this paper is f*cking insane
the strongest quant firms are not trying to predict every market move
they isolate small repeatable edges, keep risk tightly controlled, and let thousands of executions do the compounding
the result: a strategy can be correct only slightly more often than random and still generate powerful long-term returns
the strange part is how unimpressive the edge can look
a short-lived mispricing
a residual far from normal
a sliver of positive expected value
one trade means almost nothing
but repeated with disciplined sizing, clean execution, and enough independent opportunities, that tiny advantage becomes a scalable system
bookmark this before it gets lost
Also love that MS is addressing Glass Core Substrate for future CPO adoptions!
_
What is glass core substrate and what are the pros and cons of adopting it?
Glass core substrate is a next-generation substrate that replaces the traditional core layer in an ABF substrate with a sheet of glass. It offers superior flatness and rigidity (low CTE) compared to today’s organic substrates, which could help suppliers overcome the physical limits of organic materials for future generations of massive AI chips (based on our supply chain checks, 150mm × 150mm could be the area size limit for organic substrates).
Glass core substrate also enables higher interconnect density through through-glass vias (TGV), which have smaller diameters and finer pitch than laser-drilled holes in organic substrates. This significantly increases routing density and connections per package—essential for advanced packaging.
Another key prospect is its use in Co-Packaged Optics (CPO): photonic components can be integrated directly into the package. Because glass is transparent, it can act as a waveguide, allowing AI chips to communicate via light traveling through the glass core itself instead of copper wiring, which reduces power consumption.
While the technical advantages are compelling, several challenges remain for the supply chain to achieve mass production at reasonable cost:
- Fragility: Glass is brittle by nature, making it prone to cracks and difficult to handle.
- TGV fabrication is costly and slow: Traditional laser drilling is too slow for mass production, so some suppliers have shifted to laser-induced deep etching (laser weakens the glass, followed by chemical etch). Filling TGVs with copper (metallization) is also technically challenging.
- Adhesion issues: Metal and glass do not naturally bond well; copper layers can peel off during repeated high-temperature heating and cooling cycles.
- Inspection difficulties: Glass transparency disrupts traditional automated optical inspection (AOI) systems.
- Immature ecosystem: The entire glass core substrate supply chain is still in its infancy, hindering standardization and cost reduction.
Overall, producing a glass core substrate is currently several times more expensive than today’s organic substrates. Over time, as technology matures and costs decline, glass core could become more mainstream, but near-term adoption is expected to remain limited.
AI is becoming a DRAM black hole.
Experts estimate that by 2026, AI workloads could effectively consume nearly 20% of global DRAM capacity.
The shift is clear: the bottleneck is no longer compute alone, but memory capacity and bandwidth.
As cloud inference scales, high-speed memory like HBM and GDDR7 is absorbing disproportionate silicon and capacity, crowding out traditional PC, smartphone, and standard server DRAM.
This isn’t a temporary spike — it’s a structural rebalancing of the memory market driven by AI.