NVIDIA

Founded April 5 by Jensen Huang (ex-LSI Logic, AMD), Chris Malachowsky (Sun) and Curtis Priem (IBM, Sun) at a Denny's in San Jose with $40K; named from Latin 'invidia', initial focus on PC graphics accelerators.

Released NV1, NVIDIA’s first chip on Sega Saturn-style quadrilateral rendering; commercial flop that nearly bankrupted the company before SGS-Thomson and Sega rescue funding.

Released RIVA 128, the company’s first commercially successful 3D graphics chip; sold 1M units in four months and saved NVIDIA from the brink of insolvency.

IPO on NASDAQ January 22 at $12/share, raising $42M; small offering reflected modest revenue (~$160M) but funded the Riva and TNT2 graphics generations.

Announced GeForce 256 August 31 and shipped October 11; coined the term 'GPU' (Graphics Processing Unit) by integrating hardware transform-and-lighting on a single chip with 23M transistors.

Acquired struggling rival 3dfx Interactive for $112M after the Voodoo card maker filed bankruptcy; NVIDIA absorbed engineering talent and IP that powered the GeForce3.

GeForce3 launched with first programmable vertex and pixel shaders, the foundational architecture for every modern GPU; Microsoft selected the chip for the original Xbox console.

Unveiled GeForce 8800 GTX (G80) with unified shaders and first CUDA-capable architecture; laid the groundwork for general-purpose GPU computing and modern AI training.

Released CUDA 1.0 SDK February 15 for Windows and Linux; opened the GPU as a programmable parallel processor and seeded the scientific-computing and deep-learning communities.

Launched Tesla brand for HPC GPUs (Tesla C870/C1060) priced at $1,500-$1,700; first credible move beyond gaming into the data-center compute market.

AlexNet won ImageNet using two NVIDIA GTX 580 GPUs, dropping image-classification error rates from 26% to 15%; the watershed moment validated GPUs as the substrate of modern deep learning.

Tesla K80 launched, GK210 Kepler-based dual-GPU accelerator targeting deep learning workloads; the chip became the workhorse of early academic and hyperscaler ML training.

Tesla P100 unveiled at GTC 2016, first Pascal data-center GPU with HBM2 memory and NVLink interconnect; targeted deep-learning training and HPC at $5,000+ per GPU.

DGX-1 announced at GTC 2016 as first purpose-built AI supercomputer-in-a-box; eight P100 GPUs at $129K, with first system hand-delivered by Jensen Huang to OpenAI.

Tesla V100 with Volta architecture announced at GTC, introducing Tensor Cores for AI acceleration at 125 TFLOPS FP16; powered the GPT-2/GPT-3 training era at OpenAI and Microsoft.

Announced $6.9B acquisition of Mellanox to bolster networking for data-center AI fabrics; preempted Intel and Microsoft bids and locked in InfiniBand for next-gen GPU clusters.

Closed $7B Mellanox acquisition after Chinese antitrust clearance, gaining InfiniBand and Ethernet IP that became the spine of every NVIDIA reference AI cluster from DGX-2 onward.

A100 launched at GTC 2020 on TSMC 7nm with 40GB HBM2e and 19.5 TFLOPS FP32; introduced multi-instance GPU (MIG) and became the dominant LLM training chip through 2022.

Announced $40B agreement to acquire Arm Holdings from SoftBank; the deal would have given NVIDIA control of the world's most-used CPU instruction set but drew immediate global regulatory scrutiny.

Abandoned $40B Arm acquisition due to FTC and global regulatory blockades; SoftBank kept the $1.25B prepayment and prepared Arm for its 2023 IPO.

Hopper H100 announced at GTC March 22 — 80B transistors on TSMC 4N, 80GB HBM3 and Transformer Engine for FP8; became the de-facto LLM training chip and pushed lead times to 11+ months by mid-2023.

H100 entered full production with hyperscaler partner systems shipping in early Q4; Microsoft, AWS, Google and Oracle placed initial orders for 100K+ GPUs each.

US BIS export controls of October 7 banned A100/H100 sales to China; NVIDIA pivoted to A800/H800 China-specific SKUs with reduced interconnect bandwidth, preserving most China revenue.

OpenAI launched ChatGPT November 30, igniting a global AI compute demand surge that Jensen Huang called AI's 'iPhone moment'; data-center orders surged from <$4B/quarter to $14B+ within four quarters.

DGX H100 systems began shipping to enterprise customers, anchoring the first wave of post-ChatGPT AI buildouts; Microsoft Azure and Oracle Cloud took the first racks for OpenAI workloads.

Crossed $1T market cap on May 30 on a record AI revenue forecast (Q1 FY24 guide of $11B vs $7.2B consensus); joined Apple, Microsoft, Alphabet and Aramco in the trillion-dollar club.

October 17 BIS update banned A800/H800 China-specific chips and added performance-density rule covering future SKUs; NVIDIA began designing H20/L20 for the China market under the new caps.

H200 announced at SC23, first GPU with HBM3e memory at 141GB and 4.8 TB/s bandwidth; targeted Q2 2024 ship date and locked in SK Hynix as lead HBM3e supplier.

Surpassed $2T market cap intraday February 23, becoming the third US firm to close above the threshold; took just nine months from $1T to $2T, the fastest such doubling in market history.

Blackwell B200 and GB200 superchip unveiled at GTC March 18, 208B transistors with up to 20 PFLOPS FP4 per GPU; built on TSMC 4NP and used CoWoS-L packaging with 192GB HBM3e.

GB200 NVL72 announced at GTC, liquid-cooled rack of 36 Grace Blackwell superchips for trillion-parameter LLMs; 1.4 exaFLOPS FP4 compute per rack, priced at ~$3M each.

Surpassed $3T market cap June 5 and briefly became the world's most valuable company on June 18 at ~$3.34T, displacing Microsoft and Apple atop the global leaderboard.

GB200 NVL72 systems began shipping to AWS, Microsoft, Google and Oracle after multi-quarter ramp delays; CoWoS-L packaging yields finally crossed acceptable threshold in Q4.

Blackwell Ultra B300 unveiled at GTC March 18, 1.5x B200 performance with 288GB HBM3e and 15 PFLOPS dense FP4; targeted as bridge product before Rubin platform in 2026.

Took $5.5B Q1 charge on H20 China inventory after April BIS rule blocking last China-tailored Hopper SKU; Jensen Huang publicly criticized the rule as an own-goal helping Huawei.

Became first $4T public company on July 9, capping a 10x rally since ChatGPT launch; market cap doubled in 13 months from the $2T close in February 2024.

US Trump administration permitted resumed sales of H20 to China, reversing April ban after diplomatic deal; NVIDIA recovered partial inventory write-down via fresh shipments.

Crossed $5T market cap intraday October 29, the first company ever to reach that milestone; just 3.5 months after the $4T close, fastest such trillion-dollar leg in history.

Amazon EC2 P6-B300 instances with Blackwell Ultra became generally available, anchoring AWS AI fleet upgrades; instance prices started around $98/hour for 8x B300 nodes.

Began Rubin GPU production at TSMC on N3 node ahead of GTC 2026 mass announcement; SK Hynix HBM4 stacks shipped from September were paired in CoWoS-L packaging.

Vera Rubin platform announced at CES, Vera CPU + dual Rubin GPU in one package with NVLink 6 and ConnectX-9; positioned as full vertically-integrated AI factory replacement for Blackwell.