ValueChain started from a simple question: which company earns its money from whom, and pays its money to whom?
Get those relationships right and you can see where a company's results are fragile and where they hold. The problem is that this data sits scattered across filings, earnings calls, and news, and most of it circulates without verification.
So instead of another flat list of companies, we built a graph of how money moves. Nodes are companies; edges are who sells to whom. It now spans 1,359 companies, but the number we watch is not the count, it is how deep each node goes. Rather than spreading wide and thin, we take one subgraph and work it all the way down until you can reason across it in several hops.
The first set is the HBM and AI-memory chain. It runs across four layers: from AI-chip demand (NVIDIA, AMD, Broadcom, Google, Tesla) to the three memory makers (SK hynix, Samsung, Micron), to the HBM back-end equipment vendors (Hanmi, Hanwha, ASMPT, BESI), and down to the foundries (TSMC, Samsung).
The point is that we never just write down a number. Every fact carries a source, a date, a unit, and a basis. One supply contract becomes one sourced row, and three years of history is those rows stacked. And we always check the counterparty. Once, Hanmi's fifth contract turned out to be with ASE, not SK. Without looking, we would have quietly misattributed it.
Trust is marked in three tiers. If a fact sits directly in the company's own regulatory filing it is confirmed; if a counterparty filing or several outlets corroborate it, cross-verified; a single mention stays estimated and labeled unverified. We check that SK sells to NVIDIA by reading it back out of NVIDIA's own 10-K.
If you let an unrefined node spread sideways, the whole thing gets shallow. So each node has to pass a score, thirteen checks run in code, before it earns the right to expand. No human waves it through. The graph grows one verified layer at a time.
Data built this way answers questions a plain model cannot: HBM customer-concentration risk, shifts in HBM4 equipment share, the second-order beneficiaries of NVIDIA's capex, foundry dependence, single-supplier bottlenecks. Each of those needs a current number, a source, and a cross-check.
It began as a tool to inform our own investing, but the data itself became a product, and now a grounding pack that research and investment agents query directly over MCP. To see who really benefits in the AI cycle, you eventually have to know who sells to whom.
For information only. This is not investment advice.