Acquired · NVIDIA · Overview
NVIDIA
CUDA as a decade-early bet that became the entire AI moat.
Major events · 1993–2026
- 1993Founding
NVIDIA founded at Denny's, San Jose
Jensen Huang, Chris Malachowsky, Curtis Priem incorporate the company. Initial $200K funding from Sequoia (Don Valentine).
Acquired episode I
- 1995Product
NV1 ships — and fails
First product uses quadratic texture mapping instead of triangles. Microsoft's Direct3D standardises on triangles; NV1 becomes obsolete on arrival.
Acquired episode I
- 1996Crisis
Sega cancels the NV2 contract
Sega had paid $7M for NV2 chip development for the Dreamcast. Mid-development they cancel and move to PowerVR. NVIDIA has months of runway left.
Acquired episode I
- 1997Product
RIVA 128 saves the company
Triangle-based GPU shipped with $80M in sales the first year. Validates Jensen's pivot and the architectural choice. The company survives.
Acquired episode I
- 1999IPO
NVIDIA IPO + GeForce 256 launch
IPO at $12/share on NASDAQ. GeForce 256 is the first product marketed as a 'GPU'; introduces hardware T&L.
Acquired episode I + NVIDIA S-1
- 2006Strategic shift
CUDA launches
General-purpose compute API for NVIDIA GPUs. The bet: someday someone will need parallel compute for non-graphics workloads. Decade-early — nobody believes for years. Becomes the core moat.
Acquired episode II
- 2012Cultural
AlexNet wins ImageNet on GeForce GPUs
Hinton, Krizhevsky, Sutskever train an 8-layer CNN on two GeForce GTX 580s. Crushes ImageNet by 10+ points. The deep-learning era begins on NVIDIA hardware. CUDA bet pays off.
Acquired episode II
- 2016Product
DGX-1 — AI supercomputer in a box
First purpose-built deep-learning system. Jensen personally hand-delivers serial number 0001 to Sam Altman at OpenAI.
Acquired episode II + 2017 OpenAI blog
- 2019Acquired co.
Acquires Mellanox for $6.9B
Buys high-speed networking (InfiniBand). Locks in the data-center bandwidth layer that AI clusters depend on. Becomes the second moat alongside CUDA.
Acquired episode II
- 2022-11Market
ChatGPT releases — every cloud needs more GPUs
OpenAI ships ChatGPT. Hyperscalers panic-buy NVIDIA. The 16-year CUDA bet compounds in 18 months.
Acquired episode III
- 2024-03Product
Blackwell architecture announced at GTC
B100/B200 chips + NVL72 rack-scale system. Jensen markets it not as a chip but as 'one giant GPU' (72 GPUs in a single coherent memory domain).
Acquired episode III + NVIDIA GTC 2024 keynote
- 2024-06Market
NVIDIA briefly becomes world's most valuable company
Market cap touches $3.6T, passing Microsoft and Apple. The CUDA + system bet is fully repriced.
Public market data
- 2026-01Acquired covered
Acquired Part III — The Dawn of the AI Era
Ben and David's third NVIDIA episode crystallises the strategy through the AlexNet → ChatGPT arc.
https://www.acquired.fm/episodes/nvidia-the-dawn-of-the-ai-era
Origin
Programmable graphics accelerators for PC gaming — and, by Jensen's stated bet at the founding, eventually a wedge into general-purpose accelerated computing.
Three engineers from Sun, AMD, and IBM met repeatedly at a Denny's near 280 in San Jose because nobody had an office. The bet: gaming would drive demand for compute that CPUs couldn't deliver, and the same architecture would later serve other parallel workloads. They survived two near-death moments — the NV1 failure in 1995 and the cancelled Sega contract in 1996 — before the RIVA 128 (1997) finally established product-market fit. Jensen has run the company for 33+ consecutive years; founder-CEO continuity at this duration is itself a strategic asset.
Key facts
Annual revenue (FY2025)
~$130B
Market cap peak
~$3.6T
Data center revenue share
~88%
Gross margin
~75%
CUDA developers worldwide
5M+
Jensen Huang founder-CEO tenure
33+ years
Annual R&D spend
~$12B
Blackwell GB200 system price
~$60-70K per GPU
Episodes · 4 covering this company
Hooks from these episodes
Jensen Huang has been NVIDIA's CEO since 1993 — through four near-death pivots, the financial crisis, crypto crashes, and the AI boom. The CUDA bet predates the use case by more than a decade.
Long tenure compounds in ways quarterly thinking can't.