Who builds what, and why it matters
Every city needs foundations, power, roads, buildings, and people using them. AI is the same, just faster and more expensive.
Written by Matthew Bernath
Before any AI can run, someone has to build the specialised hardware. Normal chips can't handle it. AI needs to do millions of calculations at once, not one at a time.
Makes the leading AI chip (the GPU). Almost everyone uses them. The arms dealer of the AI war.
No, and this surprises a lot of people. NVIDIA designs the chips but outsources all manufacturing to TSMC in Taiwan. NVIDIA is a fabless company, no factories, just engineers. The same is true for AMD and Apple. The actual physical manufacturing is extraordinarily capital intensive and TSMC has spent decades becoming essentially irreplaceable at it.
A CPU (Central Processing Unit) is a general purpose chip. It's fast at doing one thing at a time and handles everything your computer does day to day. A GPU (Graphics Processing Unit) was originally designed for gaming because rendering graphics requires doing thousands of simple calculations simultaneously. It turns out that's exactly what AI needs too, millions of parallel calculations at once. That's why NVIDIA, which dominated gaming GPUs, ended up accidentally dominating AI infrastructure.
Makes competing chips. Cheaper than NVIDIA and catching up fast.
Trying to get back in the game after being late to AI. A turnaround story.
Actually manufactures the chips NVIDIA and AMD design. Nobody else builds at their scale. The factory everyone depends on.
Because TSMC, the company that manufactures most of the world's advanced chips, is based in Taiwan. Taiwan also sits in one of the world's most geopolitically contested regions. If anything were to disrupt TSMC's operations, a natural disaster, a conflict, a blockade, the global AI buildout would effectively stop. This is why the US and EU are spending billions subsidising domestic chip manufacturing. It's not just economics. It's national security.
No, and this surprises a lot of people. NVIDIA designs the chips but outsources all manufacturing to TSMC in Taiwan. NVIDIA is a fabless company, no factories, just engineers. The same is true for AMD and Apple. The actual physical manufacturing is extraordinarily capital intensive and TSMC has spent decades becoming essentially irreplaceable at it.
Makes the machines that make the chips. One company, total monopoly, Dutch. No ASML, no modern chips.
Effectively no, and the reason is one of the more extraordinary stories in modern industrial history. ASML makes the machines that use extreme ultraviolet (EUV) light to etch circuit patterns onto silicon wafers. The light source operates at a wavelength of 13.5 nanometres, just above the X ray range. To generate it, they fire a laser at a tiny tin droplet 50,000 times per second, which creates a plasma that emits the right wavelength of light. Then they focus and direct that light using mirrors so precise that if you scaled them up to the size of Germany, the biggest bump would be less than a tenth of a millimetre high. It took ASML roughly 30 years and billions in R&D to get this working. The supply chain involves over 5,000 suppliers across dozens of countries. A single EUV machine costs around $200 million, weighs 180 tonnes, and ships in roughly 40 freight containers. Canon and Nikon have tried to compete at various points. Neither has cracked EUV. ASML has a monopoly not because regulators gave it to them or because they crushed competitors, but because what they do is so technically complex that nobody else has managed to replicate it after decades of trying. The geopolitical dimension is significant too. The Dutch government, under pressure from the US, now restricts ASML from selling its most advanced machines to China. That single export restriction is arguably one of the most powerful economic weapons in the US China technology war. One Dutch company, making one type of machine, in one city (Veldhoven), is a genuine chokepoint in the global AI race.
Designs the chip architecture blueprint that almost every device uses. They license the design, they don't build the chips.
Make the networking chips that connect all those AI chips together at speed. The glue between the GPUs.
Make the memory chips. AI needs to store and retrieve data at extreme speed. These companies make that possible.
AI data centres use enormous amounts of electricity. A single large data centre can use as much power as a small city. Someone has to generate that power reliably.
Makes fuel cells that can power data centres without relying on the grid. Always on, no outages.
Training a large AI model requires running thousands of chips at full capacity for weeks or months. Chips generate enormous heat. Data centres cool that heat using water based cooling systems. A single large training run can use millions of litres of water and enough electricity to power a small town. This is why energy infrastructure, Bloom Energy, nuclear plays, alternative energy, is a genuine AI investment thesis, not a peripheral one.
Nuclear plays, including small modular reactor designs aimed at powering data centres directly. AI needs baseload power that solar can't guarantee on its own.
Training a large AI model requires running thousands of chips at full capacity for weeks or months. Chips generate enormous heat. Data centres cool that heat using water based cooling systems. A single large training run can use millions of litres of water and enough electricity to power a small town. This is why energy infrastructure, Bloom Energy, nuclear plays, alternative energy, is a genuine AI investment thesis, not a peripheral one.
You need physical buildings stuffed with chips, cooled constantly, and connected to the internet at enormous speed. This is the real estate of the AI economy.
Own the biggest data centres on earth through AWS, Azure, and Google Cloud. They rent compute power to everyone else.
A newer data centre company built specifically for AI workloads. Rents NVIDIA GPUs at scale to anyone who needs them.
Started as crypto miners. Already own the buildings, power contracts, and cooling systems. Now pivoting to rent that infrastructure to AI companies.
Pure coincidence of infrastructure. Bitcoin and Ethereum mining requires cheap electricity, large buildings, and lots of cooling, which is exactly what AI data centres need. When crypto revenues collapsed, miners were left with sunk cost infrastructure sitting idle. Rather than write it off, the smart ones pivoted to renting that same infrastructure to AI companies. Core Scientific, IREN, Applied Digital, and Cipher Mining are all playing this transition.
Data has to move between chips, servers, and data centres at very high speed. These companies build the pipes and roads that make that movement possible.
Builds the switches that move data around inside data centres. The internal road network.
Coherent and Lumentum make photonics components, the fibre optic connections that move data at the speed of light between data centres. Lightwave Logic is a pre-revenue research company developing polymer based optical materials, a more speculative bet on the same trend.
Satellite connectivity infrastructure. The long range road that connects remote areas and emerging markets.
This is where the actual AI gets built and run. The companies building the models that everyone else uses, and the platforms hosting those models.
Built Gemini. Owns DeepMind. Has the best data in the world via Search. The incumbent with the most to lose and the most to gain.
A foundation model is a large, general purpose AI model trained on vast amounts of data that can then be adapted to specific tasks. GPT 4, Gemini, Llama, and Claude are all foundation models. They're called foundation models because everything else gets built on top of them, applications, tools, agents. The companies building foundation models are making a very expensive bet that being at the base of the stack is where the lasting value accrues.
Training is when an AI model learns, you feed it enormous amounts of data and it adjusts billions of internal parameters until it gets good at the task. It's extremely expensive and only happens once per model version. Inference is when the trained model actually answers your question, it's much cheaper and happens billions of times a day. Most of the chip and data centre demand you're hearing about is driven by inference at scale, not just training.
Built Llama (open source). Owns Instagram and WhatsApp, which generate enormous training data. Playing a different game to everyone else.
It means Meta has released the underlying model weights and code publicly so anyone can download, run, and modify it. The alternative is closed source, like OpenAI's GPT 4, which you can only access via their API. Open source AI is controversial because it democratises access but also means anyone, including bad actors, can use it without guardrails.
A foundation model is a large, general purpose AI model trained on vast amounts of data that can then be adapted to specific tasks. GPT 4, Gemini, Llama, and Claude are all foundation models. They're called foundation models because everything else gets built on top of them, applications, tools, agents. The companies building foundation models are making a very expensive bet that being at the base of the stack is where the lasting value accrues.
Owns a large stake in OpenAI (ChatGPT). Azure is how most businesses access AI. The enterprise distribution channel.
A foundation model is a large, general purpose AI model trained on vast amounts of data that can then be adapted to specific tasks. GPT 4, Gemini, Llama, and Claude are all foundation models. They're called foundation models because everything else gets built on top of them, applications, tools, agents. The companies building foundation models are making a very expensive bet that being at the base of the stack is where the lasting value accrues.
This is the next big shift. Most AI today is reactive, you ask it something, it answers. Agentic AI takes actions autonomously on your behalf. You give it a goal and it figures out the steps, uses tools, browses the web, writes code, sends emails, and loops until the job is done, without you prompting each step. Think less chatbot and more digital employee. Palantir, ServiceNow, and Microsoft are all building agentic products. It's where most of the enterprise money is heading right now.
A European AI cloud company that emerged from the old Yandex international business. After Yandex sold its Russian operations in 2024, the Dutch parent kept the global assets, rebranded as Nebius, and refocused on AI infrastructure. Early stage but interesting positioning.
Data platforms. AI is useless without clean, accessible data. These companies store, manage, and serve that data to the AI.
Companies building useful products on top of the AI models. These are the applications people and businesses actually interact with.
AI analytics for governments and large enterprises. Heavy defence and intelligence contracts. Controversial but deeply embedded.
This is the next big shift. Most AI today is reactive, you ask it something, it answers. Agentic AI takes actions autonomously on your behalf. You give it a goal and it figures out the steps, uses tools, browses the web, writes code, sends emails, and loops until the job is done, without you prompting each step. Think less chatbot and more digital employee. Palantir, ServiceNow, and Microsoft are all building agentic products. It's where most of the enterprise money is heading right now.
Enterprise workflow software baking AI into everything. The quiet but essential layer of corporate IT.
This is the next big shift. Most AI today is reactive, you ask it something, it answers. Agentic AI takes actions autonomously on your behalf. You give it a goal and it figures out the steps, uses tools, browses the web, writes code, sends emails, and loops until the job is done, without you prompting each step. Think less chatbot and more digital employee. Palantir, ServiceNow, and Microsoft are all building agentic products. It's where most of the enterprise money is heading right now.
Enterprise software with AI automation. Slower moving and less exciting, but well embedded in large organisations.
As much an AI company as a car company. Full Self Driving and the Dojo supercomputer are massive AI bets sitting inside a car manufacturer.
Fintech that benefits from retail enthusiasm around AI stocks. A meta play on the whole trend.
Possibly, at the margins. Valuations on some names are clearly pricing in perfection. But the infrastructure buildout, chips, power, data centres, is happening regardless of which AI application wins. The picks and shovels argument is that you don't need to know who strikes gold, you just need to own the people selling the shovels. That said, some of the application layer bets are speculative and the market is not uniformly rational right now. Caveat emptor.
AI applications in biotech and insurance. Niche bets on AI transforming specific industries.
More AI means more attack surface. More data means more to steal. AI powered attacks need AI powered defence. The security layer grows in direct proportion to everything else.
Uses AI to detect and stop threats in real time. The endpoint security leader. Every laptop in a big company is probably running this.
Broad cybersecurity platform covering networks, cloud, and endpoints. One of the most comprehensive security stacks available.
Cloud native security. Protects companies where the perimeter no longer exists, when everyone works from anywhere.
Autonomous AI security platform. Detects and responds to threats without needing a human in the loop.
Investor & portfolio thinker
Founder, Beaumont Sycamore Media
I spend a lot of my time thinking about where capital flows in technology, both as a practitioner and as an investor trying to understand which businesses will compound over the next decade.
The AI ecosystem is where I spend a lot of my thinking time, both as a practitioner building on these technologies and as an investor trying to understand where the real durable value gets created. There's a lot of noise. Most of it is either hype or fear. The actual story is more interesting than either.
My view is that the infrastructure layer (chips, power, data centres, connectivity) is where the most defensible businesses are being built right now. The application layer will produce enormous winners too, but it's harder to pick them early. The picks and shovels tend to win regardless of who discovers the gold.
This explainer is my attempt to map the ecosystem in plain language. Not for analysts. For anyone who wants to understand what's actually being built and why it matters.