NVIDIA · The Techno Optimist

Analyst view

Vera Rubin in production in 2026; cloud partners get it from the second half of the year.
Licensed Groq's inference technology and hired its leaders for about $20B (Dec 2025).
Plans to invest up to $100B in OpenAI, tied to at least 10 GW of NVIDIA systems (Sep 2025).

As reported

Parallelism

Thousands of small cores do the matrix math of AI at once, using low-precision numbers for extra speed.

Scale-up

NVLink switches join 72 GPUs in a rack so a big model can run as if on one giant chip.

Software

Almost 20 years of CUDA software make NVIDIA chips the default target for AI code.

Annual cadence

A new GPU or rack generation arrives about every year: Blackwell, Rubin, then Feynman.

Programs

NVIDIA, three programs

Updated 16d ago

Research · Proto · Pilot · Deploy · Scaling

3 items

Jan 2025
paperCosmos World Foundation Model Platform for Physical AI (opens arxiv.org)
Open models that generate synthetic video to train robots and self-driving cars.
Sep 2022
paperFP8 Formats for Deep Learning (opens arxiv.org)
Proposed the 8-bit number formats that became the standard for fast AI training.
Sep 2019
paperMegatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism (opens arxiv.org)
Showed how to split one model across many GPUs. Most large AI models are trained this way.

Open · Active · Partial

Power and cooling densityActive research
AI racks went from about 40 kW to well over 100 kW. They need liquid cooling and new power designs.
Memory bandwidth wallActive research
Chips compute faster than memory can feed them, so big-model inference often waits on memory.
Export controlsOpen
US limits on China sales cut off a big market and caused a $5.5B charge in 2025.

Company-wide

AI chips · 4

2 suppliers · 2 customers · 2 partners

Supplier
TSMC
Manufactures NVIDIA GPUs and CoWoS packaging.
Supplier
SK hynix
Primary supplier of HBM memory.
Customer
OpenAI
Letter of intent for at least 10 GW of systems; NVIDIA invested $30B in the 2026 round.
Partner
Intel
NVIDIA invested $5B (Sep 2025) to co-develop x86 CPUs with NVLink and RTX SoCs.
Partner
Groq
Non-exclusive licence to Groq inference technology (Dec 2025).
Customer
CoreWeave
Large GPU cloud customer; NVIDIA is also a shareholder.

Key people