The giants behind the scenes

The labs get the headlines. The companies below own the buildings, the chips, the phones and the office software, and between them they decide what most people actually touch. Dated September 2026, because these relationships move.

Microsoft

Runs Azure, the second biggest cloud, and has been OpenAI's main backer since 2019. After OpenAI reorganised in late 2025, Microsoft ended up owning about 27 percent of it. In April 2026 the two loosened the deal: OpenAI can now sell through other clouds, and Microsoft's licence to OpenAI's models runs to 2032 but is no longer exclusive. Microsoft reported about $24 billion of revenue from OpenAI in its 2026 financial year.

What you see: Copilot, OpenAI's models inside Windows, Office and Edge. If your workplace "has AI", it is often this.

Amazon

Runs AWS, the biggest cloud, and makes its own AI chips, Trainium. It does not train a headline model. Instead it has become Anthropic's largest backer: about $8 billion invested by 2025, then in April 2026 an agreement to put in up to $25 billion more, with Anthropic committing to spend over $100 billion on AWS over ten years and securing up to five gigawatts of capacity for Claude.

What you see: Claude and other models available through Amazon Bedrock for businesses, and Alexa slowly becoming something more capable.

Google

The only company on every layer. Makes the TPU chips (and sold Anthropic access to up to a million of them), runs Google Cloud, trains Gemini through Google DeepMind, and ships Gemini inside Search, Gmail, Docs and Android. Gemini's app passed 900 million monthly users in 2026.

What you see: Gemini everywhere Google is, AI Overviews at the top of search results, and, since 2026, inside Siri.

Apple

Makes no frontier model and, for a long time, no clear AI story. In January 2026 it signed a multi-year deal with Google, reported at about a billion dollars a year, for custom Gemini-based models to run the new Siri. The models were unveiled at WWDC in June 2026 and roll out through the year. Apple's angle is privacy: as much as possible runs on the device, and what leaves it goes through Apple's own servers.

What you see: Apple Intelligence and a Siri that can finally act across your apps, with ChatGPT as a secondary option for harder questions.

Meta

Trains Llama and gives the weights away, which no other big lab does at the frontier. Llama 4's fully open release landed in August 2026. Meta's reasons are its own (it does not sell a cloud or an API), but the effect is that anyone can run a strong model on their own machines, which keeps prices honest across the industry.

What you see: Meta AI inside WhatsApp, Instagram and Messenger, and Llama underneath a surprising number of other companies' products.

Nvidia

Makes the chips. Roughly 80 percent of the data-centre AI accelerators sold in 2026 are Nvidia's, down from about 92 percent in 2023 as the clouds build their own, but the market roughly quadrupled in that time, so the smaller slice is a far bigger pie. In September 2026 Nvidia agreed to buy Hugging Face, the site where most open models are published, for about $13 billion.

What you see: nothing directly, and everything indirectly. Every answer you get from every chatbot ran on a Nvidia chip or one built to compete with it.

The chip challengers, briefly

AMD sells a rival line of GPUs and holds a single-digit share, with a new generation (MI400) due by the end of 2026. Broadcom helps Google and others design custom chips. Google's TPUs and Amazon's Trainium are the real threat to Nvidia, because they let the two biggest buyers stop paying Nvidia's prices.

What to remember

• Microsoft owns a quarter of OpenAI and ships it as Copilot. Amazon backs Anthropic and rents Claude to businesses. Google does everything itself.
• Apple's Siri now runs on Gemini. Meta gives Llama away.
• Nvidia sells the shovels, and in 2026 bought the town square where open models are shared.
• None of this changes what you build here. It changes who is standing behind the app you paste it into.