The Substrate Underneath

Underneath the silicon and underneath the interposer there is a board. The package substrate fans thousands of connections out from the die complex to the motherboard and carries power back in, built by laminating alternating layers of insulating film and copper onto a core. It is the least photographed component in an accelerator and one of the hardest to buy.

How hard is readable from what suppliers are spending, which is a better instrument than any forecast because it is board-approved and filed. IBIDEN approved approximately 500 billion yen of capital investment across fiscal 2026 to 2028, roughly 220 billion of it centred on its Gama plant in Ogaki, explicitly for AI and high-performance server substrates. AT&S went further, in a form that carries more weight: an ad-hoc disclosure — the regulated announcement a listed company must make when information moves its price — stating agreements with AMD and one unnamed technology company to expand high-end substrate capacity in Kulim, Malaysia, with the 1.5 to 2.0 billion euros required fully financed by long-term customer commitments. The customer is funding the supplier's factory. That is what a buyer does when it is more worried about getting the part than about its price.

The technical problem is area. As accelerator packages grow, the substrate beneath them has to stay flat across a wider span while carrying more layers, and organic laminate warps. For scale: one estimate puts NVIDIA's Rubin Ultra package at roughly 7,470 square millimetres, about nine reticles of silicon and memory, against around 2,739 for Blackwell. Round interposer wafers waste a growing share of their area on rectangular products at that size, and CoWoS interposer wafers run on the order of $10,000 each, comparable to a processed 7nm wafer.

The proposed answer is to replace the organic core with glass. Intel announced it in September 2023 with more than a billion dollars behind it, claiming a tenfold increase in interconnect density and 50% less pattern distortion. The mechanism is thermal expansion: a glass core can be tuned to roughly 3 to 10 parts per million per degree Celsius against silicon's 2.6, which cuts warpage by about half, and a rectangular 510 by 515 millimetre panel devotes more than 75% of its area to large die against roughly 50% for a round 300 millimetre wafer.

Three years on it is in no commercial product, and every timeline has slipped. Absolics planned mass production for the first half of 2024 and is now targeting the end of 2026; a U.S. government award summary written earlier expected first customer deliveries in 2025. Intel has moved to licensing its patents and showing demonstration vehicles, with deployment pointed at around 2030. The obstacles are specific: glass chips and cracks at the edges during drilling and dicing, and while edge coating has cut measured edge stress from 95 MPa to 49 MPa, metallising vias below ten microns and holding nanometre-scale flatness across a half-metre panel remain unsolved.

The most useful thing anyone has written about the field is a sourcing observation rather than an engineering one: no production design exists, and every customer publicly attached to glass — AMD, Broadcom, AWS, NVIDIA — comes from industry reporting rather than from confirmation by anyone involved. TSMC, whose decision matters most, is going panel-level first on 310 by 310 millimetre panels with glass integration described by an equipment supplier as under review rather than committed.

What is deliberately absent is named. The ABF shortage is among the most repeated stories in this supply chain and among the worst sourced. Estimates of one company's share of the build-up film market that surfaced in a session's searching ran from 58% to over 95%, all from syndicated reports that sit behind purchase and disagree with each other. None is used, and no market-share, market-size, shortage-percentage or price-increase figure is asserted anywhere on this page.