Glass, not lithography: Korea's HBM lead rests on a packaging step it does not own

The two scaling rates the report puts side by side are the packaging argument, whether or not glass ends up winning it. Nvidia's logic transistor count rose 6.2 times between the A100 80GB and Rubin ^[7], while the HBM travelling with it rose 3.6 times ^[8], so memory per unit of compute fell by about 42% ^[9]. Fewer bytes per transistor means every byte has to move faster, and speed there is wiring density, which lives in the substrate rather than in the DRAM cell.

SEMI's own two numbers do not describe the same market. A 67.2% annual growth rate ^[13], compounded across the twelve years to 2040, multiplies a market roughly 477 times, so the $13 billion projected for around 2040 ^[14] implies a starting base near $27 million ^[10]. Begin instead from a billion dollars and the trip to $13 billion works out at 23.8% a year ^[11], about a third of the advertised rate. The growth figure describes a pilot line; the market figure describes an industry.

Park Sung-soo, CEO of JWMT, told the paper that CoWoS is reaching its technical limits and that glass is Korea's chance to step out of TSMC's shadow ^[17]. He is also describing the market he sells into, and no Korean spending figure appears beside the claim. TSMC has taken glass CoWoS development to Ibiden in Japan and Innolux in Taiwan ^[20], which is where any Korean packaging franchise would have to displace incumbents rather than open a category. What Korea has on the record is a quote: Samsung Electro-Mechanics opened the year at 270,000 won and has traded into the 1.3 million won range ^[6], a 4.8 times move ^[12] that prices an option without qualifying a substrate.

The 2013 episode is worth reading for mechanism rather than nostalgia. Apple moved iPhone AP production to TSMC ^[2], Nvidia now places most of its top-tier chip work there ^[3], and a senior equipment industry official quoted anonymously by the paper locates the cause inside Samsung, which treated the foundry division as an outsider while chasing front-end miniaturisation ^[16]. Memory looked insulated from all that, because stacking DRAM was Korea's own process. It is less insulated once the logic die inside an HBM stack, and the substrate beneath it, determine what the product does ^[18]. A supplier can own the best cell in the industry and still have its content per socket set by whoever owns the platform it plugs into.