Lesson 3 of 5 · Understanding AI
Lesson 3
The tools and the players
Which tools you can actually use. The labs behind them. The cloud providers and hardware companies underneath. A map of the AI industry, in plain language.
By the end of this lesson, you will:
- Know the major AI tools available to consumers today and which ones are good for which job.
- Understand who builds these tools — the labs, the cloud providers, the hardware companies — and how they relate to one another.
- Be able to read news about "company X invests in AI" or "country Y bans AI service Z" with a clearer mental map of what is actually going on.
The consumer tools
Five categories of tool that anyone with a browser can use, mostly free.
General-purpose chatbots
The "talk to it about anything" assistants. The big three are Claude (from Anthropic), ChatGPT (from OpenAI), and Gemini (from Google). All three have free tiers; paid tiers offer access to more capable models and higher usage limits. They overlap in capability but differ in personality and in particular strengths — Claude is widely considered the strongest writer and reasoner, ChatGPT has the largest user base and the deepest tool ecosystem, Gemini is tightly integrated with Google services.
A smaller serious contender is Grok (from xAI). The strongest Chinese models — DeepSeek, Alibaba's Qwen — are openly available; they are competitive on capability and disrupted Western pricing in 2025.
Coding assistants
For software developers. GitHub Copilot (Microsoft + OpenAI), Cursor, Claude Code, and Windsurf are the main ones in 2026. They sit inside the developer's editor and write, debug, and refactor code as a kind of pair programmer. Most professional software engineers now use at least one.
Image generators
Midjourney, DALL·E (inside ChatGPT), Stable Diffusion (open-source, runs locally if you want), ImageFX (Google). Type a sentence, get an image. The quality has improved enormously between 2022 and today, though hands and text within images remain occasionally awkward.
Video and audio generators
Sora (OpenAI), Veo (Google DeepMind), and Runway generate short video clips from text. ElevenLabs is the leader in voice generation and cloning. Suno and Udio generate music. Video quality in 2026 is good enough for short marketing clips and storyboards; not yet good enough for feature-length film.
Embedded AI in everyday products
Tools you already use have AI in them, often quietly. Microsoft Copilot (built into Office). Google's AI features in Workspace. Adobe's Generative Fill in Photoshop. Slack and Notion AI assistants. iPhone's Apple Intelligence and Android's Gemini integration. Often the most-used AI in any given week is one you did not consciously choose.
Aside · Which one should you use?
If you are starting out and want one tool to learn with, pick Claude.ai or ChatGPT. Both have free tiers that are good enough to begin. Try one; spend a week using it for whatever feels useful (drafting emails, summarising documents, brainstorming, helping with a hobby). Switch to the other if you are curious. The skills transfer between tools — they are more alike than different.
The labs that build the models
The systems above all sit on top of large neural networks called foundation models. There are a small number of labs that build the most capable foundation models. They are the centre of the industry. Knowing who they are makes the news much easier to parse.
OpenAI. Built ChatGPT, launched it in 2022 in what is now seen as the moment that started this era of AI. Based in San Francisco. Closely allied with Microsoft, which has invested heavily and provides much of OpenAI's computing power. Famous people: Sam Altman (CEO).
Anthropic. Founded in 2021 by ex-OpenAI researchers, including Dario and Daniela Amodei. Builds Claude. Particularly focused on AI safety — the question of how to make these systems reliable and honest. Has investments from Amazon and Google. Famous people: Dario Amodei (CEO).
Google DeepMind. The combined research arm of Google's AI work. Built Gemini, AlphaFold, Veo, and many of the most important AI research papers of the last decade. Based across London and Mountain View. Famous people: Demis Hassabis (CEO, won the Nobel Prize in Chemistry in 2024 for AlphaFold).
Meta. Releases open-weight models in the LLaMA family — meaning the model weights themselves are downloadable, allowing other people to run and modify them. This has reshaped the open-source AI ecosystem. Famous people: Yann LeCun (Chief AI Scientist, one of the three pioneers who won the 2018 Turing Award).
xAI. Elon Musk's AI company. Builds Grok. Notable for its association with X (Twitter) and for the rapid construction of one of the world's largest computing clusters ("Colossus") in Memphis.
The Chinese labs. DeepSeek (which startled the industry in early 2025 with strong, cheaply-trained models), Alibaba (Qwen), Baidu (Ernie), and Moonshot AI are the most-watched. Chinese AI is a serious technical force; whether and how it is available outside China depends on shifting regulation in both directions.
Mistral. A French lab, the most prominent European AI company. Builds open-weight models. Important politically and culturally for European AI sovereignty.
The cloud providers
Training and running large AI models requires vast quantities of computing power, far beyond what any single company can build on its own. Three companies own most of that infrastructure.
Microsoft Azure. The largest cloud partner for OpenAI; the most-used cloud for AI in enterprises. Microsoft has invested over $13B in OpenAI to date.
Amazon Web Services (AWS). The largest cloud overall; major investor in Anthropic; serves a huge portion of the AI compute market.
Google Cloud Platform. Powers Google's own models; growing third-party AI business; has its own internal hardware (TPUs).
Almost every AI service you use today runs on one of these three. When you read "company X partners with cloud Y for AI", what is happening is that company X is buying compute from cloud Y to train or serve its models.
The hardware underneath
The cloud providers are themselves dependent on a smaller number of hardware companies. The dominant story of AI hardware in 2026 is one company.
Nvidia. Designs and sells the specialised chips (called GPUs) that almost all AI training uses. The company's flagship chips — H100, H200, the newer B200 and Blackwell-series — are the most coveted hardware in the industry. Nvidia became one of the most valuable companies in the world during 2024 and 2025. Its CEO, Jensen Huang, is now one of the best-known figures in technology.
Other hardware players exist but are smaller. AMD is the closest competitor. Google designs its own TPUs for internal use. AWS has its own Trainium and Inferentia chips. Several start-ups (Cerebras, Groq, Tenstorrent) are building alternatives. China's Huawei is the most significant non-Western player. None of these has come close to displacing Nvidia in 2026, but the competition is intensifying.
How it all fits together
Here is the picture in one paragraph. Hardware companies (mainly Nvidia) make the specialised chips. Cloud providers (Microsoft, Amazon, Google) buy enormous quantities of those chips and assemble them into data centres. AI labs (OpenAI, Anthropic, Google DeepMind, Meta, xAI, the Chinese labs, Mistral) rent that compute and use it to train large models. Those labs sell access to their models via apps (Claude.ai, ChatGPT) and APIs (which other companies use to build their own products). Those other companies — Microsoft Copilot, Adobe, Notion, dozens of start-ups — wrap the underlying models with specific features and sell them on. The end user is you, paying for ChatGPT Plus, or using a Slack AI feature, or having your phone summarise your text messages.
Money flows the other direction: from you, to the application company, to the AI lab, to the cloud, to Nvidia. Most of the eye-watering valuations in AI right now sit somewhere along this chain.
The AI stack in one diagram (text form)
┌──────────────────────────────────────────────┐
│ YOU │
│ (the end user, the buyer) │
└──────────────────────────────────────────────┘
▲ uses
┌──────────────────────────────────────────────┐
│ APPLICATIONS │
│ Claude.ai · ChatGPT · Copilot · Cursor │
│ Notion AI · Adobe Firefly · industry apps │
└──────────────────────────────────────────────┘
▲ built on
┌──────────────────────────────────────────────┐
│ FOUNDATION MODELS │
│ Claude · GPT · Gemini · LLaMA · Grok │
│ DeepSeek · Qwen · Mistral │
└──────────────────────────────────────────────┘
▲ trained on
┌──────────────────────────────────────────────┐
│ CLOUD COMPUTE │
│ Microsoft Azure · AWS · Google Cloud │
└──────────────────────────────────────────────┘
▲ runs on
┌──────────────────────────────────────────────┐
│ HARDWARE │
│ Nvidia (dominant) · AMD · Google TPU │
│ AWS Trainium · Huawei · start-up alternatives│
└──────────────────────────────────────────────┘
Exercise — Try two tools in twenty minutes
- Open Claude.ai and ask it to help with one real task — drafting a message, summarising something you read, brainstorming names for a project, planning a meal for the week. Anything. Spend 10 minutes.
- Open ChatGPT and ask it the same task. Spend 10 minutes.
- Compare. Which one felt more useful? Was the difference about capability, or about personality, or about the interface? There is no wrong answer; this is your first experience of judging tools rather than just hearing about them.
Self-check
- Name the three major general-purpose chatbots and the companies behind them.
- What does a foundation model do, and where in the AI stack does it sit?
- Why is Nvidia so important to the AI industry?
- What is the difference between Meta's models and OpenAI's, in terms of how they are released?
Looking ahead
Lesson 4 looks at where AI is going. The trajectory. What experts agree on, what they disagree on, and the honest uncertainty about how fast and how far. Agents that act on the world. Multi-modal models that see, hear, and reason. The questions about reasoning. The frontier, in plain language.