Semiconductors: Everything You Wanted to Know

KEYPOINTS

• 01

  Semiconductors are the unsung building blocks of the digital economy.

  Software may be eating the world, say NZS Capital investors Jon Bathgate and Brinton Johns, but semiconductors do the work underneath. Nearly every major trend of the next 10 to 15 years runs on chips: electric and autonomous cars, AI, the connected factory, drug discovery. The COVID-19 genome was mapped in January, before the virus really hit the US. Their test for strategic importance: if a company were wiped out, would it matter? Lose CVS and you go to Walgreens. Lose a couple of firms in the chip value chain and the world goes backwards 10 or 20 years.

• 02

  The industry went horizontal — and a passed-over CEO candidate created the foundry.

  Early players like Texas Instruments did everything: made equipment, fabbed chips, sold calculators and the Speak & Spell. As the industry matured, it split into layers. Morris Chang, passed over for TI's CEO job, returned to Taiwan. Asked to build a chip industry on an earthquake-prone island, he realised fabrication was customers' pain point. So he founded TSMC to make other companies' chips without competing with them: the foundry model. Today, when Apple 'makes' its chips, TSMC fabs them; Tesla's are fabbed by Samsung in Austin.

• 03

  A few companies hold chokepoints across the value chain.

  Chip designers lay out chips with electronic design automation software from just two firms, Synopsys and Cadence. TSMC fabs nearly all bleeding-edge digital chips, with Samsung as second source. Equipment comes from a handful of suppliers: ASML in the Netherlands, Lam Research, KLA and Applied Materials in the US, and Tokyo Electron. Everyone collaborates closely, and there's margin throughout the chain. Doing the whole stack yourself every two years became too hard, even Intel outsourced its design software.

• 04

  Fall behind and the race gets exponentially harder.

  GlobalFoundries, spun out of AMD and backed by Middle Eastern sovereign wealth, had effectively unlimited resources. Yet facing the move to 10nm, it gave up and settled for older technology. China's SMIC has spent tens of billions over 20 years and remains about a decade behind. Its plan for a fab with no US equipment targets a 28nm line, roughly 10-year-old technology by the time it's running. Skipping nodes is almost impossible, Johns says. Combine the same machines and 'one plus one plus one could give you blue,' and it takes a year to trace why.

• 05

  EUV lithography is one of humanity's hardest engineering feats.

  Leading-edge chips are 5nm; a nanometre is about two to three silicon atoms. Traditional lithography light is 193nm, stretched with lenses and water immersion until around 10nm. ASML's extreme ultraviolet light is about 10.5nm. To make it, drops of molten tin are hit with a laser in a vacuum, about 50,000 times a second, creating plasma. Because EUV is absorbed by all known mirrors, ASML had to invent new ones to bounce it off about 20 of them. Only one private German company makes the laser. Nikon gave up on EUV, so ASML has a natural monopoly. One system costs about $200 million and needs four 747s to ship.

• 06

  Classic Moore's Law has ended — 'More than Moore' hasn't.

  Gordon Moore predicted transistor density would double every one to two years at the same price, which democratised chips into everything. Bathgate thinks that literal definition has empirically ended against the laws of physics. But computing performance can keep doubling through chiplets, stacked chips, 3D transistors (enabled by Lam Research), new materials and specialised architectures. ASML's high-numerical-aperture tools, about $300 million each, map a path to around 2 or 1.5nm by decade's end. Even Intel, which led Moore's Law for some 40 years, fell off around 2015: a two-year node took five.

• 07

  GPUs and CUDA turned Nvidia into an AI powerhouse.

  If a CPU is a big V8 engine, a GPU is lots of little engines doing parallel work, as when rendering screens. When Jensen Huang started Nvidia there were about 100 GPU competitors; now there are two, both fabbed at TSMC. About 15 years ago Nvidia released CUDA, letting anyone write code for parallel GPU architectures. That carried GPUs into data centres, where Nvidia's business grew from a couple of hundred million dollars to about $4 billion in five years. AI model parameters were doubling every three to four months, needing GPU clusters with only a few CPUs. Meanwhile ARM licenses low-power designs to Apple and Qualcomm, and open-source RISC-V is emerging.

• 08

  Chips are the ultimate geopolitical leverage.

  The US cut ZTE off from chips over Iran sanctions for a few months, crippling it. Huawei, once about 20% of TSMC's sales and a national champion with over $100 billion in revenue, was later blocked: TSMC couldn't build for it using US equipment, and nothing can be built without it. US companies earn about half of semiconductor revenue; China buys over half of chips. But about 70% of chips pass through Taiwan at the front or back end. Less than 15% are now made in the US. Rebuilding requires a whole ecosystem, which moves in decades.