Unboxing the AI Stack
Organization · L08 · Chips & Memory

NVIDIA

NVIDIA is a US chip designer founded on April 5, 1993, by Jensen Huang, Chris Malachowsky and Curtis Priem, with the aim of bringing 3D graphics to gaming and multimedia.1 Its SEC filings give its business address as Santa Clara, California.2 The company’s own timeline credits it with inventing the graphics processing unit (GPU) in 1999 and dates its CUDA parallel-computing architecture to 2006.1 On MetaPrompt it appears mostly as the maker of data-center accelerators: the Hopper-generation H200, the Blackwell generation it announced on March 18, 2024 as Hopper’s successor,3 and the later Blackwell Ultra and Rubin.

In AI Accelerators With the Most HBM per Chip (October 2026), NVIDIA’s Rubin GPU shares #1 with AMD’s MI350X and MI355X at 288 GB, and it is the only shipping part in that table with HBM4: NVIDIA’s HGX page lists 288 GB of HBM4 at 22 TB/s per GPU in HGX Rubin NVL8, and says Vera Rubin is ramping into full production.4 The older H200 carries 141 GB of HBM3E at 4.8 TB/s.5 The H200 is also the chip that the US export rule of January 15, 2026 names, alongside AMD’s MI325X, as an example of what can now get case-by-case license review for China and Macau, as How US AI Chip Export Controls Work explains.6 In Stargate’s Sites: Operating vs Under Construction vs Planned, Crusoe says Oracle began delivering the first NVIDIA GB200 racks to the Abilene campus in June 2025.7 And in The Memory Crunch, by the Numbers, TrendForce names NVIDIA among the AI chip suppliers that high memory costs give “greater justification for raising product prices.”8

NVIDIA also publishes open models. Nemotron 3 Ultra, a hybrid Transformer-Mamba mixture-of-experts model with 55B active parameters out of 550B, is served on OpenRouter9 and appears in Most-Used AI Models on OpenRouter by Weekly Tokens. GR00T N1, a foundation model for humanoid robots first posted on March 18, 2025, is one of the designs compared in How Vision-Language-Action Robot Models Work.10 Parakeet TDT 0.6B v2, a 600-million-parameter English speech-recognition model, is among the systems in What Word Error Rate Measures, and Why Speech-to-Text WERs Disagree.11

Illustration: the ten layers of the AI stack, with Chips & Memory, where NVIDIA’s GPUs sit, pulled out.

Key facts

FoundedApril 5, 19931
FoundersJensen Huang, Chris Malachowsky, Curtis Priem1
HeadquartersSanta Clara, California, US2
Data-center GPU generations coveredHopper (H200), Blackwell, Blackwell Ultra, Rubin34
Most HBM per shipping GPU288 GB HBM4 (Rubin) Vendor-stated4
Rubin memory bandwidth (HGX Rubin NVL8)22 TB/s per GPU Vendor-stated4
H200 memory141 GB HBM3E, 4.8 TB/s Vendor-stated5

As of . How figures are labeled.

Appears in

Sources

  1. NVIDIA, Corporate Timeline, retrieved 2026-10-13.
  2. US Securities and Exchange Commission (EDGAR), NVIDIA Corp company record (business address), retrieved 2026-10-13.
  3. NVIDIA Newsroom, “NVIDIA Blackwell Platform Arrives to Power a New Era of Computing” (March 18, 2024), retrieved 2026-10-13.
  4. NVIDIA, NVIDIA HGX Platform (HGX Rubin NVL8, HGX B300 and HGX B200 specifications), retrieved 2026-10-13.
  5. NVIDIA, NVIDIA H200 GPU, retrieved 2026-10-13.
  6. Bureau of Industry and Security, “Revision to License Review Policy for Advanced Computing Commodities,” final rule, 91 FR 1684, January 15, 2026, govinfo PDF, retrieved 2026-10-13.
  7. Crusoe, “Crusoe Announces Flagship Abilene Data Center is Live” (September 30, 2025), retrieved 2026-10-13.
  8. TrendForce, “Memory Prices Soar; DRAM and NAND Flash to Account for 68% of Major CSP CapEx in 2027, Says TrendForce”, press release, August 25, 2026, retrieved 2026-10-13.
  9. OpenRouter, NVIDIA: Nemotron 3 Ultra (free), retrieved 2026-10-13.
  10. NVIDIA, GR00T N1: An Open Foundation Model for Generalist Humanoid Robots, arXiv 2503.14734 (v1 March 18, 2025), retrieved 2026-10-13.
  11. Hugging Face (NVIDIA), nvidia/parakeet-tdt-0.6b-v2 model card, retrieved 2026-10-13.