Questions people
actually ask.

No, and this surprises a lot of people. NVIDIA designs the chips and outsources manufacturing, mostly to TSMC in Taiwan. It is a fabless company: engineers and intellectual property rather than factories. AMD and Apple work the same way. Physical manufacturing is extraordinarily capital intensive, and TSMC has spent decades becoming very difficult to replace at the leading edge.Explore this layer: Semiconductors & Manufacturing →

CPUs are optimised for flexible, general purpose and low latency computing. They handle branching, unpredictable work and the everyday business of running an operating system. GPUs contain many more, simpler processing units and are especially effective at performing large numbers of similar calculations in parallel. That parallel arithmetic is most of what training and inference consist of, which is why a design originally refined for rendering graphics turned out to suit AI so well.Explore this layer: Semiconductors & Manufacturing →

Yes. AI can run on conventional processors, but specialised accelerators make large modern AI workloads dramatically more efficient. The difference is not capability so much as cost, speed and energy. A model that takes seconds on an accelerator might take minutes or hours on a general purpose CPU, and at scale that gap becomes the entire economics of the business.Explore this layer: Semiconductors & Manufacturing →

High Bandwidth Memory stacks memory dies vertically and places them next to the accelerator rather than out on the board. That gives far more bandwidth, which matters because a large model has to stream billions of parameters for every token it produces. If memory cannot supply data fast enough, the accelerator waits. That is why memory bandwidth, not raw arithmetic, is often the real constraint, and why HBM supply from SK hynix, Micron and Samsung has repeatedly limited how many AI systems could be built in a given year.Explore this layer: Semiconductors & Manufacturing →

Training a large model means running many thousands of accelerators at high utilisation for weeks. Those chips turn nearly all that electricity into heat, and the heat has to be removed, which is where cooling and in some designs water consumption come in. This is why energy, grid connections and cooling have become genuine parts of the AI investment conversation rather than background details. In several markets, securing power is now harder than securing chips.Explore this layer: Energy, Grid & Cooling →

Not in the strict sense. Meta's Llama is an open-weight model family. Its weights can be downloaded and run independently, although Meta's licensing terms mean it is not considered open source under the Open Source Definition maintained by the Open Source Initiative. The distinction matters commercially. Open weights let a company run a capable model on its own infrastructure, which puts real pricing pressure on paid APIs, but the licence still places conditions on how it may be used.Explore this layer: Models & AI Platforms →

Most AI use is still reactive: you ask, it answers. An agent is given a goal instead, and works out the intermediate steps, calling tools, reading data and looping until it decides the task is finished. That shift is why permissions, audit trails and identity suddenly matter so much. A chatbot that is wrong wastes your time. An agent that is wrong can take an action on your behalf, which is a different category of problem.Explore this layer: Applications, Agents & AI Software →

Training is not a single event. Pre-training is the long, expensive phase where a model learns general patterns from very large amounts of data. Post-training shapes that raw model into something useful and better behaved, often using human feedback and reinforcement learning. Fine-tuning adapts an existing model to a narrower task or domain, and can be done repeatedly and comparatively cheaply. Inference is the model actually answering, which happens billions of times a day. Much of the current demand for chips, power and data centres is driven by inference at scale rather than by headline training runs alone.Explore this layer: Models & AI Platforms →

Infrastructure overlap. Bitcoin mining, and historically Ethereum mining, required many of the same ingredients now sought by AI infrastructure developers: large power allocations, suitable sites and industrial cooling. Ethereum itself moved from proof of work to proof of stake in September 2022, so it is no longer mined. When mining economics weakened, several operators held energised sites that would take years to permit from scratch. Converting them for AI tenants is a substantial rebuild rather than a relabel, but the scarce ingredient, power, was already secured.Explore this layer: Data Centres & Cloud Infrastructure →

Parts of it may be. Some valuations clearly assume everything goes right. But the infrastructure spending is happening regardless of which application company wins, which is the argument for owning the suppliers rather than guessing the winners. The honest position is that the ecosystem contains both durable businesses and speculative ones, and the market is not currently doing a careful job of distinguishing them. That is a reason to read the economics of each layer rather than to treat AI as one trade.Explore this layer: Users, Enterprises & Distribution →

A large, general purpose model trained on a very broad dataset, which can then be adapted to particular tasks. The GPT, Claude, Gemini and Llama families are all foundation models. They are called foundations because applications, tools and agents get built on top of them. Training one is an expensive bet that sitting at the base of the stack is where lasting value accrues, which is precisely the question the rest of this site is trying to help you think about.Explore this layer: Models & AI Platforms →

Because TSMC, which manufactures most leading edge chips, is based there, and Taiwan sits in a contested region. A serious disruption to its operations would slow the global AI build out considerably, and there is no quick substitute for that capacity. That is the reason the United States, the European Union and Japan are all subsidising domestic manufacturing. The motivation is industrial policy and national security as much as economics.Explore this layer: Semiconductors & Manufacturing →

ASML is currently the world's only supplier of production EUV lithography systems, which are required for manufacturing many of the most advanced chips. Other companies make lithography equipment for less advanced processes, so this is not a monopoly on lithography as a whole. The EUV machines themselves are remarkable. Light at a wavelength of 13.5 nanometres, close to the X ray range, is produced by firing a laser at tin droplets tens of thousands of times a second. The light is then steered by mirrors so smooth that if you scaled one to the size of Germany, the biggest bump would be less than a tenth of a millimetre high. A single system costs in the region of two hundred million dollars and ships in dozens of freight containers. Canon and Nikon have both competed in lithography and neither has brought a production EUV system to market. The advantage is technical and cumulative rather than granted by regulators. The Dutch government, under pressure from the United States, also restricts sales of the most advanced systems to China, which makes one company in one Dutch city a genuine chokepoint in the technology relationship between two superpowers.Explore this layer: Semiconductors & Manufacturing →