Products SEO & GEO Agent Paid Ads Agent Website Agent Conversion Agent
How it works Pricing FAQ Blog Book a call

AI Chip Companies: Who Makes the Chips Behind AI Tools

Updated · 6 min read

Short answer Nvidia still sells most AI chips, but AMD, Google, Amazon and Microsoft now ship their own, mainly to cut what each AI answer costs. You never buy these chips: you rent them through tools like ChatGPT, Gemini or Copilot. Their competition is why the price per AI answer keeps falling, so recheck usage-based AI prices every year.

Want help applying this to your business? Talk it through on a free 30-minute call.

Book a call

Who makes AI chips in 2026?

Six large companies and a few startups. Most AI tools a business uses run on chips from one or more of them, in a data center you never see.

CompanyCurrent AI chips (November 2026)Where you meet them
NvidiaBlackwell GPUs; Vera Rubin, "ramping into full production" as of August 2026 (NVIDIA)The big clouds: the same release names Google Cloud, Microsoft Azure, Oracle Cloud and others
AMDInstinct MI450 series, launched in the third quarter of 2026 (AMD)OpenAI agreed to deploy 6 gigawatts of AMD GPUs, the first 1 gigawatt from the second half of 2026 (AMD)
GoogleTPU Ironwood (7th generation); TPU 8t for training and TPU 8i for answering, due later in 2026 (Google)Gemini; Anthropic (maker of Claude) plans to use up to one million TPUs (Anthropic)
Amazon (AWS)Trainium3, generally available since December 2025 (Amazon)Amazon Bedrock, which already serves production work on Trainium3
MicrosoftMaia 200, announced January 2026 (Microsoft)Microsoft 365 Copilot and OpenAI's GPT-5.2 models on Azure
IntelCrescent Island, a GPU for answering only, with customer samples planned for the second half of 2026 (Intel)Business servers, once it ships
StartupsGroq licensed its technology to Nvidia in December 2025, and its founder joined Nvidia; GroqCloud keeps running (Groq). Cerebras sells AI answers by the token (Cerebras)Developer tools that promise very fast replies

Designing a chip is not the same as making it. Microsoft, for example, says Maia 200 is built on TSMC's 3-nanometer process (Microsoft). TSMC is a Taiwanese contract chipmaker.

What changed since 2025?

Every big name has a new generation, and the focus moved from training (teaching a model) to inference (the model answering you, millions of times a day).

  • Nvidia moved from Blackwell to Vera Rubin.
  • AMD moved past the MI300 and MI350 to the MI450.
  • Google split its TPU line into two chips: one for training, one for inference.
  • Microsoft's Maia 200 is built for inference, not training.
  • Intel's newest data center chip is also inference-only.
  • Groq, once a pure Nvidia rival, now licenses its technology to Nvidia.

How far ahead is Nvidia?

Far, by sales. Nvidia's data center revenue was $89.0 billion in the quarter that ended July 26, 2026, up 117% in a year (NVIDIA). AMD's whole data center segment, which also includes its server processors, was $8.5 billion in its third quarter of 2026, up 97% (AMD). The quarters don't line up exactly, but the gap is roughly tenfold.

Google, Amazon and Microsoft don't sell their chips. They rent them inside their clouds, so no public sales figure compares them with Nvidia.

Why do AI chips change what I pay for AI tools?

Because chips are the factory behind every answer. When chips get cheaper to run, each AI answer costs the tool maker less, and competition pushes some of that saving into lower prices or bigger usage limits.

The cost drop has been steep. Stanford's AI Index found:

"the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024." (Stanford HAI: AI Index 2025)

The same report adds: "At the hardware level, costs have declined by 30% annually, while energy efficiency has improved by 40% each year."

And the chipmakers now sell their new chips on cost per answer, not just speed:

  • Microsoft says Maia 200 has "30% better performance per dollar than the latest generation hardware in our fleet today," and that it serves Microsoft 365 Copilot (Microsoft).
  • Google says its new TPU 8i, the chip for answering, will "deliver 80% better performance-per-dollar compared to the previous generation" (Google).
  • Amazon says customers "are reducing training and inference costs by up to 50% with Trainium" (Amazon).

These are the companies' own claims. What reaches you depends on how the tool charges:

How the tool chargesWhat falling chip costs do for you
Per use (per token, per call, per minute)The price per unit tends to drop with each new model generation
Flat monthly planThe price usually stays; you tend to get more use or a stronger model for it
Annual contractYou are locked in; savings reach you only at renewal

What does this mean for a small business?

Three practical points:

  1. Don't choose a tool by its chip. You can't see which chip answers your request, and it can change without notice. Judge the result and the price.
  2. Recheck usage-based prices once a year. An AI receptionist, chatbot or writing tool billed per use may be overpriced against today's rates. Ask the vendor what changed.
  3. You probably don't need an "AI PC". Microsoft defines a Copilot+ PC by an NPU (a chip for AI on the laptop itself) that can perform "more than 40 trillion operations per second" (Microsoft Learn). That powers Windows features that run on the device. ChatGPT, Gemini and Claude run in the cloud, so any recent laptop with a browser works.

For which AI model fits which marketing job, see AI marketing models.

Example

A cleaning company on a chat widget

It pays a chatbot vendor $0.02 per conversation, set two years ago. The owner asks at renewal, learns the vendor's own model costs fell, and negotiates $0.01. At 3,000 chats a month, that is $30 a month saved. The chip news didn't save money by itself; asking did.

A law firm buying laptops

The office manager is offered Copilot+ laptops at $300 more each. The firm uses ChatGPT in a browser and no on-device Windows AI features. It buys the cheaper laptops and spends the difference elsewhere.

A marketing agency on the API

It runs 5 million tokens a month through an AI model, at $10 per million tokens: $50 a month. A newer model on newer chips does the same work at $4 per million. Switching saves $30 a month, after testing that quality held.

FAQ

Is Nvidia the only AI chip company?

No. AMD, Google, Amazon, Microsoft and Intel all make AI chips, and startups such as Cerebras and Groq sell AI answers on their own hardware. Nvidia sells far more than AMD, its nearest rival in chips for sale.

What is the difference between a GPU and a TPU?

A GPU (graphics processing unit) started as a gaming chip and now does most AI work; Nvidia, AMD and Intel make them. A TPU (tensor processing unit) is Google's own chip, built only for AI and rented through Google Cloud.

Who makes the chips for ChatGPT?

Several companies. OpenAI agreed to deploy AMD's MI450 GPUs from the second half of 2026, and Microsoft says its own Maia 200 serves OpenAI's GPT-5.2 models.

Will AI tools keep getting cheaper?

The cost of each AI answer has fallen fast, and every new chip generation promises more per dollar. Whether a given tool's price falls is the vendor's choice, so compare prices when you renew.

Different for you?

  • You pay for an AI SEO tool or agent: see AI SEO agent pricing for what to compare.
  • You want AI to run tasks, not just answer: see agentic AI; agents make many model calls per task, so per-use prices matter more.
  • You want the work done, not the tools picked: see SEO & GEO.

Keep learning

Ask us one question

One line is enough. We reply by email.

Sources

  1. NVIDIA: Financial results for second quarter fiscal 2027 (August 26, 2026)
  2. AMD: Third quarter 2026 financial results (November 3, 2026)
  3. AMD: AMD and OpenAI announce 6 gigawatt partnership (October 6, 2025)
  4. Google: Our eighth generation TPUs (April 2026)
  5. Anthropic: Expanding our use of Google Cloud TPUs (October 23, 2025)
  6. Amazon: Trainium3 UltraServers now available (December 2, 2025)
  7. Microsoft: Maia 200, the AI accelerator built for inference (January 26, 2026)
  8. Intel: Intel to expand AI accelerator portfolio with new GPU (October 14, 2025)
  9. Groq: Groq and Nvidia enter non-exclusive licensing agreement (December 24, 2025)
  10. Cerebras: Inference pricing
  11. Stanford HAI: AI Index Report 2025
  12. Microsoft Learn: Copilot+ PCs developer guide