What a semiconductor is
A semiconductor is a material whose ability to conduct electricity can be switched and controlled. Silicon dominates because it is abundant, forms a perfect insulating oxide, and can be purified to extraordinary levels. Pattern billions of microscopic switches — transistors — onto a wafer of it and you get an integrated circuit: a chip that can store, transform and move information. When people say a chip is "5 nanometre" or "2 nanometre", they are referring to the manufacturing generation of that patterning, each generation packing transistors more densely and generally switching them with less energy.
The industry is best understood as three layers. Design companies (NVIDIA, AMD, Qualcomm, Apple's silicon teams, hundreds of others) architect what a chip does. Electronic-design-automation software and core intellectual property (Synopsys, Cadence, Arm) provide the tools and building blocks designs are made from. Foundries (TSMC, Samsung Foundry, Intel's contract arm, SMIC in China) actually fabricate the wafers. Most "chip companies" you can name design but do not manufacture — a split called the fabless model that reshaped the industry over thirty years.
How a chip is made
Manufacturing starts with a silicon wafer and builds the circuit layer by layer. The defining step is lithography: light projects the pattern of one layer onto the wafer through a mask, the exposed material is etched away or built up, and the process repeats for dozens of layers. Leading-edge fabs use extreme-ultraviolet (EUV) lithography — machines only ASML makes, each with mirrors polished to atomic smoothness. After fabrication, wafers are cut into dies, the good ones are packaged and connected to the outside world, and every part is tested. A leading-edge fab is one of the most complex factories humans operate: billions of dollars of equipment, cleanrooms thousands of times cleaner than a hospital operating theatre, and process recipes with thousands of steps.
That complexity explains the industry's economics. A new leading-edge fab costs tens of billions of dollars and a leading node's development cost runs into the hundreds of millions before the first wafer ships. Only a few firms can amortize that across global demand, so each generation concentrates production further. It also explains the industry's strategic weight: chips are dual-use, and export controls now govern who can buy the most advanced manufacturing equipment and AI accelerators.
Where AI fits in
AI workloads are dominated by matrix multiplication — the arithmetic at the heart of neural networks. GPUs and dedicated AI accelerators perform those operations across thousands of parallel execution lanes, which is exactly what general-purpose CPUs are not built for. When large language models and image generators took off, demand followed the arithmetic: training clusters bought accelerators by the tens of thousands, and inference — actually serving models to users — multiplied it again. Deloitte's 2026 industry outlook estimated that generative-AI chips would account for roughly half of total industry revenue in 2026 despite being a small fraction of units shipped: a small number of very expensive, very strategic parts.
That demand reshaped the supply chain. AI accelerators need enormous memory bandwidth, so high-bandwidth memory (HBM) — DRAM dies stacked and connected through advanced packaging — became as strategic as the logic die itself. CoWoS-style packaging capacity, which glues logic and HBM into one accelerator module, has repeatedly been the true constraint on how many GPUs can ship in a quarter. The AI chip story is as much a packaging and memory story as a transistor story.
The state of the industry in 2026
Three developments define the current cycle. First, TSMC began volume production on its 2-nanometre-class process with demand from flagship smartphone and AI customers, while planning a 1.6-nanometre generation behind it — sustained multi-billion-dollar investment in Taiwan, Arizona, Japan and Europe. Second, advanced packaging and HBM supply, not lithography, became the bottleneck analysts watch: IDC has forecast the overall semiconductor market passing the trillion-dollar mark ahead of earlier consensus, driven substantially by AI inference. Third, geopolitics hardened: export controls restrict the most advanced equipment and accelerators from reaching certain markets, pushing China to invest heavily in domestic alternatives and pushing buyers to diversify where their chips are made. None of these trends is speculative — each is documented in the sources linked below — but how they balance out is exactly what the industry's own forecasts disagree about.
Advantages and limits of the technology
Semiconductors are the most successful scaling technology in engineering history: decades of exponentially rising transistor counts, falling per-transistor cost and rising efficiency. But the physics and economics both tighten at the leading edge. New nodes deliver smaller gains per generation than they once did while costing more per wafer; photon scaling in the original Dennard sense ended years ago, which is why efficiency now comes from architecture (specialised accelerators, chiplets, packaging) as much as from density. For AI specifically, the binding constraints have shifted from compute to power, cooling and memory bandwidth — a data centre can only draw so many megawatts, and moving data costs more energy than computing with it.
The practical takeaway for readers following the space: treat single-number roadmaps ("2nm next year!") as marketing until a manufacturer confirms volume production, watch packaging and memory capacity as closely as logic nodes, and remember that fab locations concentrate risk in ways no single buyer can diversify away. Our AI hardware evolution coverage and the technology desk track these developments as they are announced.
Why this matters beyond chips
Every technology this publication covers sits on this foundation. Model releases are possible because accelerators got faster; the quantum-computing story is partly a cryogenics and control-chip story; electric vehicles (see our battery coverage) are as much power-semiconductor stories as battery stories. Understanding who makes chips, where, and under whose export rules is the cleanest way to understand which AI and hardware headlines are real capacity and which are announcements.
Frequently asked questions
- What is a semiconductor?
- A semiconductor is a material — usually silicon — that conducts electricity better than an insulator but worse than a conductor. By doping and layering it in precise patterns, engineers build billions of switches (transistors) on a single chip. Those switches carry out every logical operation a computer performs, which is why semiconductors sit inside phones, cars, data centres and medical devices.
- Why does AI need special chips?
- Training and running AI models mostly means multiplying huge matrices over and over. GPUs and dedicated AI accelerators do that in parallel across thousands of cores, which general-purpose CPUs do far less efficiently. That parallelism — not raw clock speed — is why AI data centres fill up with GPUs and custom accelerators, and why demand for those chips sets the pace of the whole industry.
- What does the nanometer number on a chip mean?
- It is a marketing shorthand for the manufacturing generation, not a literal measurement anymore. A 2-nanometer-class process has densely packed transistors with different physical gate structures than older nodes. What matters practically: each new node packs more transistors into the same area and generally switches with less energy — which is why flagship chips reserve the newest nodes.
- Who actually manufactures the world's most advanced chips?
- The most advanced logic chips are manufactured overwhelmingly by TSMC, with Samsung also producing at leading-edge nodes and Intel building toward parity. Most chip companies — NVIDIA, AMD, Qualcomm, Apple's chip teams — design chips but do not fabricate them (the "fabless" model). One company, ASML, is the only supplier of the extreme-ultraviolet lithography machines leading-edge fabs require, which is why the supply chain is strategically concentrated.
- What is advanced packaging and why did it become a bottleneck?
- Advanced packaging connects multiple silicon dies into one component — stacking memory beside a processor, for example. It became the bottleneck because AI accelerators need enormous memory bandwidth, and advanced packaging and high-bandwidth memory capacity — not transistor density — is often what limits how many accelerators can actually ship.
- Are chip prices going to fall as supply catches up?
- Capacity for leading-edge logic and advanced packaging has been expanding, but each new node costs more to build and operate, and demand from AI data centres has been absorbing new supply. Chip prices depend on the segment — commodity chips face cycles, while leading-edge AI silicon has stayed scarce. We do not publish price forecasts; for current figures, check the manufacturers' own earnings materials.