The AI companies you should know, and why

Most people can name two or three AI companies. That is not enough to understand what is happening to work, because the companies that matter most sit in layers most people never see. This page is a map of those layers and the companies that define each one. It says what each company is, what it is betting on, and what that bet tells you about which human skills hold their value. It avoids the numbers that make pages like this go stale, because the shape of the industry changes far more slowly than the headlines about it.

How to read the map#

Think of AI as five stacked layers.

At the bottom is the substrate: the firms that make chips, memory, the machines that make chips, and the power and cooling that keep them running. Above that are the landlords: clouds and data centre builders who rent that capacity out. Above them are the labs that train the models. Above the labs are the wrappers: products that turn a model into a tool for one job. And at the top are the converted: ordinary companies, from carmakers to banks, that have rebuilt themselves around all of it.

Money flows down the stack. Every pound a lab raises ends up with a chipmaker, a foundry or a power company. That is why the most reliably profitable AI businesses have tended to be the ones nobody has heard of.

Two more things hold across the whole map. First, the direction of travel at every layer is from answering to doing: from a model that replies to a question, to an agent that operates a computer, writes the code, takes the call or places the order. Second, wherever that shift lands, the scarce human input becomes the same three things: knowing what to ask for, judging what comes back, and taking responsibility for it.

The labs#

The labs train the frontier models. There are only a handful, because training at the frontier costs more than almost any company can afford, and each has a distinct bet.

OpenAI made generative AI a mass phenomenon with ChatGPT and remains the reference point for consumer AI. Its bet is that scale wins: more compute, more capital, more users, and a product that moves steadily from answering questions to carrying out tasks. It is also the lab that talks most openly about building general intelligence. If OpenAI is right, the premium in human work moves from doing a task to specifying it and owning the outcome.

Anthropic was founded by former OpenAI researchers and makes Claude. It built its business by selling to companies and developers rather than consumers, with a particular strength in coding and long, careful knowledge work, and it positions itself as the lab willing to hold a model back when it judges the risk too high. Its bet is on AI as a trusted colleague rather than a consumer toy, which rewards judgement, domain context and the ability to supervise.

Google DeepMind is Alphabet's merged research and product arm and the maker of Gemini. Its parent is the only lab owner with its own chips, its own distribution to billions of people through Search, Android and Workspace, and a science lab that has won a Nobel prize. Its bet is ambient AI: intelligence built into tools people already use. The skill that demands is knowing when to trust the default answer and when to go deeper.

Meta was the champion of open models through Llama, then moved its flagship models behind closed doors and reorganised its research around a superintelligence lab. It has since released smaller models openly again while keeping the flagship closed. Its bet is that owning the model matters more than sharing it. The lesson for everyone else is that free access to frontier models can never be assumed, so building on a single provider is a risk to hedge.

Microsoft is OpenAI's largest backer and the company that made AI a default feature of office work through Copilot. Its relationship with OpenAI has loosened over time, and it now trains models of its own. Its bet is distribution: AI inside the tools most people already use. That makes fluency with an assistant a baseline expectation rather than a differentiator.

xAI, Elon Musk's lab and the maker of Grok, was folded into SpaceX and sits alongside X, the social network. It is the most aggressively capitalised lab outside the top two and has been the most controversial, because Grok has repeatedly shipped with thinner guardrails than its rivals. It is a live demonstration of what a powerful model looks like without them, and a reason why verifying what you see is now a basic workplace skill.

Mistral, in Paris, is Europe's frontier lab. Its bet is sovereignty: governments and regulated industries will pay for a model they can host in their own jurisdiction. It makes procurement, compliance and vendor judgement core AI skills for any European organisation.

Beneath these sit a shifting set of research labs founded by people who left the big ones: some ship no product at all, some release open models, some are built around a single founder's theory of how intelligence should be trained. Their names change. Their role does not: they are where the next surprise comes from, and they are why no position on this map is permanent.

The wrappers#

The labs sell models. A separate kind of company grows by wrapping those models into a tool for one job, and these are the companies most people actually use at work.

Coding was the first profession where AI moved from assistant to primary tool, and Cursor is the company that proved it, growing faster than almost any software business before it and then being bought by SpaceX in the summer of 2026. Lovable and Replit took the next step: letting people who cannot code build software by describing it. Software creation is moving to people who never learned to program.

Manus is a general-purpose agent that runs multi-step tasks on its own. It matters less for its product than for its story: founded in Beijing, moved to Singapore, agreed to sell to Meta, blocked by Beijing, and set loose again as an independent company. It is the clearest case of governments treating an AI agent company as a strategic asset.

Perplexity is an answer engine that takes on Google search directly, and its browser agent has become the test case for whether an AI may act on your behalf inside another company's website. That question is being fought in court, round by round, and the answer will shape what agents are allowed to do for you.

Voice, video and images have their own leaders: ElevenLabs for voices and agents that handle calls, Synthesia for avatar video used in corporate training, Runway for generated video, and Midjourney for images and for the copyright fights that follow. In professional services, Harvey is the legal AI that put the billable hour under pressure, and OpenEvidence is the reference tool doctors adopted faster than hospitals could govern it. Character.AI, the companion chatbot, is the case that forced the first rules on AI and minors, through settlements, state laws and an under-18 ban rather than a judgment.

The pattern across all of them is the same. The work did not disappear. It moved from producing the thing to briefing, checking and taking responsibility for it.

The converted#

The most instructive companies on the map are not AI companies at all. They are companies from other industries that decided AI was the business.

Tesla is a car company that repositioned itself as an AI and robotics company: self-driving, a small robotaxi service, and a humanoid robot, with a car production line already torn out and rebuilt to make robots. Its valuation is a bet on machines that replace labour. SpaceX is the strangest AI company on the list because it is a rocket company; it bought xAI, listed with proceeds earmarked for AI compute, and then bought Cursor. Capital markets now price space launch, satellite internet and AI compute as one business.

Palantir sells the plumbing that puts AI into existing workflows for governments and large companies, and its growth is the evidence that in enterprise AI the winners often sell integration, not models.

Salesforce was the first large company to substitute AI agents for support staff on the record and to say publicly that engineering headcount would stay flat because of AI. Shopify told staff to prove AI could not do a job before asking for headcount, and its chief executive later coined a name for the cost, "slop grenades": unreviewed AI output pushed onto colleagues. Mandating AI is easy. Policing its quality is the new management task.

Klarna is the cautionary tale: it automated customer service aggressively, announced the savings, admitted it had cut too far, and rehired humans. Duolingo declared itself AI-first and was punished by its own users for looking as though it was replacing people. Between them they mark the boundary: customers tolerate AI doing the work, and they punish companies that seem to hold people in contempt. What happened to the companies that cut staff for AI follows both cases through to what their chief executives said afterwards.

Apple is the most striking admission on the map. After its own AI effort stumbled, it chose to build the new Siri on Google's models. Even Apple, which makes almost everything itself, decided that frontier intelligence is something you buy.

JPMorgan shows what AI does inside a bank: hundreds of use cases, around half its staff using internal models weekly, and open acknowledgement that some units have shed a large share of their headcount. The large banks are restructuring around AI while insisting on redeployment rather than mass layoffs, and how honest that insistence proves to be is one of the things to watch.

China#

Chinese AI is usually reported through one name, DeepSeek. The reality is three layers, and the answer to "which AI is deployed across the whole of China" depends on which layer you mean.

For ordinary people, the answer is ByteDance's Doubao, the assistant from the company behind TikTok and Douyin. ByteDance won the consumer layer the way it wins everything: distribution. Doubao is what people use.

For the state and heavy industry, the answer is DeepSeek. Its cheap open models caused the shock of early 2025 that showed the frontier could be reached with far less money than the US labs were spending, and within months provincial governments, carmakers, phone makers, hospitals and banks had embedded them. It later took state money and adapted its models to run on Huawei chips. Its consumer app faded; its models became the state-favoured backbone. DeepSeek is what China runs.

For developers, at home and abroad, the answer is Alibaba's Qwen, the open model family that most of the world's builders actually download. Alibaba also put Qwen to work inside its own commerce: an assistant with agentic access to the entire Taobao and Tmall catalogue, and agent tools for the cross-border sellers who use AliExpress and its international arm. Qwen is what people build on.

So: Doubao for people, DeepSeek for the state, Qwen for builders.

Behind them sits Huawei, the hardware backbone of Chinese self-sufficiency and the only Chinese stack running frontier models at scale. Tencent owns the distribution channel everyone else lacks, WeChat, and is the place where agents inside a super-app, with payments and a social graph attached, will work first if they work anywhere. Baidu is the search incumbent whose core business shrank while its robotaxis and chips became the parts worth watching. A second tier of labs, among them Moonshot (Kimi), Zhipu and MiniMax, produce open models that compete with the best in the world, particularly in coding, and two of them, Zhipu and MiniMax, were the first Chinese AI labs to list publicly. Xiaomi, the phone and car maker, builds its own models rather than renting one.

The state frames all of this through an explicit national policy of AI adoption across every major sector, a national computing network built on domestic chips, and capital from state funds that now anchors the leading labs. US export controls shape which chips China can buy and have pushed the whole ecosystem towards Huawei.

What it means outside China: expect ever cheaper frontier-class open models, with Chinese state capital and Chinese data law behind them.

India and the rest of the world#

India is the largest AI market that most coverage ignores, and the answer to "which AI is most widely deployed across India" is ChatGPT, not anything Indian, with Google's Gemini behind it and Meta AI inside WhatsApp probably undercounted. India is the second-largest market for the US labs and the hardest to make money in.

The Indian state chose to subsidise compute rather than pick a champion: the IndiaAI Mission rents a national pool of GPUs to startups at low rates and funds a spread of home-grown models. The nearest thing to a champion is Sarvam AI in Bengaluru, which builds open models for Indian languages and sells mainly to government and enterprise. Krutrim, Ola's AI unit and once the most hyped private effort, retreated from models to renting out GPUs. It is the cautionary tale.

Reliance Jio positioned itself as India's landlord for foreign models rather than a model builder: data centres in Jamnagar with Google and Meta, and AI subscriptions bundled for its mobile customers. Tata is India's biggest seller of AI services through TCS and the country's only chipmaker, though its first fab starts many generations behind AI-grade silicon. Infosys, Wipro and the rest of the IT services industry show the squeeze most clearly: AI grows their deal sizes while shrinking their billable hours, so AI revenue rises and total revenue stays flat. Language is where Indian labs have a real edge, through Bhashini and the BharatGen consortium's multilingual models covering all 22 scheduled languages.

Elsewhere the story is state money and chip allocations. In the Gulf, G42 in the UAE and Humain in Saudi Arabia build sovereign compute with US chips and US lab partners. In Japan, SoftBank is one of OpenAI's largest shareholders and a builder of its data centres, and Sakana AI is the country's most valuable startup. In South Korea the weight is in memory chips rather than models: Samsung and SK hynix supply the high-bandwidth memory beside every AI chip in the world.

The substrate#

If you learn three names from this page, make them these.

TSMC, the Taiwanese contract manufacturer, fabricates nearly every advanced AI chip: Nvidia's, AMD's, Broadcom's, Apple's. AI supply, and Taiwan risk, sit on one company.

ASML, the Dutch firm, is the sole maker of the extreme ultraviolet lithography machines used to print those chips. No ASML, no new leading-edge fabs anywhere.

SK hynix, together with Micron and Samsung, makes the high-bandwidth memory without which the chips cannot run. Memory, not logic, has repeatedly turned out to be the binding constraint.

Above them, Nvidia designs the GPUs and the software that train and run most models, and roughly the whole industry's compute budget flows through it. Broadcom designs the custom chips through which the big labs reduce their dependence on Nvidia, including Google's own. AMD is the credible second source of GPUs that buyers want for leverage. Arm licenses the processor architecture that sits beside every GPU in a server. Cerebras and Groq made alternative chips, and their fates show what happens to alternatives: one listed, the other licensed its technology to Nvidia and lost its leadership to it, in a deal now under antitrust scrutiny.

The landlords are the newest and most leveraged part of the stack. CoreWeave and its peers rent GPU capacity financed by debt. Oracle turned itself from a database company into the largest single landlord for OpenAI, with a backlog resting heavily on one customer's ability to pay. Amazon Web Services is Anthropic's primary training partner and builds its own chips. Scale AI and Surge AI supply the human-labelled data that trains the models, a large industry few people see.

Finally, power. Vertiv and Schneider Electric make the cooling and switchgear without which GPUs are useless. Constellation is restarting a reactor at Three Mile Island for Microsoft, and Meta and Google have signed long-term nuclear deals of their own. In the long run, power, not chips, is the bottleneck, and local resistance to data centres is growing.

How to tell whether a company belongs on this list#

Names on this page will change. The tests do not.

A company belongs on the map if removing it would break the stack (TSMC, ASML, the memory makers), if it trains frontier models rather than renting them (the labs), if it changed how a whole profession works (Cursor for coders, Harvey for lawyers), if it is the one a government treats as strategic (DeepSeek, Manus, Sarvam), if it is what most people in a large country actually use (Doubao, ChatGPT in India), or if it is an ordinary company whose AI bet is large enough to change its identity (Tesla, Apple, Salesforce).

A company does not belong on the map merely because it raised money, shipped a model that topped a benchmark, or announced an AI strategy. Those things happen every week.

What this means for your skills#

Read the map top to bottom and one thing repeats. At every layer the winning companies are pushing AI from answering towards doing: agents that operate computers, write code, take calls, place orders and run for hours. That is the shift that makes some human skills more valuable and others less. The value moves away from producing the first draft and towards knowing what to ask for, judging what comes back, and taking responsibility for it. Shopify's warning about unreviewed AI output and Klarna's rehiring are the same lesson from opposite directions: the bottleneck is now human judgement, and organisations that cut it pay for it.

The evidence on which skills gain value as AI improves is covered in the site's graded studies. This page is context for that question, not an answer to it.

A note on sources#

Everything above is drawn from company announcements, regulatory filings and reporting by Reuters, Bloomberg, the FT, CNBC, SCMP, Caixin, TechCrunch and the Economic Times, checked in September 2026. Corporate events described in the past tense (acquisitions, listings, policy decisions, product reversals) are historical and remain true; descriptions of roles and bets are the author's reading of the map and should be re-read once a year.