Nvidia weighs glass chip substrates — and the vias aren't ready
The weekend wire is thin on hard news, so this brief goes upstream — to the packaging layer where the next round of AI performance is being decided, and to a small bet that light, not copper, carries the data between accelerators.
Nvidia is weighing glass-core substrates for its next accelerators, and the supplier building the equipment says the hard part isn't the glass — it's the metal inside it.
Germany's SCHMID Group, which sells the plating and wet-process tools that go into substrate production, told investors on its first-half earnings call that it is working on glass-core substrate equipment with key suppliers inside the Intel, Nvidia and AMD supply chains. Chief strategy officer Roland Rettenmeier named the bottleneck in the same conversation: metallizing through-glass vias — the microscopic vertical connections drilled through the panel — is not solved, and final customer qualification is not complete. That second sentence is the story. Glass beats the plastic substrate it would replace on the properties that matter for big AI packages: flatter, stiffer, stable at temperature, and industries' own figures have it lifting speed roughly 40% while halving power draw. But a substrate whose vias don't work yet is an expensive coaster.
Korean reporting has Nvidia asking substrate makers to accelerate development, with adoption possibly pulled forward to before 2028. Treat that date as soft, because every timeline in this segment has slipped. Absolics, SKC's US subsidiary, planned mass production for early 2024 and is still in package-level qualification after an extra ₩92.1 billion of investment; Samsung Electro-Mechanics' glass joint venture targets the second half of 2027; Intel has shifted to licensing its patents with its own deployment pointed at around 2030; and TSMC's panel-level CoPoS line has glass integration "under review, not committed," with TrendForce putting commercial glass cores after 2030. Tom's Hardware's read of the roadmaps is blunt: no production design exists, and every customer name attached to the technology — AMD, Broadcom, AWS, Nvidia — comes from industry reporting rather than a signed announcement. Nvidia has confirmed nothing.
Which is why the version of this story now moving through Chinese media — "Nvidia partners with SCHMID to lay out glass substrates" — deserves to be read slowly. SCHMID said it works with suppliers in Nvidia's chain. Nobody announced the partnership in that headline. The honest state of play is that the equipment exists, the qualification doesn't, and the company with the most to gain keeps signalling interest without signing anything.
PicoJool raised a $27.5 million Series A led by Socratic Partners to build laser links between AI accelerators. The Palo Alto company, backed by former Intel chief executive Pat Gelsinger, is sampling 200G VCSELs and smaller microVCSELs it says scale optical connectivity from 800G toward 1.6T and 3.2T, plus a 64G configuration aimed at optical modules on PCIe 6.0, a portfolio it says is already being qualified with customers. Hudson River Trading joined the round, which follows a $12 million seed led by Playground Global and takes total funding to $39.5 million.
The pitch is the wall the industry keeps hitting: clusters add processors faster than they add bandwidth or power budget, so every generation makes moving a bit more expensive. We covered that thesis in a much larger round — Ayar Labs adds $150M more as copper interconnects hit the wall — and the difference here is manufacturing economics. VCSELs are cheap to produce at volume, so PicoJool's real claim is cost per lane at data-center scale, not the fastest possible link.
The cardinal Pope Leo XIV put in charge of the Vatican's AI commission does not use AI. Cardinal Michael Czerny, appointed in May to chair the Inter-Dicasterial Commission on Artificial Intelligence, told POLITICO's Brussels Playbook Week Ender podcast that he just hasn't had the time — and that he considers it "my moral duty also to use it so that I have personal experience, and I hope it helps me with my work." He wants European bishops and governments to make frontier-model risk a public conversation, warning that the alternative is saying "five months or five years from now, 'Oh, I felt there was a disaster coming, but there was nothing I could do.'"
The Chinese rewrites that carried this over the weekend blur Czerny with Paolo Benanti, the friar POLITICO calls the Vatican's adviser on AI — a distinction worth keeping as the Church's AI shop takes shape.
What to watch: whether Absolics publishes package-level reliability results by the end of the year, or any Korean glass maker names a first customer.
If the people writing the rules for AI haven't used the tools, whose experience should the rules be built on? Tell us in the comments.
