Isaac GR00T N1.7, a GR00T N2 preview, Alpamayo 1.5, Newton 1.0 general availability, Isaac Sim 6.0, the Vera Rubin DSX reference design and the Omniverse DSX blueprint, all on one day. On 26 July 2026, PhysicsNeMo was re-architected into agent-callable libraries and folded into the Agent Toolkit alongside cuISS, cuDSS and cuEST. In October 2026, GTC Berlin carried the same narrative into Europe and turned part of it into shipping software: on 21 October NVIDIA announced Isaac Lab 3.0 general availability, Isaac OS, the SimReady Central Library, the Human 1M dataset and three open reference workflows for DGX Station, and made the FOX factory-operations blueprint available. The Cosmos platform was extended on the same day with Cosmos Evaluator 26.10, an updated Cosmos 3 NIM, and the Cosmos Data Factory and Cosmos Framework.

Four Common Misreadings

Misreading one: treating NVIDIA as a CAE solver vendor. NVIDIA does not sell structural, fluid or electromagnetic solver licences. It sells the layer that makes other vendors' solvers run faster on GPUs and become callable by agents. The solvers still belong to Ansys, Siemens, Dassault, Cadence, Synopsys and COMSOL, and most of those companies are simultaneously NVIDIA partners.

Misreading two: reading the Isaac Lab 3.0 status from a single press-release wording. An NVIDIA blog retrospective on GTC in April 2026 described Newton 1.0, Isaac Sim 6.0 and Isaac Lab 3.0 as generally available, which was premature for Isaac Lab 3.0 at that date. The repository timeline then showed beta on 17 March, beta 2 on 17 June, Early Access on 16 September, with general availability targeted for the end of October 2026, which in turn invited the opposite error of treating the framework as still unreleased. The checked position is narrower than either. The v3.0.0 tag appeared on GitHub on 20 October 2026, the NGC Isaac Lab container 3.0.0 carries the same date, and NVIDIA announced general availability at GTC Berlin on 21 October 2026. The release targets Isaac Sim 6.2, Python 3.12, PyTorch 2.11, CUDA 13.0, Warp 1.17, Newton 1.6.1 and RSL-RL 5.5.1, and it changes the quaternion convention to x, y, z, w, which makes the framework a migration event for existing 2.x code rather than a drop-in upgrade.

Misreading three: treating the March announcement as the delivery event for Cosmos 3. On 16 March 2026 the official wording was that Cosmos 3 is expected to come soon, with no date. The checkpoints are in fact available: the Cosmos 3 collection on Hugging Face and the Cosmos 3 repository both carry a release date of 31 May 2026, and the NVIDIA NGC container for Cosmos3 lists the same date. What is not available is a single model. The release covers Cosmos 3 Nano with 16 billion parameters for workstation inference and Cosmos 3 Super with 64 billion parameters for datacenter deployment, with Edge variants and DROID policy variants added later under the OpenMDW 1.1 licence.

Misreading four: adding China revenue back into guidance. The 108.0 billion dollar guidance for the third quarter of fiscal 2027 states explicitly that it assumes no Data Center compute revenue from China. Any model that restores a China opportunity to that guidance overstates a single quarter by several billion dollars.

Technology

The Numerical Foundation: Four New Components in CUDA-X

Roughly ninety percent of the compute bottleneck in physics simulation sits in solving sparse linear systems. In July 2026 NVIDIA pushed four components forward.

cuDSS is a direct sparse solver aimed at large sparse systems in electronic design automation and scientific computing, emphasising numerical robustness and scaling across multiple GPUs and nodes, and it is generally available. cuISS is a newly introduced iterative sparse solver library offering composable solvers and preconditioners for the ill-conditioned matrices common in physics simulation. cuEST is an electronic structure library that moves density functional theory and post-DFT methods onto GPUs. AmgX, cuSPARSE and cuFFT sit beneath and beside them.

The significance is that the solver kernel, historically locked inside commercial CAE packages, becomes a library any developer can call. Keysight states publicly that it achieved up to 10x improvement in electromagnetic simulation using cuDSS, Samsung applies PhysicsNeMo to chip-scale thermal stress analysis, and TSMC and Synopsys integrate cuEST into quantum chemistry workflows where NVIDIA quotes an optimistic figure of up to 50x. These are vendor and partner claims without an independent test matrix, but the direction is unambiguous: the GPU is moving from accelerator to host for the solver itself.

Warp: A Language Written for Physics

Warp is the key to NVIDIA's physical AI strategy and the least discussed part of it. It is a simulation domain-specific language fronted by Python and compiled to the GPU, and its central property is differentiability. A simulation can be differentiated with respect to its parameters, so gradients pass straight through the physics solver into optimisation, inversion and reinforcement learning.

Without differentiability, physics simulation and machine learning are two languages joined by fitted data. With it, simulation becomes part of the network. Writing Newton entirely in Warp moves the physics engine from the C++ world of CPUs into the Python world of GPUs.

PhysicsNeMo: From Framework to Agent Skill

PhysicsNeMo began as a physics machine learning framework, formerly Modulus, dedicated to learning PDE solution operators and building surrogates. On 26 July 2026 NVIDIA re-architected it into a set of agent-friendly libraries inside the Agent Toolkit, paired with Nemotron 3 Ultra to form what the company calls an autonomous AI engineer stack.

The substance of that move is to demote the task of training a physics surrogate from a researcher's project to a tool call an agent can make. Announced partners span Cadence, Synopsys, Siemens, Samsung, TSMC, Keysight, Silvaco and ChipAgents, with Cadence claiming up to 20x improvement in advanced packaging and printed circuit board multiphysics workflows.

A caution is warranted. PhysicsNeMo's most prominent public result previously was SimScale's centrifugal pump foundation model in March 2025, claimed at roughly 2,700 times faster than conventional CFD. What such numbers share is that the reference case, mesh size and accuracy tolerance are all undisclosed.

Newton: Handing the Physics Engine to a Foundation

Newton 1.0 is the most concrete artefact of GTC 2026. It is open source under Apache-2.0, governed by the Linux Foundation, developed jointly by NVIDIA, Google DeepMind and Disney Research, with Toyota Research Institute joining later.

Four technical characteristics are worth recording. First, the solver set is a matrix rather than a single engine. MuJoCo Warp serves as the primary rigid-body solver, while Featherstone, XPBD, VBD, Implicit MPM and Style3D cover articulated mechanisms, soft bodies, deformables, granular materials and cloth, and Kamino, contributed by Disney Research, handles closed-loop mechanisms such as parallel-linkage legs. Second, collision uses signed distance fields and ingests CAD meshes directly, avoiding the geometric loss of convex hull approximation. Third, contact modelling uses hydroelastic contact inherited from Toyota Research Institute's Drake engine, producing distributed pressure across contact patches. Fourth, deformables couple bidirectionally with rigid bodies, so VBD lets a robot genuinely bend a cable during assembly simulation.

The performance claim is that MuJoCo Warp on an RTX PRO 6000 Blackwell reaches up to 252x Google DeepMind's MJX on locomotion tasks and up to 475x on manipulation tasks. Both figures hold only against MJX as the baseline and cannot be generalised into a claim of being 475 times faster than all physics engines.

Public production deployments currently number three. Skild AI trains GPU rack assembly policies using Newton's signed distance field collision and hydroelastic contacts. Samsung uses its cable simulation to generate vision-language-action training data. Lightwheel uses it to calibrate SimReady assets. All three are early adopters and do not constitute evidence of scale.

Isaac Sim 6.0, Isaac Lab 3.0 and Omniverse

Isaac Sim 6.0 has shipped, with Newton available as a swappable physics backend. Isaac Lab 3.0 was the item that required the most precise treatment in this report, and that precise treatment has now changed. The v3.0.0 tag appeared on GitHub on 20 October 2026, the NGC Isaac Lab container 3.0.0 carries the same date, and general availability was announced at GTC Berlin on 21 October 2026, after which NVIDIA Robotics posted the release note on 22 October. The release targets Isaac Sim 6.2 rather than the 6.0 used by the earlier betas, requires Python 3.12 and PyTorch 2.11 with CUDA 13.0, and pins Warp 1.17, Newton 1.6.1 and RSL-RL 5.5.1. Physics, rendering and visualisation became independent choices, selectable as newton_mjwarp, ovphysx or isaacsim_physx for physics and newton_renderer, ovrtx or isaacsim_rtx for rendering, and the Newton, OVPhysX and OVRTX workflows run without launching Isaac Sim at all. The framework also adds Isaac Capture teleoperation, LEAPP policy export, and seventeen browser-based interactive demos that run Newton through WebAssembly on a local CPU.

On the Omniverse side the material change is DSX. Vera Rubin DSX is an AI factory reference design, and the Omniverse DSX blueprint provides physically accurate digital twins of AI factories, generally available on 16 March 2026. It decomposes into six parts. DSX Sim validates layout, power, cooling and network design. DSX OS takes over operations once power is on. DSX Max-Q raises tokens per watt inside a fixed power budget. DSX Flex lets a factory interact with the grid. DSX Exchange bridges IT and operational technology signals. DSX SimReady supplies three-dimensional assets carrying physical properties. The official claim is that a 500 megawatt AI factory can host up to 30 percent more GPUs within the same power envelope using Max-Q.

The implication for the CAE industry is underrated. DSX turns the data centre into an object requiring multiphysics simulation across power, thermal, structural and fluid domains, at a moment when per-gigawatt investment has risen from 20 to 30 billion dollars toward 80 to 100 billion dollars. Simulation stops being a research cost and becomes a risk control instrument for capital expenditure.

The same GTC Berlin announcement widened the platform around the framework. Isaac OS is an open-source runtime built on Holoscan that connects AI models, sensors, control systems and robot hardware, so that behaviour developed in simulation reaches physical machines with deterministic real-time scheduling. The SimReady Central Library, published on Hugging Face, offers more than 2,000 simulation-ready robot and factory assets from Comau, FANUC, ForgeCAD, Franka Robotics, LG Electronics, Palatial, Robotiq, Standard Bots, Universal Robots and Vertiv, and the Human 1M dataset carries diverse human activity with synchronised three-view video and 3D hand tracking, with the complete dataset available to Cosmos Coalition partners rather than to the general download. Three open reference workflows for DGX Station cover reinforcement-learning fine-tuning of GR00T 1.7, training of dexterous manipulation skills in Isaac Lab, and post-training of the Cosmos world model with ChatGPT Astra to generate synthetic data, with 1X, Agile Robots, Humanoid, Maven Robotics, Mentee, MicroAGI, Rhoda, Sereact and Skild AI supporting them. Agile Robots used the same event to publish Agile ONE SN, a reference humanoid standing 1.40 metres and weighing 40 kilograms with 31 degrees of freedom and a 3 kilogram single-arm payload, two WUJI Hand 2 five-finger hands at 20 degrees of freedom each with 1 kHz force control and tactile sensing, and Jetson Thor providing 2,070 FP4 TFLOPS and 128 GB of unified memory, with availability planned for early 2027. Separately, the FOX factory-operations blueprint became available, packaging CUDA-X, Omniverse and open models into a production blueprint for European industrial deployment.

World Models and Policy Models

Cosmos 3 is the first world foundation model unifying synthetic world generation, physical AI reasoning and action simulation, and it has been open since 31 May 2026. Its architecture is a Mixture-of-Transformers design with two towers: an autoregressive vision-language reasoner tower that interprets images, video and text, and a diffusion-based generator tower that produces physics-aware video and action sequences conditioned on the reasoner. The reasoner can be called on its own, while guided generation always activates both. Cosmos 3 Nano with 16 billion parameters targets workstation inference on devices such as the RTX PRO 6000, and Cosmos 3 Super with 64 billion parameters targets datacenter deployment on Hopper and Blackwell. NVIDIA reports leading positions on VANTAGE-Bench, PAI-Bench, R-Bench Physics-IQ and RoboLab, and open datasets, training recipes and post-training code are published for robotics, physics simulation, human motion, driving and warehouse environments. The deployment side runs as Cosmos 3 NIM microservices, and an updated version delivered up to 2x faster inference with no quality loss in validated configurations, with the reasoner NIM built on vLLM and offering BF16, FP8 and NVFP4 checkpoints, the 4-bit path producing the stated speedup.

Isaac GR00T N1.7 is a three-billion-parameter open reasoning vision-language-action model pretrained on 20,854 hours of human egocentric video, with open weights on Hugging Face from 17 April 2026 and early access under commercial licensing announced at GTC. GR00T N2 is based on DreamZero research and uses a world action model architecture; NVIDIA claims more than twice the success rate of leading vision-language-action models on new tasks in new environments, first place on both MolmoSpaces and RoboArena, and availability by the end of the year. As of 29 September 2026 it had not been released.

The platform around Cosmos 3 grew on 21 October 2026. Cosmos Evaluator 26.10 automates grading of synthetic video outputs across four reference checks: hallucination, object consistency, attribute verification and physical plausibility. Cosmos Data Factory generates synthetic training datasets and Cosmos Framework handles training, distillation and quantisation, with MicroAGI reporting that it distilled a 32-billion-parameter Cosmos 3 reasoner down to an 8-billion-parameter Nano model without accuracy loss. For adjacent video workloads, Metropolis Blueprint for video search and summarisation 3.3, built on Cosmos NIM microservices, is stated to monitor 40 percent more video streams on the same GPU, and ADNOC is cited as running Fogsphere's MistIQ agentic platform on Cosmos across 65 sites.

Alpamayo 1.5 is a reasoning vision-language-action model for autonomous driving, adding navigation guidance, prompt conditioning, flexible multi-camera support and configurable camera parameters. A separate healthcare line includes Open-H-Embodiment, an open medical robotics dataset of roughly 778 hours contributed by 35 organisations; GR00T-H, a policy model built on a two-billion-parameter Cosmos-Reason 2 backbone post-trained on roughly 600 hours of surgical data; and Cosmos-H, a surgical simulator.

Layering Technical Credibility

Sorting these claims by verifiability produces a steadier conclusion.

Verifiable: the Newton 1.0 code and licence, with the v1.0.0 tag dated 13 April 2026 on GitHub; the Isaac Lab 3.0 release timeline in GitHub releases; the open weights of GR00T N1.7, downloadable from Hugging Face with checkable parameter count and pretraining volume; and the general availability status of cuDSS.

Partially verifiable: the Isaac Sim 6.0 feature list, the component structure of the DSX blueprint, and the partner roster around Omniverse.

Not verifiable: every performance multiple, including 252x, 475x, 10x, 20x, 50x and 2,700x, because test matrices, baseline definitions and accuracy tolerances are absent; and every statement about production scale, because customer counts and contract values are absent.

Product and Commercial Model

Commercialisation in physical AI is a three-tier structure with markedly different maturity in each tier.

The first tier is hardware, and it is mature. DGX and HGX systems, GB200 and GB300 racks, and the Vera Rubin platform constitute the compute itself, with Vera Rubin entering full production in the second quarter of fiscal 2027. Revenue at this tier is the 89.0 billion dollar Data Center line.

The second tier is software subscription, at undisclosed scale. NVIDIA AI Enterprise is licensed per GPU per year, Omniverse carries its own subscription, and the Isaac family ships with the platform. Financial statements never break this out, placing it inside an unsegmented software and services discussion. Any specific Omniverse revenue figure is an external estimate.

The third tier is reference design, and its function is demand generation. The Vera Rubin DSX reference design, the Omniverse DSX blueprint and Isaac robot reference workflows are not directly monetised. Their purpose is to make NVIDIA's stack the default when a customer builds a factory, a line or a robotics team. The GTC 2026 partner roster illustrates the mechanism: Cadence, Dassault Systemes, Eaton, Jacobs, Nscale, Phaidra, Procore, PTC, Schneider Electric, Siemens, Switch, Trane Technologies and Vertiv contribute platforms, SimReady assets or software integration, while Emerald AI, GE Vernova, Hitachi and Siemens Energy address the power side. Half of that list consists of companies that compete with NVIDIA in traditional CAE and industrial software.

One set of ecosystem numbers is worth recording. FANUC, ABB, YASKAWA and KUKA together operate an installed base above two million robots, and they are integrating Omniverse libraries and Isaac simulation frameworks into virtual commissioning while adding Jetson modules to controllers. On the humanoid and general-purpose side, 1X, AGIBOT, Agility, Agile Robots, Boston Dynamics, Figure, Hexagon Robotics, Humanoid, Mentee and NEURA Robotics use Cosmos, Isaac Sim and Isaac Lab. Among robot brain developers, FieldAI, Skild AI, World Labs and Generalist AI also appear.

A side effect of the model deserves stating. When NVIDIA supplies the same reference design to a customer and to that customer's competitor, neutrality becomes an operating requirement. That is the structural reason for placing Newton under the Linux Foundation and opening Cosmos and GR00T weights: neutrality converts into ecosystem share.

Competitive Landscape

Against the traditional CAE leaders, partnership dominates competition. On 16 March 2026, NVIDIA and Cadence, Dassault Systemes, PTC, Siemens and Synopsys jointly announced bringing CUDA-X and Omniverse into industrial software. Those companies need GPUs for their solvers to keep pace with customer demand for scale and iteration speed, and NVIDIA needs them to hold the professional solver entry point. Real tension appears one layer up: once PhysicsNeMo lets agents build surrogates themselves and Omniverse lets agents build twins themselves, the interface and workflow value of a CAE vendor compresses.

Against cloud providers, the relationship mixes customer and substitute. Hyperscale contributes 55 percent of Data Center revenue, while AWS, Azure and Google Cloud all develop their own accelerators and offer simulation and twin services on their platforms. NVIDIA's response is to productise the AI factory. DSX is not a GPU sale; it packages power, cooling, networking and operations into a reference design that cloud providers follow when expanding capacity.

Against robot model companies such as Physical Intelligence, Figure and Skild AI, the relationship is both customer and competitor. NVIDIA trades open weights in GR00T N1.7 and Cosmos for ecosystem position. The cost is continued model investment; the benefit is that training and inference land on its hardware as those companies scale. The risk is that if world models consolidate into a few closed vendors, the open strategy becomes building infrastructure for others.

Against the open simulation ecosystem, including OpenFOAM, MuJoCo and FEniCS, the pattern is absorption. After Google DeepMind open-sourced MuJoCo, NVIDIA made it GPU-native through MuJoCo Warp and adopted it as Newton's primary rigid-body solver. The pattern shows how NVIDIA handles open source: it takes over the performance implementation layer and thereby controls the de facto standard without owning the upstream project.

Against quantum computing and new architectures, near-term competition is absent. cuEST moves quantum chemistry onto GPUs, which in effect completes high-accuracy electronic structure work on classical hardware before quantum computers become practical.

Financing and Capital Structure

NVIDIA has no external venture financing history. It was founded in 1993, listed on Nasdaq in January 1999, and subsequent capital activity has centred on equity compensation, buybacks and dividends.

In fiscal 2026 the company returned 41.1 billion dollars to shareholders through repurchases and dividends, leaving 58.5 billion dollars of authorisation at year end, with a quarterly dividend of 0.01 dollar per share. In the second quarter of fiscal 2027 it returned approximately 26.0 billion dollars, leaving about 99.0 billion dollars of authorisation, and raised the quarterly dividend to 0.25 dollar per share, twenty-five times the prior level, payable on 1 October 2026. That jump is a public statement about confidence in cash flow durability.

The key financial sequence, on a GAAP basis unless noted, runs as follows. The fourth quarter of fiscal 2026 produced 68.1 billion dollars of revenue, up 20 percent sequentially and 73 percent year on year, with Data Center at 62.3 billion, up 75 percent, a gross margin of 75.0 percent and diluted earnings per share of 1.76 dollar. Full fiscal 2026 revenue was 215.9 billion dollars, up 65 percent, with a 71.1 percent gross margin and diluted earnings per share of 4.90 dollars.

The first quarter of fiscal 2027 produced 81.62 billion dollars of revenue, a 74.9 percent gross margin and diluted earnings per share of 2.39 dollars. The second quarter produced 96.22 billion dollars, up 18 percent sequentially and 106 percent year on year, a 75.0 percent gross margin, net income of 59.69 billion dollars and diluted earnings per share of 2.46 dollars, or 2.22 dollars non-GAAP. Third quarter guidance is 108.0 billion dollars plus or minus 2 percent, with a 74.0 percent gross margin plus or minus 50 basis points and no China Data Center compute revenue.

Valuation requires care. Using a market capitalisation of about 5.14 trillion dollars on 22 October 2026 against annualised second-quarter fiscal 2027 revenue of roughly 384.9 billion dollars gives a static price-to-sales ratio of about 13.4 times, which is this report's calculation rather than a company disclosure. Within the same week, different sources quote 5.10 trillion, 5.14 trillion and roughly 4.9 trillion dollars, reflecting differences in pricing time, share count basis and treatment of in-period repurchases.

One non-equity transaction deserves separate mention. In December 2025 NVIDIA obtained a non-exclusive licence to Groq LPU technology and reports it within Compute and Networking. Such licences are invisible in capital structure but may matter more to the inference cost curve than an ordinary investment.

The Hugging Face agreement is the largest transaction in the company's history by a wide margin and is still pending. NVIDIA signed a definitive agreement on 2 September 2026 and filed an 8-K on 3 September, with the announcement the same day. The headline consideration given by the chief executive is about 12.93 billion dollars, and the securities filing breaks that into about 11.9 billion dollars payable to Hugging Face stockholders subject to adjustment, plus an equity-based retention programme for employees worth up to about 1 billion dollars. Completion is expected in the first half of 2027 subject to regulatory approval, and the platform remains legally independent until then. On 3 September NVIDIA also stated that Hugging Face would remain an open platform, that NVIDIA compute would not be required to build on or deploy through it, and that multi-cloud and multi-accelerator support would continue, a commitment the filing repeats. The scale figures attached to the platform are the vendor's own: more than 18 million developers, researchers and creators, more than 3 million models, 500,000 datasets, 1 million applications and more than 200,000 companies.

DISCREPANCIES

1. The Isaac Lab 3.0 delivery date has three defensible readings, and the earlier reading in this report is superseded. (Source grade: OFFICIAL versus CODE versus MEDIA) An NVIDIA blog on 17 April 2026, reviewing GTC, described Newton 1.0, Isaac Sim 6.0 and Isaac Lab 3.0 together as generally available, which was premature for Isaac Lab 3.0 at that date. The repository timeline subsequently showed beta on 17 March 2026, beta 2 on 17 June 2026, Early Access on 16 September 2026, and general availability targeted for the end of October 2026. The actual general-availability date now has three anchors: the v3.0.0 GitHub tag and the NGC container 3.0.0 dated 20 October 2026, the official announcement at GTC Berlin on 21 October 2026, and the NVIDIA Robotics community post on 22 October 2026. An earlier revision of this report concluded from the timeline alone that Isaac Lab 3.0 remained in early access, and that conclusion is withdrawn. This report now records general availability as delivered on 20 October 2026, with the announcement on 21 October.

2. Newton speedups exist in two framings. (Source grade: OFFICIAL versus MEDIA) NVIDIA and first-hand technical coverage report MuJoCo Warp against MJX on an RTX PRO 6000 Blackwell at up to 252x on locomotion and up to 475x on manipulation. Some secondary outlets compress this into Newton being 475 times faster than MJX, dropping both the task dimension and the hardware premise. This report retains both figures and names the baseline rather than adopting a single multiple.

3. The parameters and data for GR00T N1 and N1.7 are conflated. (Source grade: MEDIA versus OFFICIAL) Some secondary coverage attributes the figures of 780,000 synthetic trajectories, eleven hours and the equivalent of nine months of human demonstration to GR00T N1.7 at GTC 2026. Those figures belong to GR00T N1, announced at GTC 2025. The published position for N1.7 is three billion parameters, 20,854 hours of human egocentric video pretraining and open weights from 17 April 2026. This report follows the Hugging Face model card and the official blog.

4. The Cosmos 3 delivery date is confirmed by release metadata rather than by a headline announcement. (Source grade: OFFICIAL versus CODE versus MEDIA) The 16 March 2026 release says expected to come soon without a date, and secondary coverage placed the weights around 1 June 2026 without an official announcement to anchor them. The official channels now carry the date in metadata: the Cosmos 3 collection on Hugging Face and the Cosmos 3 repository both show 31 May 2026, and the NGC Cosmos3 container lists the same release date, with the Edge and DROID policy variants added later. An earlier revision of this report treated Cosmos 3 as announced with delivery unconfirmed, and that treatment is withdrawn. This report records Cosmos 3 as open since 31 May 2026.

5. GR00T N2 availability has been moved forward in coverage. (Source grade: OFFICIAL versus MEDIA) The official position is availability planned for the end of the year, and as of 29 September 2026 it had not been released. Some coverage states in headlines that it has shipped or is available. This report treats it as previewed and unreleased.

6. Market capitalisation figures differ across sources. (Source grade: DB) Within the week of October 2026, sources quote 5.103 trillion, 5.143 trillion and roughly 4.9 trillion dollars, reflecting pricing time, share count basis and treatment of buybacks. This report standardises on about 5.14 trillion dollars as of 22 October 2026 and states the time point in any calculation.

7. Fiscal-year framing is systematically confused. (Source grade: OFFICIAL versus MEDIA) Fiscal 2026 ended on 25 January 2026 and fiscal 2027 is a 53-week year. Chinese-language coverage frequently abbreviates the second quarter of fiscal 2027 as the second quarter of 2026, misaligning the year-on-year base. This report always writes fiscal 2027 second quarter and states the corresponding calendar interval.

8. The Hugging Face consideration appears in three figures. (Source grade: OFFICIAL versus FILING versus MEDIA) The chief executive's announcement on 3 September 2026 gives a total of about 12.93 billion dollars, and the 8-K filed the same day breaks that into about 11.9 billion dollars payable to stockholders, subject to adjustment, plus an equity-based retention programme worth up to about 1 billion dollars. One secondary outlet instead reports a total of about 17.9 billion dollars, with about 16.5 billion for stockholders and about 1.4 billion for retention. This report adopts the announcement and filing figures, which are mutually consistent, and records the larger figure as an unreconciled secondary claim rather than a company disclosure.

Risks and Diligence Gaps

Revenue concentration is the first risk. Of 89.0 billion dollars of Data Center revenue in the second quarter of fiscal 2027, Hyperscale accounts for 48.7 billion, or 55 percent. The capital expenditure rhythm of a small number of hyperscale customers therefore drives quarterly variance directly. Filings do not disclose how the top customers' share has moved, which is the first primary datum any diligence exercise must add.

Policy and geopolitical risk is already priced into guidance. Third quarter guidance of 108.0 billion dollars assumes no China Data Center compute revenue, and the 10-K risk factors describe both Chinese government encouragement of domestic chip purchasing and restrictions on US data centre products. The company itself is treating that revenue as zero.

Supply chain and power are physical constraints. Vera Rubin ramp depends on HBM and advanced packaging capacity. The DSX narrative depends on power. The official framing is that per-gigawatt AI factory investment has moved from 20 to 30 billion dollars toward 80 to 100 billion dollars, with 100 gigawatts of AI factories expected online before the end of the decade. That forecast is simultaneously demand narrative and risk exposure: any delay in grid interconnection shifts the entire DSX deployment schedule.

Physical AI commercialisation is not yet validated. Three things can be confirmed: Newton has three publicly named early production users in Skild AI, Samsung and Lightwheel; GR00T N1.7 is available through both open weights and commercial licensing; and DSX has a long partner roster. What is missing is paid customer counts, contract values, renewal rates and any customer-disclosed return on investment.

The durability of open-source commitments is uncertain. Newton under the Linux Foundation and open weights for Cosmos and GR00T all require sustained investment without direct revenue. Whether NVIDIA maintains the same intensity if compute demand growth slows is an open question.

The Hugging Face transaction carries a regulatory question that no filing can settle. The concern is structural: an owner that supplies the dominant training and inference silicon also controls a distribution point used by rivals. Reported enquiries from the United States Department of Justice focus on whether NVIDIA could use Hugging Face to disadvantage competing model companies or gain leverage over them. Such an enquiry is not a finding of harm and may end without action, and the acquisition has not closed. The measurable obligation that follows from closing is the platform's continued support for competing silicon providers, and no post-closing evidence on that exists yet.

Diligence gap list, ordered by importance. First, paid seats and annual revenue for Omniverse and Isaac, under any split. Second, downstream customer counts and typical deployment size for PhysicsNeMo. Third, signed DSX projects and per-project values. Fourth, how Newton's production users expand beyond three, with published sim-to-real transfer results. Fifth, counterparties and fee structure of GR00T commercial licences. Sixth, a time series of top-five customer revenue share. Seventh, actual China revenue by quarter in fiscal 2026 and fiscal 2027, to calibrate how conservative the zero assumption is. Eighth, the outcome of the regulatory review of the Hugging Face acquisition, and any evidence on whether competing accelerators remain supported on the platform after closing.

Composite Assessment

techScore 9.0. NVIDIA is the only company holding a differentiable simulation language in Warp, an open-source physics engine in Newton, sparse solver libraries in cuDSS and cuISS, world models in Cosmos and policy models in GR00T and Alpamayo inside one stack, with code and weights that can be inspected. The deduction reflects the absence of third-party reproduction for performance claims, the short release cycle of Isaac Lab 3.0 that has left benchmark data for the current version thin, and the migration burden created by the quaternion-convention change in the 3.0 line.

bizScore 8.5. Revenue grew 106 percent year on year in the second quarter of fiscal 2027 at a 75.0 percent gross margin, with 26.0 billion dollars returned to shareholders in a single quarter, so commercial execution is not in dispute. The deduction comes from customer concentration at 55 percent Hyperscale and from the complete absence of disclosure on physical AI software revenue.

fundScore 9.5. The company needs no external financing, returned 41.1 billion dollars in fiscal 2026 and 26.0 billion dollars in one quarter of fiscal 2027, holds 99.0 billion dollars of remaining repurchase authorisation and raised its dividend twenty-five fold. This is the strongest dimension in the assessment.

productScore 9.0. The four-tier structure from silicon to reference design is complete, and every tier has a usable open or commercial entry point. The deduction reflects opaque pricing and packaging relationships between tiers, which make total cost of ownership hard for a customer to compute.

marketScore 9.5. A market capitalisation of 5.14 trillion dollars, first in the world, with definitional authority over physical AI and default-platform status across both industrial software and humanoid robotics ecosystems. The deduction reflects the regulatory and geopolitical exposure that this position itself creates, now compounded by the antitrust review of the Hugging Face acquisition.

References

  1. NVIDIA financial results for the fourth quarter and fiscal 2026, 25 February 2026, investor.nvidia.com, including fiscal 2026 revenue of 215.9 billion dollars, fourth quarter revenue of 68.1 billion dollars, full-year gross margin of 71.1 percent and 41.1 billion dollars returned to shareholders
  2. NVIDIA financial results for the second quarter of fiscal 2027, 26 August 2026, investor.nvidia.com, including revenue of 96.22 billion dollars, Data Center of 89.0 billion dollars, a 75.0 percent gross margin and third quarter guidance of 108.0 billion dollars excluding China Data Center compute
  3. NVIDIA Form 10-K for fiscal 2026, filed 25 February 2026, including 42,000 employees in 38 countries, 31,000 in research and development, 3.7 percent turnover, 76.7 billion dollars of cumulative research and development, headquarters address, Delaware incorporation and fiscal year definition
  4. NVIDIA Form 10-Q for the quarter ended 26 July 2026, filed 26 August 2026, including segment revenue of 88.3 billion dollars for Compute and Networking and 7.9 billion for Graphics, the platform view of 48.7 billion Hyperscale, 40.3 billion AI Clouds, Industrial and Enterprise and 7.2 billion Edge Computing, the 53-week fiscal 2027, and the non-exclusive Groq LPU licence
  5. NVIDIA press release, NVIDIA Expands Open Model Families to Power the Next Wave of Agentic, Physical and Healthcare AI, 16 March 2026, including the official wording for Cosmos 3, GR00T N1.7, the GR00T N2 preview and Alpamayo 1.5
  6. NVIDIA press release, NVIDIA and Global Robotics Leaders Take Physical AI to the Real World, 16 March 2026, including the two-million-plus installed base at FANUC, ABB, YASKAWA and KUKA, Isaac Lab 3.0 early access and Newton 1.0
  7. NVIDIA press release, NVIDIA Releases Vera Rubin DSX AI Factory Reference Design and Omniverse DSX Digital Twin Blueprint, 16 March 2026, including the six DSX components and the partner roster
  8. NVIDIA press release, NVIDIA Expands NVIDIA Agent Toolkit With NVIDIA PhysicsNeMo and CUDA-X Libraries, 26 July 2026, including the PhysicsNeMo re-architecture, cuISS, cuDSS and cuEST, and the Cadence 20x and Keysight 10x claims
  9. Isaac Lab repository release timeline, beta on 17 March 2026, beta 2 on 17 June 2026, Early Access on 16 September 2026, general availability targeted for the end of October 2026, and the Newton repository v1.0.0 tag dated 13 April 2026, github.com
  10. Hugging Face model card nvidia/GR00T-N1.7, open weights from 17 April 2026, three billion parameters, pretrained on 20,854 hours of human egocentric video
  11. NVIDIA blog, National Robotics Week: Latest Physical AI Research, Breakthroughs and Resources, April 2026, including the general availability wording for Newton 1.0, Isaac Sim 6.0 and Isaac Lab 3.0 that is addressed in DISCREPANCY 1
  12. CompaniesMarketCap NVIDIA market capitalisation page with data as of 22 October 2026, showing 5.14 trillion dollars and a share price of about 211 to 213 dollars, together with the Databahn NVIDIA Sales Intelligence Report updated 26 September 2026 as a secondary consolidation of the 10-K and 10-Q
  13. NVIDIA developer blog, Develop Physical AI Reasoning, World, and Action Models with NVIDIA Cosmos 3, including the Mixture-of-Transformers architecture, the Nano 16-billion and Super 64-billion parameter sizes, the benchmark positions and the NIM optimisation options
  14. NVIDIA NGC container listing for cosmos3, with the Hugging Face and GitHub release date of 31 May 2026, the container model list including the Edge and DROID policy variants, and the OpenMDW 1.1 licence
  15. NVIDIA blog, NVIDIA Advances Isaac Platform for General Purpose Robotics with New Reference Humanoid Frameworks, Simulation Pipelines and Data, GTC Berlin, 21 October 2026, including Isaac Lab 3.0 general availability, Isaac OS, the SimReady Central Library of more than 2,000 assets, the Human 1M dataset, the three DGX Station reference workflows and the Agile ONE SN specification
  16. NVIDIA blog, NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models, 21 October 2026, including Cosmos Evaluator 26.10, the up-to-2x Cosmos 3 NIM inference claim, Cosmos Data Factory, Cosmos Framework, Metropolis Blueprint for video search and summarisation 3.3 and the ADNOC deployment
  17. Isaac Lab repository release page for v3.0.0, released 20 October 2026, together with the NVIDIA NGC Isaac Lab container 3.0.0, including the Isaac Sim 6.2, Python 3.12, PyTorch 2.11, CUDA 13.0, Warp 1.17, Newton 1.6.1 and RSL-RL 5.5.1 pins and the quaternion-order change
This report is compiled from public sources for informational purposes only and does not constitute investment advice. Data is current as of 2 November 2026; please refer to company announcements and periodic reports.

Related reports66

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.