Preparing your workspace
Jobbit
Guides10 min read

What Does Nvidia Do? AI Chips Explained in 9 Plain Answers

What does Nvidia do? A plain-English guide to the chips behind ChatGPT, Claude and Gemini: who buys them, what they cost and why they price your AI tools.

What Does Nvidia Do? AI Chips Explained in 9 Plain Answers
Read in:

Nvidia designs the chips that train and run almost every well-known AI system, including the ones behind ChatGPT, Claude and Gemini. It began in 1993 making graphics chips for PC games. In its last full financial year it sold $215.9 billion of products, and about 90% of that came from data centres. This guide explains what the chips do, who buys them and why they decide what your AI tools cost.

You run a business, not a trading desk, yet the same name keeps appearing in the news. In November 2026 it was there again: Nvidia's long-time chief executive resigned, the shares fell for a sixth week running and the company dropped below a $4 trillion valuation for the first time since July 2025. It is fair to wonder why a chip maker in California should matter to a cafe owner in Leeds or a plumber in Bristol. What does Nvidia do, and why does it reach into your monthly subscriptions? The short version: every answer an AI tool gives you is worked out on hardware, and most of that hardware is Nvidia's.

Below: how we checked this guide, the company in one table, 9 plain answers, the rivals, the link to the prices you pay and what a small business should take from it. Nothing here is investment advice.

You do not need to follow chip news to use AI well. Jobbit shows what each task costs in credits and lets you choose the model. Free plan, 100 credits a day, no card.

How we checked this guide

  1. Nvidia's own documents, read on 14 November 2026. Its annual report for the year to 25 January 2026, its quarterly results releases and its company timeline.
  2. Other companies' own releases for what they spend: Amazon, Meta, Google, Microsoft and Anthropic.
  3. Independent bodies for context. The International Energy Agency on electricity, Stanford's AI Index on costs, IBM for a neutral definition of the chip.
  4. Chip prices are analysts' estimates. Nvidia does not publish a price list, and we say so wherever a price appears.
  5. Timing. Nvidia's next results are due on 17 November 2026, after this was written, so the figures here stop at the quarter that ended in July.

Nvidia at a glance

QuestionAnswerSource
Founded5 April 1993, by Jensen Huang, Chris Malachowsky and Curtis PriemNvidia timeline
BasedSanta Clara, CaliforniaAnnual report
StaffAbout 42,000 in 38 countriesAnnual report
First business3D graphics for games and multimediaNvidia timeline
Sales, year to January 2026$215.9 billion, up 65%Results release
From data centres$193.7 billion, about 90%Results release, our arithmetic
Sales, quarter to July 2026$96.2 billion, up 106% on a year earlierResults release
Gross margin, same quarter75%Results release
Current chip familyBlackwellResults release
Next chip familyVera Rubin, in production since summer 2026Results release
Next results17 November 2026Nvidia announcement
Chief executiveColette Kress, interim, since 12 November 2026Company announcement

The 9 plain answers

1. What does Nvidia actually make?

Chips, and the computers and software built around them. The chip is called a graphics processing unit, or GPU. IBM defines it as "an electronic circuit designed to speed computer graphics and image processing". Nvidia's own timeline says it invented the GPU in 1999. For years that meant better-looking video games.

2. Why is a games chip good at AI?

Drawing a game means working out the colour of millions of dots at the same instant. A normal processor works through tasks mostly one after another. A GPU, in IBM's words, has "hundreds, even thousands, of cores" doing sums side by side. Training an AI model turns out to be the same kind of job: an enormous number of small sums that can all run at once. Picture one brilliant accountant against a hall of ten thousand people with calculators. For this work, the hall wins.

3. What is CUDA, and why do people say it matters?

CUDA is Nvidia's software layer, unveiled in 2006, that lets programmers use its chips for sums that have nothing to do with graphics. Researchers spent the following years writing their AI tools on top of it. That history is a large part of Nvidia's advantage: a rival has to match the chip and also persuade people to move years of software.

4. Does Nvidia make ChatGPT?

No. Nvidia does not make the assistants. OpenAI makes ChatGPT, Anthropic makes Claude, Google makes Gemini. Nvidia sells the hardware they run on, a little like the firm that sells ovens to every bakery in town. If you want the difference between a chatbot and an agent, see what an AI agent is.

5. Who buys the chips?

A small number of very large customers. Nvidia's annual report says its sales are concentrated among customers who buy directly or indirectly. They are mostly the cloud companies and the AI labs.

  • Amazon says it expects to spend about $220 billion on capital projects in 2026.
  • Meta expects $135 billion to $145 billion, and Nvidia says its partnership with Meta covers "millions" of chips.
  • Google told investors its spending will "increase significantly" in 2027, the Associated Press reported.
  • OpenAI signed a letter of intent with Nvidia in September 2025 for at least 10 gigawatts of systems, with Nvidia to invest up to $100 billion as they are deployed.

6. What does one of these chips cost?

Nvidia publishes no prices. Analysts' estimates reported by Tom's Hardware put a top Blackwell chip at up to about $70,000 and a full server rack at about $3 million or more. Treat those as estimates. They explain the gross margin of 75% and why only the largest companies buy in bulk.

7. How big is Nvidia?

By sales, $215.9 billion in its last full year. By stock market value, it was the first company to be valued at $5 trillion, in October 2025, and CNBC reported on 13 November 2026 that it had closed below $4 trillion for the first time since July 2025. A share price is an opinion about the future, and opinions about AI are swinging widely this month.

8. Why can Nvidia not sell freely to China?

The United States restricts exports of advanced AI chips to China on national security grounds. The rules began in October 2022 and have been tightened and loosened several times since. In December 2025 the president said one older chip could be sold there with a 25% fee, while the newest families stayed excluded, Reuters reported. Nvidia's own forecast for the latest quarter assumes no data centre chip sales to China at all.

9. Why is Nvidia in the news in November 2026?

Three reasons. Its chief executive since 1993, Jensen Huang, resigned on 12 November 2026 and the finance chief, Colette Kress, became interim chief executive. Its shares fell about 13% in that week, a sixth weekly fall in a row, CNBC reported, with investors already worried about how much debt the AI build-out needs. And its quarterly results arrive on 17 November, with analysts expecting about $108.9 billion of sales.

The rivals, in one line each

  • Google makes its own AI chips, called TPUs. Anthropic plans to use up to one million of them.
  • Amazon makes Trainium. Its chief executive said in December 2025 that the second version was already a multi-billion-dollar business with more than a million chips in production.
  • Microsoft announced Maia 200 in January 2026, claiming 30% better performance per dollar than the newest hardware in its own fleet.
  • AMD sells competing GPUs and agreed a large supply deal with OpenAI in October 2025, according to press reports.
  • Huawei is building China's alternative, the Ascend family, with a published roadmap through 2026.

Anthropic is a useful example of where this is heading. It says it runs on three platforms: Google's chips, Amazon's chips and Nvidia's. The biggest buyers do not want to depend on one supplier, for the same reason you should not.

Why chips decide what your AI tools cost

No study measures an exact pass-through from chip prices to your subscription. The chain of cause and effect is still easy to follow.

  • Every answer costs computing time. When you send a message, a chip somewhere does the work. A long document or a video takes far more than a short question.
  • Computing time is scarce. In September 2026 OpenAI paused new sign-ups to its $200 a month plan because demand for a new model outran its capacity, Fortune reported. Google told investors in October that usage was growing "even with capacity constraints".
  • So providers ration. Message caps, slower responses at busy times, and the best models reserved for higher plans are all ways of sharing out limited chips. Our guide to what free AI plans allow lists the current limits.
  • The cost of the same quality falls fast. Stanford's AI Index found the cost of running a model at a fixed level of ability fell more than 280 times between November 2022 and October 2024. Nvidia claims its next chips cut the cost per answer by up to ten times.
  • The newest model is always the expensive one. Prices fall for last year's ability and stay high for this year's.
  • Electricity sits underneath it all. The International Energy Agency says data centres used about 415 terawatt hours in 2024, roughly 1.5% of the world's electricity, and expects that to more than double by 2030.

What a small business should take from this

  1. Use the smallest model that does the job. A summary of an email does not need the most powerful model. Smaller models cost less and answer faster.
  2. Expect limits to move. Caps and prices change when chip supply changes. Keep a second tool you can switch to.
  3. Watch what a task costs. Tools that show usage per task let you spot waste. Our pricing guide explains how credits work.
  4. Do not buy on headlines. A falling share price says nothing about whether a tool is useful to you on Monday morning.
  5. Learn the vocabulary once. Our AI glossary covers tokens, models and the rest in plain English.

How an AI agent hides the hardware from you

ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse all run on chips like these, and so does every other AI product. What differs is how much of the machinery you have to think about.

  • A choice of models. Jobbit's paid plans let you pick between models from several providers, with prices shown, so a light task can use a light model. See the plans.
  • A visible meter. Credits show what each task used.
  • A low entry price. Go costs £7 a month with 2,000 credits and hosting for one small app on your own domain.
  • Finished work, not answers. The agent builds the site, the document or the spreadsheet, and a human professional can take over through the network when the job needs one.

Frequently asked questions

What does Nvidia do in simple terms?

Nvidia designs computer chips called GPUs and the systems around them. They were made for video game graphics and turned out to be ideal for AI, because they do huge numbers of small sums at once. Most well-known AI systems are trained and run on them.

Why is Nvidia so valuable?

Because nearly every company building AI needs its chips, supply is limited and its profit on each sale is high: a gross margin of 75% in the quarter to July 2026. Its software, CUDA, also makes customers slow to switch. Whether the valuation is justified is a matter of opinion and not something this guide judges.

Does ChatGPT use Nvidia chips?

Yes, among others. OpenAI signed a letter of intent with Nvidia in 2025 for at least 10 gigawatts of systems, and has also announced deals with other chip makers. Most large AI labs now use more than one supplier.

Who are Nvidia's competitors?

AMD sells rival GPUs. Google, Amazon and Microsoft each design their own AI chips for their own data centres. Huawei leads China's effort. None has yet matched Nvidia's share of the market for training the largest models.

Who is the chief executive of Nvidia now?

Colette Kress, the finance chief, became interim chief executive on 12 November 2026 after Jensen Huang, who co-founded the company in 1993, resigned. The board has said it is searching for a permanent successor.

Will AI get cheaper or more expensive?

Both. The cost of a fixed level of ability has fallen very fast, more than 280 times in two years by Stanford's measure. The newest and most capable models stay expensive and are often rationed. The practical answer is to match the model to the task. Try Jobbit free with 100 credits a day and no card.

Related guides