For decades chip progress had a single mantra: make transistors smaller. But today the bottleneck slowing the AI revolution is no longer (only) there. It's in a field few people know but that is worth billions: advanced packaging — the way chips are assembled and connected to each other.

What advanced packaging is

Traditionally a chip was a single piece of silicon inside a casing. With advanced packaging, instead, several "pieces" of silicon — so-called chiplets — are joined into a single high-performance module. Technologies like TSMC's CoWoS let compute logic (GPUs) sit alongside ultra-fast HBM memory on the same substrate, delivering the bandwidth and density an AI accelerator needs.

Why it became the bottleneck

The figure that explains it all: memory and packaging together now account for 60-70% of the cost of an AI accelerator. Logic silicon, once the dominant item, no longer is. Demand for these assemblies is so high that supply can't keep up.

Grafico interattivo: TradingView · NASDAQ:SMH

TSMC leads the market with a share of around 40% in advanced packaging for AI chips, and is expanding CoWoS capacity from about 80,000 to 120,000-130,000 units per month, with NVIDIA alone absorbing roughly 60%. Despite capacity growing at a dizzying pace, demand keeps outstripping supply.

Why it matters for investors

  • Value is shifting: it's no longer only those who design chips that matter, but those who assemble them and supply the memory. This reshapes the semiconductor value chain.
  • Concentration and risk: very few companies dominate these technologies, creating strategic (and geopolitical, given Taiwan's centrality) bottlenecks.
  • Emerging standards: initiatives like UCIe aim to standardise chiplet interconnection, a key piece for future growth.

The semiconductor ETF chart (SMH) offers a reference for the sector's trend; read it as an illustrative tool, not trading guidance.

The bottom line

Advanced packaging is the least visible but increasingly decisive part of the chip race: the point where memory, assembly and thermal engineering determine how fast AI can run. Understanding it helps you read the semiconductor sector with fresh eyes — and recognise where value is really moving.

Disclaimer: this article is for information purposes only and does not constitute financial advice. Any investment decision should be assessed against your own circumstances and, if needed, with a qualified professional.