NVIDIA

"Software is eating the world, but AI is eating software."

NVIDIA is the dominant designer of graphics processing units (GPUs) and the company whose chips have become the essential infrastructure of the AI revolution. Founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem, NVIDIA evolved from a gaming graphics company into the most important semiconductor firm of the AI era — with a market capitalization that has at times exceeded $3 trillion. NVIDIA spans six layers of the agentic economy — a breadth of presence that makes it far more than a chipmaker.

The GPU Monopoly

NVIDIA's dominance in AI compute is extraordinary. Its H100 and successor Blackwell GPUs power the vast majority of large language model training worldwide. The CUDA software ecosystem — NVIDIA's proprietary parallel computing platform — has created a moat that competitors like AMD and Intel struggle to breach, because decades of AI research tooling has been built on CUDA. Every major AI lab — OpenAI, Anthropic, Google DeepMind, Meta — depends on NVIDIA silicon.

Full-Stack AI Platform

NVIDIA has been quietly building upward through the entire technology stack, transforming itself from a chip company into a full-stack AI platform. NeMo — NVIDIA's agent development toolkit, along with the NeMo Claw open-source agent platform announced at GTC 2026 — positions them in agent creation and orchestration. NIM microservices provide optimized inference deployment infrastructure, making it easy to serve AI models at scale. Nemotron, NVIDIA's family of open models optimized for agentic AI, gives NVIDIA its own foundation models. And DGX Cloud offers managed compute for training and inference. The strategic logic is clear: if you make the silicon that everything runs on, you have a natural advantage in optimizing every layer above it.

In 2025, NVIDIA signaled how seriously it takes the foundation model layer by committing $26 billion to training its own open-weight AI models — a figure first disclosed in financial filings and reported by Sherwood News. The investment dwarfs what most standalone AI labs spend on training and represents a strategic inflection point: NVIDIA is no longer content to supply the picks and shovels — it intends to mine the gold itself. Open-weight distribution (rather than closed APIs) fits their platform logic: widely adopted NVIDIA-trained models create downstream demand for NVIDIA inference hardware, reinforcing the same flywheel that CUDA established for training.

AI Infrastructure Provider

NVIDIA has expanded from chip design into a full AI infrastructure stack: DGX systems for data center AI, networking (InfiniBand and NVLink), software frameworks (TensorRT, NEMO), and cloud partnerships with every major cloud provider. The company's inference platform is becoming critical as AI shifts from training to deployment — making NVIDIA central to the economics of the agentic web. The Vera Rubin platform, the generation after Blackwell, is now in full production, continuing the cadence of exponential performance improvements.

3D and the Metaverse

NVIDIA's Omniverse platform is a USD-based (Universal Scene Description) collaboration and simulation platform for 3D content and digital twins. Combined with RTX ray-tracing technology that brought real-time cinematic rendering to consumer GPUs, NVIDIA provides both the real-time rendering pipeline and the industrial simulation infrastructure for the metaverse.

Compute Capital

NVIDIA GPUs have become a form of capital — access to H100 clusters determines who can train frontier AI models. This dynamic has spawned an entire industry of GPU cloud providers like CoreWeave and shaped the compute capital markets that Jon Radoff has analyzed. The $26 billion foundation-model commitment raises a deeper question: when the infrastructure monopolist starts training its own frontier models on its own hardware, it competes directly with the customers who made it dominant. Whether the market fully accepts NVIDIA as a software platform company — not just a chipmaker — is one of the most consequential bets in the agentic economy.

2026 Developments

Q2 fiscal 2027 results

For its fiscal second quarter of 2027, which ended July 26, 2026, NVIDIA reported revenue of $96.2 billion, up 106% from a year earlier. Data center revenue was $89.0 billion, up 117%. The company guided to revenue of $108.0 billion, plus or minus 2%, for the third quarter. Its purchase commitments rose from $119 billion to $279 billion, a sign of how much supply it is locking in for the next wave of demand. NVIDIA said Vera Rubin is "now in full production." These results sit at the center of the debate over AI capex and a possible bubble: they show the money is real, but one company's revenue is another's capital spending.

China

China has largely dropped out of NVIDIA's numbers. Hopper shipments to China accounted for less than 1% of data center revenue in the quarter, and the company's outlook assumes no China data center compute revenue at all. The growth is coming almost entirely from elsewhere.

Nemotron 3 Ultra

NVIDIA's own model effort produced a notable release on June 4, 2026: Nemotron 3 Ultra, a hybrid Mamba-Transformer mixture-of-experts model with about 550 billion parameters, roughly 55 billion of them active per token, released with open weights. According to RuntimeWire, it scored 47.7 on the Artificial Analysis Intelligence Index, making it the top US open-weight model on that measure at release.

Isaac GR00T Reference Humanoid

On June 1, 2026, NVIDIA announced the Isaac GR00T Reference Humanoid for academic research labs. It combines a Unitree H2 Plus body, Sharpa tactile hands and a Jetson Thor computer, with 75 degrees of freedom. The goal is to give university labs a common, NVIDIA-powered platform for humanoid robot research, extending the Isaac stack from simulation into standard hardware and pulling robotics research further toward physical AI built on NVIDIA.